You are on page 1of 972

Edward H.

James J. Cimino Editors


Computer Applications in
Health Care and Biomedicine
Fourth Edition


Biomedical Informatics

(021) 66485438 66485457

Edward H. Shortliffe • James J. Cimino

Biomedical Informatics
Computer Applications in Health
Care and Biomedicine

Fourth Edition

(021) 66485438 66485457

Edward H. Shortliffe, MD, PhD James J. Cimino, MD
Departments of Biomedical Informatics Bethesda, MD
at Columbia University USA
and Arizona State University
New York, NY

ISBN 978-1-4471-4473-1 ISBN 978-1-4471-4474-8 (eBook)
DOI 10.1007/978-1-4471-4474-8
Springer London Heidelberg New York Dordrecht

Library of Congress Control Number: 2013955588

© Springer-Verlag London 2014
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or
part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way,
and transmission or information storage and retrieval, electronic adaptation, computer software,
or by similar or dissimilar methodology now known or hereafter developed. Exempted from this
legal reservation are brief excerpts in connection with reviews or scholarly analysis or material
supplied specifically for the purpose of being entered and executed on a computer system, for
exclusive use by the purchaser of the work. Duplication of this publication or parts thereof is
permitted only under the provisions of the Copyright Law of the Publisher's location, in its
current version, and permission for use must always be obtained from Springer. Permissions for
use may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable
to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are
exempt from the relevant protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility
for any errors or omissions that may be made. The publisher makes no warranty, express or
implied, with respect to the material contained herein.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (

(021) 66485438 66485457

Dedicated to Homer R. Warner, MD, PhD, FACMI
A Principal Founder of the Field of Biomedical Informatics

The Fourth Edition of Biomedical Informatics: Computer Applications in
Health Care and Biomedicine is dedicated to the memory and professional
contributions of Homer R. Warner. Homer was not only a pioneer in biomedi-
cal informatics but a sustained contributor who is truly one of the founders of
the field that mourned his loss in November of 2012. Homer’s publications on
the use of computers in health care span 50 years, from 1963 to 2012, but he
can claim an additional decade of informatics research that predated digital
computer use, including the use of analog computers and mathematical models
ranging from details of cardiac function all the way up to medical diagnosis.1
He is best known for his development of the Health Evaluation through
Logical Processing (HELP) system, which was revolutionary in its own right
as a hospital information system, but was truly visionary in its inclusion of
the logical modules for generating alerts and reminders. The HELP system,

Warner, H. R., Toronto, A. F., Veasey, L. G., & Stephenson, R. 1961. A mathematical
approach to medical diagnosis. Application to congenital heart disease. JAMA: The Journal
of the American Medical Association, 177, 177–183.

(021) 66485438 66485457

begun in 1968, is still running today at the LDS Hospital in Salt Lake City;
innovations are continually added while commercial systems struggle to rep-
licate functions that HELP has had for almost half a century. Homer’s other
contributions are far too numerous to recount here, but you will find them
described in no less than six different chapters of this book.
Homer’s contributions go far beyond merely the scientific foundation of bio-
medical informatics. He also provided extensive leadership to define informatics
as a separate academic field. He accomplished this in many settings; locally by
founding the first degree-granting informatics department at the University of
Utah, nationally as the President of the American College of Medical Informatics,
and internationally as the founding editor of the well-known and influential jour-
nal Computers and Biomedical Research (now the Journal of Biomedical
Informatics). But perhaps his greatest impact is the generations of researchers
and trainees that he personally inspired who have gone on to mentor additional
researchers and trainees who together are the life blood of biomedical informat-
ics. Homer’s true influence on the field is therefore incalculable. Just consider the
convenience sample of this book’s 60 chapter co-authors: the following diagram
shows his lineage of professional influence on 52 of us.2
Both of us were privileged to have many professional and personal inter-
actions with Homer and we were always struck by his enthusiasm, energy,
humor, generosity, and integrity. In 1994, Homer received the American
College of Medical Informatics’ highest honor, the Morris F Collen Award of
Excellence. We are proud to have this opportunity to add to the recognition of
Homer’s life and career with this dedication.

James J. Cimino
Edward H. Shortliffe

Homer R. Warner

Scott Narus
Valerie Florance Stanley M. Huff Clement J. McDonald Charles P. Friedman
Scott Evans
Reed M. Gardner
Paul D. Clayton Terry Clemmer G. Octo Barnett W. Edward Hammond Peter Szolovits
Randolph A. Miller
David Vawdrey
Judy G. Ozbolt
Roger B. Mark Issac Kohane
William A. Yasnoff
Kenneth W. Goodman
Carol Friedman James J. Cimino Robert A. Greenes
Lynn Vogel

Kenneth Mandl
George Hripcsak Vimla L. Patel Edward H. Shortliffe William Hersh

David R. Kaufman Peter Embi
Adam Wilcox
Noémie Elhadad David W. Bates
Justin B. Starren
Douglas K. Owens Parvati Dev
James Brinkley
Robert Rudin Paul C. Tang Kevin B. Johnson Blackford Middleton

Mark A. Musen Mark E. Frisse Suzanne Bakken
Philip Payne Michael Chiang Holly Jimison Russ B. Altman

Jonathan Silverstein
Patricia Dykes
Jessica Tenenbaum Sean D. Mooney Nigam Shah
Daniel L. Rubin

Ian Foster

Paul Clayton and Peter Szolovits provide important connections between Homer Warner
and ten coauthors but, while they are informatics leaders in their own right, they are not
contributors to this edition of this book.

(021) 66485438 66485457

Preface to the Fourth Edition

The world of biomedical research and health care has changed remarkably in
the 25 years since the first edition of this book was undertaken. So too has the
world of computing and communications and thus the underlying scientific
issues that sit at the intersections among biomedical science, patient care, pub-
lic health, and information technology. It is no longer necessary to argue that
it has become impossible to practice modern medicine, or to conduct modern
biological research, without information technologies. Since the initiation of
the human genome project two decades ago, life scientists have been generat-
ing data at a rate that defies traditional methods for information management
and data analysis. Health professionals also are constantly reminded that a
large percentage of their activities relates to information management—for
example, obtaining and recording information about patients, consulting col-
leagues, reading and assessing the scientific literature, planning diagnostic
procedures, devising strategies for patient care, interpreting results of labora-
tory and radiologic studies, or conducting case-based and population-based
research. It is complexity and uncertainty, plus society’s overriding concern
for patient well-being, and the resulting need for optimal decision making, that
set medicine and health apart from many other information-intensive fields.
Our desire to provide the best possible health and health care for our society
gives a special significance to the effective organization and management of
the huge bodies of data with which health professionals and biomedical
researchers must deal. It also suggests the need for specialized approaches and
for skilled scientists who are knowledgeable about human biology, clinical
care, information technologies, and the scientific issues that drive the effective
use of such technologies in the biomedical context.

Information Management in Biomedicine

The clinical and research influence of biomedical-computing systems is
remarkably broad. Clinical information systems, which provide communica-
tion and information-management functions, are now installed in essentially
all healthcare institutions. Physicians can search entire drug indexes in a few
seconds, using the information provided by a computer program to anticipate
harmful side effects or drug interactions. Electrocardiograms (ECGs) are
typically analyzed initially by computer programs, and similar techniques are
being applied for interpretation of pulmonary-function tests and a variety of


(021) 66485438 66485457

viii Preface to the Fourth Edition

laboratory and radiologic abnormalities. Devices with embedded processors
routinely monitor patients and provide warnings in critical-care settings, such
as the intensive-care unit (ICU) or the operating room. Both biomedical
researchers and clinicians regularly use computer programs to search the
medical literature, and modern clinical research would be severely hampered
without computer-based data-storage techniques and statistical analysis sys-
tems. Advanced decision-support tools also are emerging from research labo-
ratories, are being integrated with patient-care systems, and are beginning to
have a profound effect on the way medicine is practiced.
Despite this extensive use of computers in healthcare settings and bio-
medical research, and a resulting expansion of interest in learning more about
biomedical computing, many life scientists, health-science students, and pro-
fessionals have found it difficult to obtain a comprehensive and rigorous, but
nontechnical, overview of the field. Both practitioners and basic scientists are
recognizing that thorough preparation for their professional futures requires
that they gain an understanding of the state of the art in biomedical comput-
ing, of the current and future capabilities and limitations of the technology,
and of the way in which such developments fit within the scientific, social,
and financial context of biomedicine and our healthcare system. In turn, the
future of the biomedical computing field will be largely determined by how
well health professionals and biomedical scientists are prepared to guide and
to capitalize upon the discipline’s development. This book is intended to meet
this growing need for such well-equipped professionals. The first edition
appeared in 1990 (published by Addison-Wesley) and was used extensively
in courses on medical informatics throughout the world. It was updated with
a second edition (published by Springer) in 2000, responding to the remark-
able changes that occurred during the 1990s, most notably the introduction of
the World Wide Web and its impact on adoption and acceptance of the
Internet. The third edition (again published by Springer) appeared in 2006,
reflecting the ongoing rapid evolution of both technology and health- and
biomedically-related applications, plus the emerging government recognition
of the key role that health information technology would need to play in pro-
moting quality, safety, and efficiency in patient care. With that edition the title
of the book was changed from Medical Informatics to Biomedical Informatics,
reflecting (as is discussed in Chap. 1) both the increasing breadth of the basic
discipline and the evolving new name for academic units, societies, research
programs, and publications in the field. Like the first three editions, this new
version provides a conceptual framework for learning about the science that
underlies applications of computing and communications technology in bio-
medicine and health care, for understanding the state of the art in computer
applications in clinical care and biology, for critiquing existing systems, and
for anticipating future directions that the field may take.
In many respects, this new edition is very different from its predecessors,
however. Most importantly, it reflects the remarkable changes in computing
and communications that continue to occur, most notably in communications,
networking, and health information technology policy, and the exploding
interest in the role that information technology must play in systems integra-
tion and the melding of genomics with innovations in clinical practice and

(021) 66485438 66485457

Preface to the Fourth Edition ix

treatment. In addition, new chapters have been introduced, one (healthcare
financing) was eliminated, while others have been revamped. We have intro-
duced new chapters on the health information infrastructure, consumer health
informatics, telemedicine, translational bioinformatics, clinical research
informatics, and health information technology policy. Most of the previous
chapters have undergone extensive revisions. Those readers who are familiar
with the first three editions will find that the organization and philosophy are
unchanged, but the content is either new or extensively updated.1
This book differs from other introductions to the field in its broad coverage
and in its emphasis on the field’s conceptual underpinnings rather than on
technical details. Our book presumes no health- or computer-science back-
ground, but it does assume that you are interested in a comprehensive sum-
mary of the field that stresses the underlying concepts, and that introduces
technical details only to the extent that they are necessary to meet the princi-
pal goal. It thus differs from an impressive early text in the field (Ledley
1965) that emphasized technical details but did not dwell on the broader
social and clinical context in which biomedical computing systems are devel-
oped and implemented.

Overview and Guide to Use of This book

This book is written as a text so that it can be used in formal courses, but we
have adopted a broad view of the population for whom it is intended. Thus,
it may be used not only by students of medicine and of the other health
professions, but also as an introductory text by future biomedical informat-
ics professionals, as well as for self-study and for reference by practitio-
ners. The book is probably too detailed for use in a 2- or 3-day
continuing-education course, although it could be introduced as a reference
for further independent study.
Our principal goal in writing this text is to teach concepts in biomedical
informatics—the study of biomedical information and its use in decision
making—and to illustrate them in the context of descriptions of representa-
tive systems that are in use today or that taught us lessons in the past. As
you will see, biomedical informatics is more than the study of computers in
biomedicine, and we have organized the book to emphasize that point.
Chapter 1 first sets the stage for the rest of the book by providing a glimpse
of the future, defining important terms and concepts, describing the content
of the field, explaining the connections between biomedical informatics and
related disciplines, and discussing the forces that have influenced research
in biomedical informatics and its integration into clinical practice and bio-
logical research.

As with the first three editions, this book has tended to draw both its examples and it con-
tributors from North America. There is excellent work in other parts of the world as well,
although variations in healthcare systems, and especially financing, do tend to change the
way in which systems evolve from one country to the next. The basic concepts are identi-
cal, however, so the book is intended to be useful in educational programs in other parts of
the world as well.

(021) 66485438 66485457

x Preface to the Fourth Edition

Broad issues regarding the nature of data, information, and knowledge
pervade all areas of application, as do concepts related to optimal decision
making. Chapters 2 and 3 focus on these topics but mention computers only
in passing. They serve as the foundation for all that follows. Chapter 4 on
cognitive science issues enhances the discussions in Chaps. 2 and 3, pointing
out that decision making and behavior are deeply rooted in the ways in which
information is processed by the human mind. Key concepts underlying sys-
tem design, human-computer interaction, patient safety, educational technol-
ogy, and decision making are introduced in this chapter.
Chapters 5 and 6 introduce the central notions of computer architectures
and software engineering that are important for understanding the applications
described later. Also included is a discussion of computer-system design, with
explanations of important issues for you to consider when you read about
specific applications and systems throughout the remainder of this book.
Chapter 7 summarizes the issues of standards development, focusing in
particular on data exchange and issues related to sharing of clinical data. This
important and rapidly evolving topic warrants inclusion given the evolution
of the health information exchange, institutional system integration chal-
lenges, and the increasingly central role of standards in enabling clinical sys-
tems to have their desired influence on healthcare practices.
Chapter 8 addresses a topic of increasing practical relevance in both the
clinical and biological worlds: natural language understanding and the pro-
cessing of biomedical texts. The importance of these methods is clear when
one considers the amount of information contained in free-text dictated notes
or in the published biomedical literature. Even with efforts to encourage
structured data entry in clinical systems, there will likely always be an impor-
tant role for techniques that allow computer systems to extract meaning from
natural language documents.
Chapter 9 is a comprehensive introduction to the conceptual underpin-
nings of biomedical and clinical image capture, analysis, interpretation and
use. This overview of the basic issues and imaging modalities serves as back-
ground for Chap. 20, which deals with imaging applications issues, high-
lighted in the world of radiological imaging and image management (e.g., in
picture archiving and communication systems).
Chapter 10 addresses the key legal and ethical issues that have arisen when
health information systems are considered. Then, in Chap. 11, the challenges
associated with technology assessment and with the evaluation of clinical
information systems are introduced.
Chapters 12–26 (which include several new chapters in this edition) survey
many of the key biomedical areas in which computers are being used. Each
chapter explains the conceptual and organizational issues in building that type
of system, reviews the pertinent history, and examines the barriers to success-
ful implementations.
Chapter 27 is a new chapter in the fourth edition, providing a summary of
the rapidly evolving policy issues related to health information technology.
Although the emphasis is on US government policy, there is some discussion
of issues that clearly generalize both to states (in the US) and to other countries.
The book concludes in Chap. 28 with a look to the future—a vision of how

(021) 66485438 66485457

Preface to the Fourth Edition xi

informatics concepts, computers, and advanced communication devices one
day may pervade every aspect of biomedical research and clinical practice.

The Study of Computer Applications in Biomedicine

The actual and potential uses of computers in health care and biomedicine
form a remarkably broad and complex topic. However, just as you do not
need to understand how a telephone or an ATM machine works to make good
use of it and to tell when it is functioning poorly, we believe that technical
biomedical-computing skills are not needed by health workers and life scien-
tists who wish simply to become effective users of evolving information tech-
nologies. On the other hand, such technical skills are of course necessary for
individuals with career commitment to developing information systems for
biomedical and health environments. Thus, this book will neither teach you
to be a programmer, nor show you how to fix a broken computer (although it
might motivate you to learn how to do both). It also will not tell you about
every important biomedical-computing system or application; we shall use an
extensive bibliography to direct you to a wealth of literature where review
articles and individual project reports can be found. We describe specific sys-
tems only as examples that can provide you with an understanding of the
conceptual and organizational issues to be addressed in building systems for
such uses. Examples also help to reveal the remaining barriers to successful
implementations. Some of the application systems described in the book are
well established, even in the commercial marketplace. Others are just begin-
ning to be used broadly in biomedical settings. Several are still largely con-
fined to the research laboratory.
Because we wish to emphasize the concepts underlying this field, we gen-
erally limit the discussion of technical implementation details. The computer-
science issues can be learned from other courses and other textbooks. One
exception, however, is our emphasis on the details of decision science as they
relate to biomedical problem solving (Chaps. 3 and 22). These topics gener-
ally are not presented in computer-science courses, yet they play a central
role in the intelligent use of biomedical data and knowledge. Sections on
medical decision making and computer-assisted decision support accordingly
include more technical detail than you will find in other chapters.
All chapters include an annotated list of Suggested Readings to which you
can turn if you have a particular interest in a topic, and there is a comprehen-
sive Bibliography, drawn from the individual chapters, at the end of the book.
We use boldface print to indicate the key terms of each chapter; the defini-
tions of these terms are included in the Glossary at the end of the book.
Because many of the issues in biomedical informatics are conceptual, we
have included Questions for Discussion at the end of each chapter. You will
quickly discover that most of these questions do not have “right” answers.
They are intended to illuminate key issues in the field and to motivate you to
examine additional readings and new areas of research.
It is inherently limiting to learn about computer applications solely by
reading about them. We accordingly encourage you to complement your

(021) 66485438 66485457

xii Preface to the Fourth Edition

studies by seeing real systems in use—ideally by using them yourself. Your
understanding of system limitations and of what you would do to improve a
biomedical-computing system will be greatly enhanced if you have had per-
sonal experience with representative applications. Be aggressive in seeking
opportunities to observe and use working systems.
In a field that is changing as rapidly as biomedical informatics is, it is diffi-
cult ever to feel that you have knowledge that is completely current. However,
the conceptual basis for study changes much more slowly than do the detailed
technological issues. Thus, the lessons you learn from this volume will provide
you with a foundation on which you can continue to build in the years ahead.

The Need for a Course in Biomedical Informatics

A suggestion that new courses are needed in the curricula for students of the
health professions is generally not met with enthusiasm. If anything, educators
and students have been clamoring for reduced lecture time, for more emphasis
on small group sessions, and for more free time for problem solving and reflec-
tion. A 1984 national survey by the Association of American Medical Colleges
found that both medical students and their educators severely criticized the
traditional emphasis on lectures and memorization. Yet the analysis of a panel
on the General Professional Education of the Physician (GPEP) (Association of
American Medical Colleges 1984) and several subsequent studies and reports
have specifically identified biomedical informatics, including computer appli-
cations, as an area in which new educational opportunities need to be developed
so that physicians and other health professionals will be better prepared for
clinical practice. The AAMC recommended the formation of new academic
units in biomedical informatics in our medical schools, and subsequent studies
and reports have continued to stress the importance of the field and the need for
its inclusion in the educational environments of health professionals.
The reason for this strong recommendation is clear: The practice of medi-
cine is inextricably entwined with the management of information. In the past,
practitioners handled medical information through resources such as the near-
est hospital or medical-school library; personal collections of books, journals,
and reprints; files of patient records; consultation with colleagues; manual
office bookkeeping; and (all-too-often flawed) memorization. Although these
techniques continue to be variably valuable, information technology is offering
new methods for finding, filing, and sorting information: online bibliographic-
retrieval systems, including full-text publications; personal computers, laptops,
tablets, and smart phones, with database software to maintain personal infor-
mation and commonly used references; office-practice and clinical information
systems to capture, communicate, and preserve key elements of the health
record; information retrieval and consultation systems to provide assistance
when an answer to a question is needed rapidly; practice-management systems
to integrate billing and receivable functions with other aspects of office or clinic
organization; and other online information resources that help to reduce the

(021) 66485438 66485457

Preface to the Fourth Edition xiii

pressure to memorize in a field that defies total mastery of all but its narrowest
aspects. With such a pervasive and inevitable role for computers in clinical
practice, and with a growing failure of traditional techniques to deal with the
rapidly increasing information-management needs of practitioners, it has
become obvious to many people that an essential topic has emerged for study
in schools that train medical and other health professionals.
What is less clear is how the subject should be taught, and to what extent
it should be left for postgraduate education. We believe that topics in bio-
medical informatics are best taught and learned in the context of health-
science training, which allows concepts from both the health sciences and
informatics science to be integrated. Biomedical-computing novices are
likely to have only limited opportunities for intensive study of the material
once their health-professional training has been completed.
The format of biomedical informatics education is certain to evolve as fac-
ulty members are hired to develop it at more health-science schools, and as the
emphasis on lectures as the primary teaching method continues to diminish.
Computers will be used increasingly as teaching tools and as devices for com-
munication, problem solving, and data sharing among students and faculty. In
the meantime, key content in biomedical informatics will likely be taught
largely in the classroom setting. This book is designed to be used in that kind
of traditional course, although the Questions for Discussion also could be used
to focus conversation in small seminars and working groups. As resources
improve in schools and academic medical centers, integration of biomedical
informatics topics into clinical experiences also will become more common.
The eventual goal should be to provide instruction in biomedical informatics
whenever this field is most relevant to the topic the student is studying. This
aim requires educational opportunities throughout the years of formal training,
supplemented by continuing-education programs after graduation.
The goal of integrating biomedicine and biomedical informatics is to pro-
vide a mechanism for increasing the sophistication of health professionals, so
that they know and understand the available resources. They also should be
familiar with biomedical computing’s successes and failures, its research
frontiers and its limitations, so that they can avoid repeating the mistakes of
the past. Study of biomedical informatics also should improve their skills in
information management and problem solving. With a suitable integration of
hands-on computer experience, computer-based learning, courses in clinical
problem solving, and study of the material in this volume, health-science
students will be well prepared to make effective use of computer-based tools
and information management in healthcare delivery.

The Need for Specialists in Biomedical Informatics

As mentioned, this book also is intended to be used as an introductory text in
programs of study for people who intend to make their professional careers in
biomedical informatics. If we have persuaded you that a course in biomedical

(021) 66485438 66485457

xiv Preface to the Fourth Edition

informatics is needed, then the requirement for trained faculty to teach the
courses will be obvious. Some people might argue, however, that a course on
this subject could be taught by a computer scientist who had an interest in
biomedical computing, or by a physician or biologist who had taken a few
computing courses. Indeed, in the past, most teaching—and research—has
been undertaken by faculty trained primarily in one of the fields and later
drawn to the other. Today, however, schools have come to realize the need for
professionals trained specifically at the interfaces among biomedicine, bio-
medical informatics, and related disciplines such as computer science, statis-
tics, cognitive science, health economics, and medical ethics. This book
outlines a first course for students training for careers in the biomedical infor-
matics field. We specifically address the need for an educational experience in
which computing and information-science concepts are synthesized with bio-
medical issues regarding research, training, and clinical practice. It is the inte-
gration of the related disciplines that traditionally has been lacking in the
educational opportunities available to students with career interests in bio-
medical informatics. If schools are to establish such courses and training pro-
grams (and there are growing numbers of examples of each), they clearly need
educators who have a broad familiarity with the field and who can develop
curricula for students of the health professions as well as of informatics itself.
The increasing introduction of computing techniques into biomedical envi-
ronments will require that well-trained individuals be available not only to teach
students, but also to design, develop, select, and manage the biomedical-
computing systems of tomorrow. There is a wide range of context-dependent
computing issues that people can appreciate only by working on problems
defined by the healthcare setting and its constraints. The field’s development has
been hampered because there are relatively few trained personnel to design
research programs, to carry out the experimental and developmental activities,
and to provide academic leadership in biomedical informatics. A frequently
cited problem is the difficulty a health professional (or a biologist) and a techni-
cally trained computer scientist experience when they try to communicate with
one another. The vocabularies of the two fields are complex and have little over-
lap, and there is a process of acculturation to biomedicine that is difficult for
computer scientists to appreciate through distant observation. Thus, interdisci-
plinary research and development projects are more likely to be successful when
they are led by people who can effectively bridge the biomedical and computing
fields. Such professionals often can facilitate sensitive communication among
program personnel whose backgrounds and training differ substantially.
It is exciting to be working in a field that is maturing and that is having a
beneficial effect on society. There is ample opportunity remaining for innova-
tion as new technologies evolve and fundamental computing problems
succumb to the creativity and hard work of our colleagues. In light of the

(021) 66485438 66485457

Preface to the Fourth Edition xv

increasing sophistication and specialization required in computer science in
general, it is hardly surprising that a new discipline should arise at that field’s
interface with biomedicine. This book is dedicated to clarifying the definition
and to nurturing the effectiveness of that discipline: biomedical informatics.

New York, NY Edward H. Shortliffe
Bethesda, MD James J. Cimino
October 2013

(021) 66485438 66485457

(021) 66485438 66485457


In the 1980s, when I was based at Stanford University, I conferred with colleagues
Larry Fagan and Gio Wiederhold and we decided to compile the first comprehen-
sive textbook on what was then called medical informatics. As it turned out, none
of us predicted the enormity of the task we were about to undertake. Our challenge
was to create a multi-authored textbook that captured the collective expertise of
leaders in the field yet was cohesive in content and style. The concept for the book
first developed in 1982. We had begun to teach a course on computer applications
in health care at Stanford’s School of Medicine and had quickly determined that
there was no comprehensive introductory text on the subject. Despite several pub-
lished collections of research descriptions and subject reviews, none had been
developed with the needs of a rigorous introductory course in mind.
The thought of writing a textbook was daunting due to the diversity of top-
ics. None of us felt that he was sufficiently expert in the full range of impor-
tant subjects for us to write the book ourselves. Yet we wanted to avoid
putting together a collection of disconnected chapters containing assorted
subject reviews. Thus, we decided to solicit contributions from leaders in the
respective fields to be represented but to provide organizational guidelines in
advance for each chapter. We also urged contributors to avoid writing subject
reviews but, instead, to focus on the key conceptual topics in their field and to
pick a handful of examples to illustrate their didactic points.
As the draft chapters began to come in, we realized that major editing would
be required if we were to achieve our goals of cohesiveness and a uniform orien-
tation across all the chapters. We were thus delighted when, in 1987, Leslie
Perreault, a graduate of our training program, assumed responsibility for rework-
ing the individual chapters to make an integral whole and for bringing the project
to completion. The final product, published in 1990, was the result of many
compromises, heavy editing, detailed rewriting, and numerous iterations. We
were gratified by the positive response to the book when it finally appeared, and
especially by the students of biomedical informatics who have often come to us
at scientific meetings and told us about their appreciation of the book.
As the 1990s progressed, however, we began to realize that, despite our
emphasis on basic concepts in the field (rather than a survey of existing sys-
tems), the volume was beginning to show its age. A great deal had changed
since the initial chapters were written, and it became clear that a new edition
would be required. The original editors discussed the project and decided that
we should redesign the book, solicit updated chapters, and publish a new
edition. Leslie Perreault by this time was a busy Director at First Consulting


(021) 66485438 66485457

Without his hard work and expertise. the chapter authors did a marvelous job. Similarly. who has been a superb and talented collaborator in this effort for the last two editions. We were once again extremely appreciative of all the authors’ commit- ment and for the excellence of their work on behalf of the book and the field. MD. Shortliffe (021) 66485438 66485457 www. Particular gratitude is owed to Maureen Alexander. With that in mind we. we embarked on the new project. as plans for a third edition began to take shape. agreed to embark on a fourth edition. Texas. And I want to offer my sincere personal thanks to Jim Cimino. We began to realize that the maintenance of a textbook in a field such as biomedical informatics was nearly a constant. attempting to cover additional key topics that readers and authors had identified as being necessary enhancements to the earlier edi- tions. The completed fourth edition reflects the work and support of many peo- ple in addition to the editors and chapter authors. our developmental editor whose rigorous attention to detail was crucial given the size and the complexity of the undertaking. putting up with editing changes that were designed to bring a uni- form style to the book. Arizona. At Springer we have been delighted to work on this edition with Grant Weston. No sooner had the second edition appeared in print than we started to get inquiries about when the next update would appear. we began working on the fourth edition.ketabpezeshki. Accordingly. then Houston. even in a young field like informatics!).com . New York. NY Edward H. Fagan. Once again the authors did their best to try to meet our deadlines as the third edition took shape. it was only a short time after the publication of the third edi- tion that we began to get queries about a fourth edition. We resisted for a year or two but it became clear that the third edition was becoming rapidly stale in some key areas and that there were new topics that were not in the book and needed to be added. finally completing the effort in early 2013. who has been extremely supportive despite our missed deadlines. in light of our knowledge of the work that would be involved. Progress was slowed by my professional moves (to Phoenix. Jim Cimino left Columbia to assume new responsibilities at the NIH Clinical Center in Bethesda. With several new chapters in mind. in consultation with Grant Weston from Springer’s offices in London. Predictably. my Columbia colleague Jim Cimino joined me as the new associate editor. With trepidation. whereas Drs. Wiederhold. and contributing excellent chapters that nicely reflected the changes in the field in the preceding decade. and Perreault continued to be involved as chapter authors. As before. ongoing process. By this time I had moved to Columbia University and the initial group of editors had largely disbanded to take on other responsi- bilities. and the need to change authors of some of the existing chapters due to retire- ments (this too will happen. with Leslie Perreault no longer available. This time we added several chapters.xviii Acknowledgments Group in New York City and would not have as much time to devote to the project as she had when we did the first edition. trying to meet our deadlines. we would still be struggling to complete the massive editing job associated with this now very long manuscript. and then back to New York) with a very busy three-year stint as President and CEO of the American Medical Informatics Association.

. . . and Use . . . . . . 3 Edward H. Shortliffe and Marsden S. . . . 211 W. . . . . . . . . . . . . . . . . Silverstein and Ian T. . . . . Patel and David R. Sox 4 Cognitive Science and Biomedical Informatics . . . . . . . Hayit Greenspan. . . Rubin. . Miller 11 Evaluation of Biomedical and Health Information Resources . Owens and Harold C. and James F. . 185 Adam B. Edward Hammond. . . Brinkley 10 Ethics in Biomedical and Health Informatics: Users. . . 255 Carol Friedman and Noémie Elhadad 9 Biomedical Imaging Informatics . . . . . . . 355 Charles P. . . . . . . . Vawdrey 7 Standards in Biomedical Informatics . . Scott P. Charles Jaffe. . . and Randolph A. . . . . . . .ketabpezeshki. . Blois 2 Biomedical Data: Their Acquisition. . . 149 Jonathan C. . . . 329 Kenneth W. . . . . . . . Octo Barnett 3 Biomedical Decision Making: Probabilistic Clinical Reasoning . . . . . . . . 285 Daniel L. . . . . . . . . . . . . . . . . . . . . . . . and Outcomes . Kaufman 5 Computer Architectures for Health Care and Biomedicine . . . . . . . . . 109 Vimla L. Standards. . . . . . Huff 8 Natural Language Processing in Health Care and Biomedicine . . . . . . . . . . . 67 Douglas K. . . . . . . . . . Reid Cushman. Cimino. . . . and Stanley M. . . . Storage. . . . . . Friedman and Jeremy C. . . . . James J. . . . . . Foster 6 Software Engineering for Health Care and Biomedicine . . . . . . . and David K. . . . Goodman. . . . Shortliffe and G. . . . . . Wyatt xix (021) 66485438 66485457 www. . . . . . Contents Part I Recurrent Themes in Biomedical Informatics 1 Biomedical Informatics: The Science and the Pragmatics . . . Wilcox. . . . . 39 Edward H. . . . . . . . . . . . . . .

. . Altman 26 Clinical Research Informatics . . . . . . . . . . . . . . . Shah. . . . . . . . . . . . . Chiang 19 Patient Monitoring Systems . . . . . . . . . . Embi. . . 443 Lynn Harold Vogel 15 Patient-Centered Care Systems . . . . and William A. . . 755 Philip R. . . . . . McDonald. . . . and James J. . . . . . . . . . . . . . . Yasnoff 17 Consumer Health Informatics and Personal Health Records . . . . . . . . . . . . Yasnoff 14 Management of Information in Health Care Organizations . . Nigam H. . . . . . Clemmer. . . . . . . . Payne. and Russ B. . . . . . . . . 695 Sean D. . 517 Kevin Johnson. . 541 Justin B. . . . . 561 Reed M. . Altman 25 Translational Bioinformatics . . . . . Thomas S. . . . . . . and Michael F. . . .ketabpezeshki. . . . . . . . . . Greenes 21 Information Retrieval and Digital Libraries . . 423 William A. . . . . and Russ B. . . . 593 Bradley Erickson and Robert A. . . . . . . . . . . . . . . . Mooney. . . . . . . . . . . Tenenbaum. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dykes 16 Public Health Informatics . . Hersh 22 Clinical Decision-Support Systems . . . . . . . . . .xx Contents Part II Biomedical Informatics Applications 12 Electronic Health Record Systems . . . . . . . . . . Suzanne Bakken. . . . . . . . . . . . . . . . . . Tang. . . . . . . . . . . and Kenneth D. . . . . . . Blackford Middleton. . . 721 Jessica D. . . . . . . . . . . . Paul C. . . . Gardner. . . . . . 503 Martin LaVenture. . . . . . Peter J.L. . . . . 643 Mark A. . . . . . . . . . . . . . . . . . . . Scott Evans. and George Hripcsak 13 Health Information Infrastructure . . . . . . Musen. . . Greenes 23 Computers in Health Care Education. . . . . . . Mandl 18 Telehealth . . 613 William R. . . . Mark 20 Imaging Systems in Radiology . . . . 675 Parvati Dev and Titus K. . . . . . . . . . Jessica D. . . Schleyer 24 Bioinformatics . and Roger G. . . . . . . . . . . . R. . Nesbitt. . . . 391 Clement J. . and Robert A. . 475 Judy Ozbolt. . . . . . . . . Terry P. . . . Starren. . Cimino (021) 66485438 66485457 www. . . and Patricia C. . David A. . Holly Brugge Jimison. . . . Ross. . . . . . . . . . . . . . . . . . .O. . . . . . . . . . . . . . Tenenbaum. . . . . . . . . . . . . . . . . . . . . . . .com . . .

. . . . . . . . 943 (021) 66485438 66485457 www. Paul C. . . . . . and David W. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . .com . Kohane Glossary . . . . . Rudin. . . . . . . . . . . . . . . . . Valerie Florance. . . . . . . 781 Robert S. . . . . . . . . . . . . . . 927 Subject Index . . . . . . . . . . 865 Name Index . . . Frisse. . . . . . . . . . . . . . . . . .ketabpezeshki. . . . . . . . . . Tang. . . . . . . Kenneth D. . . . . . . . . . . . . . . . . . . . . . . . 797 Mark E. . . . . . . . . . . . . . . Bates 28 The Future of Informatics in Biomedicine . . . . . . . . . . and Isaac S. . . . . . . . . .Contents xxi Part III Biomedical Informatics in the Years Ahead 27 Health Information Technology Policy . . . . Mandl. . . . . . . .

com .ketabpezeshki.(021) 66485438 66485457 www.

USA Parvati Dev. MD. OH. MA Department of Ophthalmology and Medical Informatics and Clinical Epidemiology. PhD.ketabpezeshki. Altman. USA G. MD. Clemmer. USA Noémie Elhadad. MS. WA. Brinkley. MD. USA Patricia C. MD. Stanford University. CA. MD. NY. Biomedical Education and Medical Education. FACMI Innovation in Learning Inc. Computer Science and Engineering. USA Reid Cushman. FACMI Laboratory of Computer Science (Harvard Medical School and Massachusetts General Hospital). Columbia University. FACMI Center for Patient Safety Research and Practice. Bethesda. MA. FACMI Laboratory for Informatics . FACMI Departments of Biomedical Informatics and Internal Medicine. USA xxiii (021) 66485438 66485457 www. DNSc.. PhD Department of Medicine. Brigham and Women’s Hospital. FACP. USA James J. Salt Lake City. Oregon Health & Science University. MA. MA. FL. Embi. The Ohio State University Wexner Medical Center. RN. Octo Barnett. FAAN. USA Suzanne Bakken. PhD. MD. FACMI Division of General Internal Medicine and Primary Care. FACMI Department of Biological Structure. Columbus. USA James F. Columbia University. USA Michael F. MSc. PhD. University of Miami. OR. Cimino. PhD. School of Nursing. Brigham and Women’s Hospital. New York. MD. NY. LDS Hospital. University of Washington. Portland. Boston. MD. Los Alotos Hills. Bates. USA David W. USA Peter J. UT. Contributors Russ B. MD Pulmonary – Critical Care Medicine. Chiang. Miami. Dykes. Department of Medicine. FACMI Departments of Bioengineering. New York. Boston. Boston. MA. PhD Department of Biomedical Informatics. CA. Seattle. NIH Clinical Center. USA Terry P. FACMI Department of Biomedical Informatics. Genetics and Medicine. Stanford.

PhD. USA Robert A. National Library of Medicine. TN. NC. MD. Gardner. Tel Aviv. MBA. Friedman. USA Valerie Florance. USA Charles Jaffe. Greenes. Foster. FACMI Division of Extramural Programs. USA William R. USA Reed . USA Ian T. Hersh. Ann Arbor. USA Carol Friedman. USA Hayit Greenspan. Rochester. New York. FACMI Medical Informatics Department. USA Stanley M. USA (021) 66485438 66485457 www. Intermountain Healthcare. FACP Department of Medical Informatics and Clinical Epidemiology. USA Holly Brugge Jimison. Colleges of Computer and Information Sciences and Health Sciences. Vanderbilt University Medical Center. Columbia University. FACMI Department of Biomedical Informatics.ketabpezeshki. Biomedical Informatics. PhD Department of Biomedical Engineering. PhD. Mayo Clinic. MI. FL. Salt Lake City. PhD. New York. USA Division of Health Sciences Research. Huff. PhD. LDS Hospital. Columbia University Medical Center. MD. CA. PhD Searle Chemistry Laboratory. USA Kenneth W. PhD Health Level Seven International. Faculty of Engineering. MD. Murray. FACMI Consortium on Technology for Proactive Care. Bethesda. Nashville. MD. MS. Edward Hammond. University of Chicago and Argonne National Laboratory. MD. UT. FACMI Duke Center for Health Informatics. Arizona State University. Boston. USA R. Intermountain Healthcare. BS. USA Charles P. FACMI University of Miami Bioethics Program. IL. Del Mar. UT. NY. Miami. AZ. University of Michigan. Duke University Medical Center. Chicago. Frisse. College of Medicine. PhD. MD. TelAviv University. MS. Israel W. USA George Hripcsak.xxiv Contributors Bradley Erickson. FACMI Department of Biomedical Informatics. UT. FACMI Schools of Information and Public Health. NY. MD FACMI. Portland. PhD. DHHS. AZ. PhD. Tempe. Scottsdale. Goodman. Durham. Oregon Health and Science University. PhD Department of Radiology and Medical Informatics. PhD. Salt Lake City. National Institutes of Health. University of Utah. MN. Scott Evans. USA Mark E. PhD. Computation Institute. Northeastern University. MA. FACMI Department of Informatics. MS. FACMI Department of Biomedical Informatics. OR. Mayo Clinic. FACMI Medical Informatics. FACMI Department of Biomedical Informatics.

J. McDonald. MD. USA (021) 66485438 66485457 www. FAAN. FACMI Minnesota Department of Health. Mandl. Lister Hill National Center for Biomedical Communications. Scottsdale. Stanford University. DSc. Kaiser Center for Primary Care and Outcomes Research/Center for Health Policy. USA Judy Ozbolt. FACMI Harvard Medical School Center for Biomedical Informatics and Children’s Hospital Informatics Program. PhD Department of Biomedical Informatics. Musen. MPH.Contributors xxv Kevin Johnson. MPH. CA. Paul. RN. USA Mark A. USA Scott P. MS. The New York Academy of Medicine. Arizona State University. St. Mark. UT. PhD Buck Institute for Research on Aging. National Institutes of Health. MPH Department of Family and Community Medicine. Center for Health Informatics. Vanderbilt University School of Medicine. VA Palo Alto Health Care System and H. USA Martin LaVenture. USA Vimla L. Vanderbilt University Medical Center. Patel. UC Davis Health System. Kaufman. University of Maryland School of Nursing. FACMI Center for Cognitive Studies in Medicine and Public Health. Boston. USA Clement J. Office of HIT and e-Health. MD. MD. Novato. Vanderbilt University Medical Center. MD. PhD. Cambridge. USA Sean D. CA. CA. Mooney. Massachusetts Institute of Technology. Nesbitt. Owens. FACMI Children’s Hospital Informatics Program. FACMI Department of Biomedical Informatics. USA David R. USA Isaac S.ketabpezeshki. Narus. FACMI Informatics Center. PhD. USA Thomas S. Miller. MD. New York. MN. MD. USA Kenneth D. TN. Department of Electrical Engineering and Computer Science (EECS). CA. PhD. MSc. MD. Nashville. Intermountain Healthcare. Boston. MD. PhD. TN. . Sacramento. Stanford University School of Medicine. Baltimore. FACMI. MD. PhD Department of Medical Informatics. MS. USA Douglas K. FACMI Department of Biomedical Informatics. Bethesda. Stanford. NY. MD. Kohane. MD. Stanford. MA. Nashville. AZ. MD. USA Roger G. Nashville. USA Randolph A. MPH. School of Medicine. FAIMBE Department of Organizational Systems and Adult Health. Boston Children’s Hospital. National Library of Medicine. Harvard Medical School. FACMI Office of the Director. FACMI Center for Biomedical Informatics Research. MA. PhD Institute of Medical Engineering and Science. TN. Murray. USA Blackford Middleton. PhD.

MS. Department of Preventive Medicine and Medical Social Sciences.O. PhD. FACMI NHII Advisors.L. Shah. Leeds. Columbus. Dartmouth College. Intermountain Healthcare. FACMI Division of Health and Biomedical Informatics. USA Lynn Harold Vogel. FACMI Center for Biomedical Informatics Regenstrief Institute. PhD. Shortliffe. Boston. Starren. Geisel School of Medicine. MD.xxvi Contributors Philip R. USA Edward H. USA Justin B. MBBS. NY. MS. USA David K. New York. VA. NC. USA Jeremy C. NH. Ridgewood. New York. Tenenbaum. MD. MACP. MD. Columbia University. PhD. D. FACMI Department of Biomedical Informatics. Mountain View. Arizona State University. Indianapolis. Rand Corporation. Weill Cornell Medical College. Silverstein. Evanston. PhD Duke Translational Medicine Institute. University of Leeds. USA Adam B. USA Harold C. Ross. Inc. Stanford. Rubin. Schleyer. FACMI Research Institute. USA Daniel L. NJ. FACMI David Druker Center for Health Systems Innovation. NorthShore University Health System. NY.ketabpezeshki. DMD. USA Jonathan C. USA Paul C. MB BS. Tang. NY. MD. and the New York Academy of Medicine. Sox. FACMI Departments of Biomedical Informatics. MACP Dartmouth Institute. USA Robert S. Rudin. Columbia University. New York. PhD. Wilcox. IL. PhD LH Vogel Consulting. PhD Health Unit. FACMI Departments of Radiology and Medicine. CA. OH. BS. Palo Alto Medical Foundation. PhD Department of Medicine. MD. IN Nigam H. Durham.Sc Public Health Informatics Institute/ The Task Force for Global Health. USA Titus K. Stanford University. Payne. USA David . Decatur. Vawdrey. Arlington. MA. IL. MD. Northwestern University Feinberg School of Medicine. MD. Stanford. Duke University. West Lebanon. Chicago. MS. PhD. FACMI Department of Biomedical Informatics. USA Jessica D.. Wyatt. SM. CA. PhD Department of Biomedical Informatics. FRCP. The Ohio State University Wexner Medical Center. PhD. FACMI Leeds Institute of Health Sciences. Yasnoff. USA (021) 66485438 66485457 www. UK William A. CA. GA. Stanford University. LLC.

Part I Recurrent Themes in Biomedical Informatics (021) 66485438 66485457 .

and biomedical research in coming years? • What do we mean by the terms biomedical 1. informa. remains unchanged in each edition of the book. society was told that • Why should health professionals. Shortliffe. Within and informatics applications? the next decade. and students of the health professions tinely as memory devices. Biomedical Informatics. 272 W 107th St #5B. J. molecular biology. bioinformatics. Blois coauthored the 1990 (1st edition) version of this chapter shortly before his death in 1988. a year tion science. see his classic volume (Blois 1984) and and The New York Academy of Medicine.ketabpezeshki. and health informatics? computers in the 1940s. these new machines would soon be serving rou- tists. 3 DOI 10. USA leadership/acmi-fellow/marsden. PhD ter. New York 10025. learn more about this important early leader in the field Weill Cornell Medical College. you should know the • How does information in clinical medicine answers to these questions: and health differ from information in the basic • Why is information and knowledge manage. Shortliffe.). Although the chapter has evolved in subsequent editions. the promotion of health. decision science. physicians and other health • How has the development of modern comput.1007/978-1-4471-4474-8_1. assisting with calcu- learn about biomedical informatics concepts lations and with information retrieval. life . Blois as a coauthor because of his seminal contributions to the field as well as to this chap- E. a tribute to him at http://www. and how might we expect them to ical practice? affect the practice of medicine. medi. clinical informatics. public (Accessed 3/3/2013). Biomedical Informatics: The Science and the Pragmatics 1 Edward H. since it is time- Departments of Biomedical Informatics less. sciences? ment a central issue in biomedical research • How can changes in computer technology and and clinical practice? the way patient care is financed influence the • What are integrated information management integration of biomedical computing into clin- environments.amia. we con- † Author was deceased at the time of publication.J.5 was written by him and. E. © Springer-Verlag London 2014 (021) 66485438 66485457 www. NY. public health After scientists had developed the first digital informatics.H. Shortliffe and Marsden S. Comes to Medicine cal computing. of informatics. professionals had begun to hear about the dra- ing technologies and the Internet changed the matic effects that such technology would have nature of biomedical computing? • How is biomedical informatics related to clinical practice.H. tinue to name Dr. MD.1 The Information Revolution informatics. medical computer science. Blois† After reading this chapter.s-blois-md-facmi e-mail: ted@shortliffe. To at Columbia University and Arizona State University. biomedical engineering. nursing informatics. Section 1. Cimino (eds. and computer science? prior to the completion of the full manuscript.

but comprehensive clinical cal advances of the last three decades—personal data are now also important for institutional self- computers and graphical interfaces. greatest influence is yet to come. new methods analysis and strategic planning. 12. today’s children show an uncanny ability to make What might that future hold for the typi- use of computers (including their increasingly cal practicing clinician? As we shall discuss in mobile versions) as routine tools for study and detail in Chap. which computational tools assist not only with combined with equally pervasive and revolution. no applied clinical comput- entertainment. the for human-computer interaction. environments of the coming decades requires a ing the “information revolution” and “big data” deep understanding of the role that information fill our newspapers and popular magazines.1 Integrated Access to Clinical in the years ahead. Yet many observers cite the health care allow them to answer questions that are crucially system as being slow to understand information important for strategic planning. Blois on clinical practice. most health care institutions only since ~1980. and wireless communication are are single-entry points into a clinical world in even more recent. Furthermore. Nonetheless.S. and many of the original related and that planning for the new health care prophesies have come to pass. as 1970s. Information: The Future When one considers the penetration of com. they do not have systems in place that effectively nectivity. and are being gradually Health care organizations have recognized that supplanted by mobile devices with wireless con. slow to incor. mobile devices. and in some cases supporting (021) 66485438 66485457 www. In the past. Yet many observers remarkable progress in computing have followed now believe that the two topics are inextricably those early predictions. and more—have information is one of the principal barriers that all combined to make the routine use of comput. This book will teach you both about our present resources and accomplishments and about what you can expect . More than six decades of deal with both issues at once. social media. This dizzying rate of change. Shortliffe and M. mation by clinicians. Stories regard. the World Wide Web dates are seeking to develop integrated computer-based only to the early 1990s. and for reporting to regu- and slow to understand its strategic importance latory agencies. local area networking has been available is discussed later). and smart phones.4 E. patient-care matters (reporting results of tests. the Internet. is the issue of electronic health records (EHRs). social information-management environments.ketabpezeshki. for such planning.1. These networking. makes it difficult for public-health plan. clinicians encounter when trying to increase their ers by all health workers and biomedical scientists efficiency in order to meet productivity goals for inevitable. A new world is already with us. it is remarkable that the first personal Encouraged by health information technology computers were introduced as recently as the late (HIT) vendors (and by the US government. facilitating access to tran- ners and health-institutional managers to try to scribed reports. petitive environment. patient offices for years. for their better technology. provider groups in their local or regional com- porate it effectively into the work environment. but its their practices. toring devices and tools. accepted (Dick and Steen 1991 (Revised 1997)). financial data were the major elements required mitment. the enormous technologi. slow to exploit it for its unique prac. Similarly. and technology is likely to play in those environments. ary changes in almost all international health care allowing direct entry of orders or patient infor- systems. innovations inefficiencies and frustrations associated with the in mass storage of data (both locally and in the use of paper-based medical records are now well “cloud”). understanding of how they compare with other tical and strategic functionalities. Is Now puters and communication into our daily lives today. wireless especially when inadequate access to clinical communications. administrative and and its resulting need for investment and com. clinical workstations ing topic is gaining more attention currently than have been available on hospital wards and in out.H. personal health moni.

Implementing electronic the nineteenth century as a highly personalized records is inherently a systems-integration task.g. elec. supported by ing the needs of modern medicine.1. supporting clinical trials. and of EHRs. including elements of process results to fill up the record’s pages. the to spreadsheets and document-management soft. analyzes the processes associated with the cre- scholarly information (e. which implies that the institutions that . Thus. most easily achieved when the computing envi- tronic clinical record.1 Biomedical Informatics: The Science and the Pragmatics 5 telemedicine applications or decision-support to ask the following questions: “What is a health functions) but also administrative and financial record in the modern world? Are the available topics (e. the clinician. To and problem solving.. and integrated with other seeks to find a specific piece of information that types of useful information to assist in planning could occur almost anywhere within the chart. providing access For example. across a wide variety of systems and resources. 1. and even office automation (e. It arose in technology (Fig. be useful. routine use of computers.. This observation forces us ronment offers a critical mass of functionality (021) 66485438 66485457 www. Rather than becoming of paper charts to serve the best interests of the “power users” of a narrowly defined software patient.3). accessing digital ation and use of such records rather than think- libraries. is (and expensive) to move to a paperless. 1. managing materials and inventory. and few data or test mentation process. Today the inability and Shortliffe 1990). EHR requires the integration of processes for ware).1). secure. clinical-care records is best appreciated if one and implementing various treatment protocols). and managing the payroll). acceptable to cally—often a severe limitation when a clinician clinicians and patients.. organization must put into place. moved around as needed within the institution. however. confidential. it “lab notebook” that clinicians could use to record is not possible to buy a medical record system for their observations and plans so that they could a complex organization as an off-the-shelf prod- be reminded of pertinent details when they next uct. products and systems well matched with the tal.g. to analyze them. analyzing the outcomes associ. tracking of patients within the hospi. the EHR is best viewed not as an object. and ing of the record as a physical object that can be providing access to drug information databases). The complexity associated with automating ing quality assurance.2 Moving Beyond the Paper with secondary users of the record who are not Record involved in direct patient care (Fig. 12 and 14).ketabpezeshki. and to share them among colleagues and 1. the record system must make it easy to access and display needed data. 1. The record re-engineering and cultural change that are inevi- that met the needs of clinicians a century ago tably involved.2). Joint development and local adaptation are saw the same patient.. or a The traditional paper-based medical record is product. and the health system has package.g. Most organizations have found it challenging Thus.g. supporting modern notions of a comprehensive health personnel functions. struggled mightily to adjust over the decades and Experience has shown that clinicians are “hori- to accommodate to new requirements as health zontal” users of information technology (Greenes care and medicine changed. no assumptions that the record purchase such systems must have local expertise would be used to support communication among that can oversee and facilitate an effective imple- varied providers of care. There were no regulatory crucial. The contents of the paper record electronic health record that is intended to be have traditionally been organized chronologi- accessible. supporting bibliographic search. on the input side (Fig. they tend to seek broad functionality become clear (see Chaps. is that at the heart data capture and for merging information from of the evolving integrated environments lies an diverse sources. The key idea. users as well as the health systems themselves?” ated with treatments and procedures. but rather as a set of processes that an now recognized as woefully inadequate for meet. record? Do they meet the needs of individual research (e. perform.

4). pointing out to a (021) 66485438 66485457 www. advantages accrue as well.1 Inputs to the clinical-care record. Other ers were constrained in the past by clumsy meth.ketabpezeshki. 2 and 12 and in and added to the high costs associated with ran- the now classic Institute of Medicine’s report on domized prospective research protocols. Most obviously. Shortliffe and M. and data obtained directly paper record is created by a variety of organizational pro. Medical . generally organized in chronological regarding direct encounters between health professionals order and patients. fraught with opportunities for error. from patients). generally relying on manual capture of ance with a research protocol. records are summarized in Chaps. The record thus becomes a merged collec- cesses that capture varying types of information (notes tion of such data. routine clinical record keeping (Fig. The approach was labor- The arguments for automating clinical-care intensive. analysis (Fig. that warrants emphasis is the importance of the 26).5). thus making much of what is required for and to gain insight into disease processes that are research data collection simply a by-product of not otherwise well understood. The traditional telephone calls or prescriptions. One argument to those carrying out clinical research (see Chap. For example. computer-based patient records (CPRs) (Dick The use of EHRs has offered many advantages and Steen 1991 (Revised 1997)).S.6 E. it helps to eliminate the EHR in supporting clinical trials—experiments manual task of extracting data from charts or fill- in which data from specific patient interactions ing out specialized datasheets. reports of that makes the system both smoothly integrated information onto datasheets that were later with workflow and useful for essentially every transcribed into computer databases for statistical patient encounter. The data needed are pooled and analyzed in order to learn about for a study can often be derived directly from the the safety and efficacy of new treatments or tests EHR. 1. Blois Fig. 1. the ods for acquiring the data needed for clinical record environment can help to ensure compli- trials. laboratory or radiologic results.H. 1.

guidelines. it a patient’s care. as well as individual provider when the protocol for a study calls for a specific groups. (021) 66485438 66485457 www.ketabpezeshki. Several government and professional advice available in the routine clinical setting. analyzing. generally in an ners. clinical guidelines. Computer-based tools for implementing effort to reduce practice variability and to develop such guidelines. We are also seeing the evidence from the literature. we need better methods for delivering the deci- senting patient descriptors.2 Outputs from the clinical-care record. there is a growing recognition that ible with the local EHR’s conventions for repre. present a means for making high-quality . Despite clinical trial protocols that can help to ensure that the success in creating such evidence-based the data elements needed for the trial are compat. and sharing information from such professionals and the patients themselves but also a wide records results from a set of processes that often varies variety of “secondary users” (represented here by the indi. Numerous providers are typically involved in information is collected in the traditional paper chart.1 Biomedical Informatics: The Science and the Pragmatics 7 Fig. unavailable when the knowledge of standard order sets. Once patient care. often putting an emphasis on using clear data about that patient. 1. so the chart also serves as a means for may be provided to a wide variety of potential users of the communicating among them. have invested heavily in guideline devel- management plan given the currently available opment. and integrating them with the consensus approaches to recurring management EHR. The mechanisms for dis- information that it contains. rather than expert development of novel authoring environments for opinion alone. as the basis for the advice. and they contain would be most valuable to practitio- clinical pathways (see Chap. 22). These users include health playing. substantially across several patient-care settings and viduals in business suits) who have valid reasons for institutions accessing the record but who are not involved with direct clinician when a patient is eligible for a study or organizations. sion logic to the point of care. Guidelines that Another theme in the changing world of health appear in monographs or journal articles tend to care is the increasing investment in the creation sit on shelves.

Shortliffe and M.8 E. The first of these issues 1. and offers solutions for the ways consistently constrained our efforts to build in which the EHR can be better joined with other effective EHRs: (1) the need for standards in the relevant information resources and clinical pro- area of clinical terminology. (Accessed 4/21/13/). As that information to those who have valid reasons for shown in Figs 1.H.S. especially within communities where ing data privacy.3 Anticipating the Future of is discussed in detail in Chap.forbes. 7. (2) concerns . object or process from the physical world. confidentiality.3 Complex processes demanded of the record. 2013). (3) patients may have records with multiple provid- challenges in data entry by physicians. which in meeting the diverse requirements for data collection both gather information to be shared and then distribute and information access that are implied by this diagram Many organizations are accordingly attempting one of the central topics in Chap. Chapter 13 examines the fourth decision support to practitioners.1 and 1. Paper-based documents are severely limited tion of a complex set of organizational processes. Issues of to integrate decision-support tools with their direct data entry by clinicians are discussed in EHR systems. the clinical chart is the incarna. and there are highly visible efforts Chaps. and privacy is Electronic Health Records 1 One of the first instincts of software devel- http://www. Blois Fig. accessing it. opers is to create an electronic version of an healthcare/.ketabpezeshki. focusing on recent trends in networked There are at least four major issues that have data integration. difficulties associated with the integration of record systems with other information resources in the health care setting. Some (021) 66485438 66485457 www.1. and (4) ers and health care systems (Yasnoff et al. 10.1 topic.2. 2 and 12 and throughout many other underway to provide computer-based diagnostic chapters as well. 1. cesses. and ibms-watson-gets-its-first-piece-of-business-in.

or a secondary research the clinicians of the rules that are defined by the research database may be created by downloading information protocol. The manual processes in mental plan Fig. The trials are generally systems for data storage and analysis. Physicians can be clinical trials can become a by-product of the routine care reminded when their patients are eligible for an experi- of the patients. collecting those data in a structured format to be left to tants.5 Role of electronic health records (EHRs) in sup. 1. Electronic Health Clinical Data Record (EHR) Repository Clinical trial design Clinical trial •Definition of data elements database •Definition of eligibility •Process descriptions Analyses •Stopping criteria •Other details of the trial Results Fig.1 Biomedical Informatics: The Science and the Pragmatics 9 Medical record Computer database Clinical trial design •Definition of data elements Data sheets •Definition of eligibility •Process descriptions •Stopping criteria Analyses •Other details of the trial Results Fig. manual processes at the point of patient care sheets for later transcription into computer databases. the (021) 66485438 66485457 www. the collection of much of the research data for ity and efficiency of the clinical trial. Alternatively. 1. interaction of the physician with the EHR permits two- porting clinical trials. way communication. Physicians who methods for analysis. or their research assis. With the introduction of EHR sys. thereby increasing compliance with the experi- from the online patient . which can greatly improve the qual- tems. and the computer system can also remind from the clinical data repository. Research data may be analyzed directly mental protocol.ketabpezeshki. the gathering of designed to define data elements that are required and the research data is still often a manual task. In addition. data managers have been hired to abstract Although modern clinical trials routinely use computer the relevant data from the chart. but it is common for the process of care for patients enrolled in trials. have traditionally been asked to fill out special data. 1.4 are thereby largely eliminated.4 Traditional data collection for clinical trials.

human ingenuity and be made about the importance of committing to creativity often lead to an evolution that extends the use of EHRs today. Technology and Health ence between today’s office automation software Data Integration and the typewriter.2 Communications Consider. and has pro- use the early airplanes for travel. But 1930s mercial organizations to join the network as well. for example. and machine-to-machine electronic that were never contemplated when the industry mail exchanges quickly became a major compo- depended on human bank tellers. integration of figures. National Science Foundation took over the task nated most of our health care environments. began to be connected to what 1930s.10 E. which was the original inspi- ration for the development of “word processors”. accessible ized in the decades ahead. network began as a novel mechanism for allow- tion. The computer can thus facil- itate paradigm shifts in how we think about such familiar concepts. the quality and vided the invisible but mandatory infrastructure (021) 66485438 66485457 www. 2000). Note of running the principal high-speed backbone that the state of today’s EHR is roughly compa. con. network in the United States. no much better than both paper records and the early longer depending on the U. numbers that we once created on graph paper.ketabpezeshki. the facilities. today’s spreadsheet pro. the Internet is ubiquitous. grammar aids. The notion of electronic mail arose soon facilitation of today’s worldwide banking in ways thereafter. To to share data files with each other and to provide take an example from the financial world. albeit had become a fully commercialized operation. spelling correc. puters. air travel seems archaic by modern standards. and the antiquated paper folders that until recently domi. government to sup- computer-based systems of the 1960s and 1970s. papers each time a minor change was made to a The Internet began in 1968 as a U. nent of the network’s traffic. A similar point can been developed. the Internet in the United States it is logical to assume that today’s EHRs. to the network (Shortliffe 1998a. Hospitals. Initially known as the ARPANET.S. Similarly. and by 1973 the first medi- implemented on computer systems and new cally related research computer had been added opportunities for its enhancement become increas. and in stantially from the days of the Wright Brothers.H. the remarkable differ. “publishing” on the . and By April 1995. An obvious opportunity for changing the role and Although the early word processors were functionality of clinical-care records in the digi- designed largely to allow users to avoid retyping tal age is the power and ubiquity of the Internet. It is clear that EHRs a decade During the 1980s. If people had failed to through mobile wireless devices. research document. will be greatly improved and further modern. Blois familiar notion provides the inspiration for a new efficiency of airplanes and air travel would not software product. even though we know that the software version far beyond what was ini. Projects Agency (ARPA) of the Department of Consider all the powerful desktop-publishing Defense. a major policy move it was decided to allow com- and air travel was becoming common. the document-management software activity funded by the Advanced Research of today bears little resemblance to a typewriter. tially contemplated.S. etc. the technology began to be from now will be remarkably different from the developed in other parts of the world. Once the software version has have improved as they have. Shortliffe and M. remote access to computing power at other loca- sider automatic teller machines (ATMs) and their tions. they need to be much better in the future. located mostly at academic institutions or grams bear little resemblance to the tables of in the research facilities of military contractors. Today. port even the major backbone connections. use ing a handful of defense-related mainframe com- of color. By that time air travel had progressed sub. 1.S. had by then become known as the Internet. mostly rable to the status of commercial aviation in the academic centers. ingly clear to us. As the technology It is accordingly logical to ask what the health matured. however. its value for nonmilitary research activi- record will become after it has been effectively ties was recognized.

modern technical issues and policy directions. nizations (see Chap. lenge. small offices. secure connectivity with the Internet. legislation was in their practices both to assist in patient care and passed in 1996 to allow new competition to to provide patients with counsel on illness pre- develop and new industries to emerge. We have vention. scientific. and inexpen. The impact on everyone has been great appeared in (Shortliffe and Sondik 2004). with companies infrastructure that will allow all clinicians. is to find a way to integrate data sive mechanisms for connecting to the Internet from such diverse practice settings. The societal impact of this communications phenomenon cannot be overstated. consider for bringing timely information to the desktop one example of how disease surveillance.. etc. especially given the international connectivity that has 1. requires no special training. Navigating the Web is highly intui. networking. There has in turn been a major upheaval in the The goal is to create an information-management telecommunications industry. had been left to the pri. 1. The full impact of this use of electronic subsequently seen the merging of technologies resources will occur when data from all such such as cable television. Many people point to the seek health-related information and it is also Internet as a superb example of the facilitating enhancing how patients can gain access to their role of federal investment in promoting innova.g. health care providers and to their clinical data.1 A Model of Integrated Disease grown phenomenally in the past two decades.1 Biomedical Informatics: The Science and the Pragmatics 11 for social. 14). and provides health. when the World integrated information resources that combine Wide Web (which had been conceived initially clinical and health data from multiple institutions by the physics community as a way of using the within regions.. systems have come to appreciate the importance ated if the research and development. and administrative systems within their orga- vate sector. regard- that used to be in different businesses (e. preven- machines and mobile devices of individuals with tion. ranging To emphasize the role that the nation’s network- from consumers to scientists to those interested ing infrastructure is playing in integrating clini- in political issues. Internet services. financial. mediated through into homes and offices are widely available. multispecialty groups. using cell phones or set-top boxes) have also 2 This section is adapted from a discussion that originally emerged. High-speed lines veillance databases (Fig. Surveillance2 Countries that once were isolated from infor- mation that was important to citizens. The Internet is a major societal Just as individual hospitals and health care force that arguably would never have been cre. 13 and 16). As you will see. records are pooled in regional and national sur- and satellite communications. cable less of practice setting (hospitals. are now finding new options cal data and enhancing care delivery. health planners and The explosive growth of the Internet did governments now appreciate the need to develop not occur until the late 1990s. community clinics. every chapter in this book. In the United States. plus the of integrating information from multiple clinical coordinating activities. political.) to use EHRs merged. (021) 66485438 66485457 www. telephone. The chal- wireless networking is ubiquitous. we introduce it in the tion that accounts for its remarkable growth as a following sections because of its importance to worldwide phenomenon.2.6). the Internet diagrams among researchers) was introduced and and the role of digital communications has there- popularized. fore become a major part of modern medicine and tive. by information and communications technology. and ultimately nationally (see Internet to share preprints with photographs and Chaps. emergency television.ketabpezeshki. and and hence it is affecting the way that individuals entertainment ventures. especially without using conventional computers (e. and care are increasingly being influenced an Internet connection. Although this topic recurs in essentially a mechanism for access to multimedia informa. military finding that their activities and technologies have bases.g. of course. tive technologies. and telephone) now .

transmission of clinical information occur The pooling and integration of data requires only if those data are encrypted. in security concerns as discussed in the text). (021) 66485438 66485457 www. over time from such information. Interestingly. The de The practical need to pool and integrate clini. • Quality control and error checking: Any sys- • HIPAA-compliant policies: The privacy and tem for accumulating. and a variety of metrics can be derived discussing here. research use. and utilizing security rules that resulted from the 1996 clinical data from diverse sources must be Health Insurance Portability and complemented by a rigorous approach to qual- Accountability Act (HIPAA) do not prohibit ity control and error checking. facto standard for such sharing. is beginning to be uni- addressed in achieving such functionality and formly adopted. HL7. guide- of the solution in achieving the vision we are lines. and utilized resources.12 E. Blois Internet Provider EHR Provider EHR Regional and National Surveillance Databases Provider EHR Provider EHR Protocols and Guidelines for Standards of Care Provider EHR Different Vendors Fig.H. political. but they do lay down policy rules and hensiveness of the data that are collected in technical security practices that must be part such repositories.6 A future vision of surveillance databases. Shortliffe and M. emphasizes the practical issues that need to be after years of work. Sharing data over networks requires that all ing to provide value-added capabilities that will developers of EHRs and clinical databases excite and attract the practitioners for whom their adopt a single set of standards for communi- EHR product is intended. It is crucial that the pooling and use of such data (see Chap. and financial rather than • Standards for data definitions: A uniform technical in nature: “envelope” for digital communication. . such as • Encryption of data: Concerns regarding pri. users have faith in the accuracy and compre- 10). cating and exchanging information. does not assure that the contents of such vacy and data protection require that Internet messages will be understood or standardized. because policies. implemented. Health Level cal data from such diverse resources and systems 7 (HL7). pooled. the adoption of standards for clinical termi- lished mechanism for identifying and authen. When informa- which clinical data are pooled in regional and national tion is effectively gathered. analyzing. and analyzed. compet. was introduced decades ago and. most of the barriers are (see Chap. nology and potentially for the schemas used to ticating individuals before they are allowed to store clinical information in databases (see decrypt the information for surveillance or Chap.S. there repositories through a process of data submission that are significant opportunities for feeding back the results occurs over the Internet (with attention to privacy and of derived insights to practitioners at the point of care since there are multiple vendors and system • Standards for data transmission and sharing: developers active in the marketplace. 7).ketabpezeshki. logistical. 7). with an estab.

observations and to gain access to informa- which can be delivered to practitioners through tion about the . Thus one should envision a records is then forwarded automatically to day when clinicians. the diagram. employability.8).7).6. or clinical research protocols.1 (Fig. Beginning at the left of Furthermore. order-entry systems that government program in Sect. an expanded view of want to have taken for the treatment or manage. Although this view is a powerful con- liberties. as is shown in Fig. Sect. interpreta- of state and federal governments will need to tion.2. organization. for payment of incentives to clinicians or hospitals cians use in their practice settings. and the political issues addressed the small cycle that is implied: patient-specific (including the concerns of some members of data and plans entered into an EHR and subse- the populace that any government role in man. it fails to include a larger view of the bases. such straightfor- to providers to enhance their decision making at ward use of EHRs for direct patient care does not the point of care (Fig. quently made available to the same practitioner or aging or analyzing their health data may have others who are involved in that patient’s care societal repercussions that threaten individual (Fig. will receive integrated. In fact. summary information can flow back in clinical-care records. 1. funding.6). clinicians caring for patients use bases can help to support the creation of evidence. and the like). adapted for execution Patients and integration into patient-specific decision Access Patient support rather than simply provided as text Information documents Provider’s • Opportunities for distributed (community.2 The Goal: A Learning Health Any adoption of the model in Fig.ketabpezeshki. You can easily understand be clarified. either in their community or more widely • Trends and patterns of public health Providers importance Caring for • Clinical guidelines.3). Information from these the feedback process. and ultimate use. These may be who implement EHRs (see the discussion of this EHRs or. increasingly. instead. the data. and a robust system of Internet integration societal value of the information that is contained with EHRs. We have been stressing the cyclical role of bases that are involved (see Chap. clinicians use to specify the actions that they Consider. 1. Knowlege and based) clinical research.2. 1. 1.6 will Care System require mechanisms for creating. electronic health records. 1. of patients. meet some of the requirements that the US govern- dards that allow such information to be integrated ment has specified when determining eligibility into the vendor-supplied products that the clini. at the point of care. 12 and 14). and maintaining the regional and national data. 1. the health surveillance model introduced in ment of their patients (see Chaps. non-dogmatic. tributor to improved data management in the care With the establishment of surveillance data. The role information—its capture. both to record their based guidelines. whereby patients Advice from Others are enrolled in clinical trials and protocol guidelines are in turn integrated with the cli- Fig.7 There is a limited view of the role of EHRs that nicians’ EHR to support protocol-compliant sees them as intended largely to support the ongoing care management of enrolled patients of the patient whose clinical data are stored in the record (021) 66485438 66485457 www. 13). 1. This assumes stan.1 Biomedical Informatics: The Science and the Pragmatics 13 • Regional and national surveillance databases: 1. supportive information regarding: Record Electronic • Recommended steps for health promotion and Patient Health disease prevention Information Records • Detection of syndromes or problems. 1.

(021) 66485438 66485457 www. driven by experience. building digital data excellent. The analyzed information from reg. physicians from what we do. 1. tive studies (see Chap. and Order-Entry A ‘’Learning and Systems Healthcare Educational Materials System’’ Fig. as well as healthy individuals. The standards for treatment in turn turning to the Internet for health information. has registries and research databases. As a result.3 Implications of the Internet istries and research studies can in turn be used to for Patients develop standards for prevention and treatment. ongoing subject of study by the Institute of aspx (Accessed 3/3/2013). in this sense physicians can often learn resources using EHRs. Blois Biomedical and Clinical Resarch Electronic Health Regional Records and National Public Providers Health and Caring for Disease Patients Registries Standards for Prevention Creation of and Information. This cycle of new knowl- records (EHRs) are routinely and effortlessly submitted to edge. and fed back to clinicians.H. 2012). 2011. The companies that provide search it is integrated seamlessly with EHRs and order. The resulting new been dubbed a “learning health care system” knowledge then can feed back to practitioners at the point regional and national registries as well as to Medicine. The concept has been dubbed a learning health care system and is an 3 http://www. Shortliffe and M. that arose due to medically related searches that the information informs patient care. where on the net. guidelines. Protocols.14 E. It is can be translated into protocols. Guidelines. it is not surprising that increasing numbers from the health records or from the pooled data of patients. engines for the Internet report that health-related entry systems.2. 11) or formal institu- tional or community-based clinical trials (see Chap. are in registries. 1.iom. sites are among the most popular ones being This notion of a system that allows us to learn explored by consumers. with major guidance from biomedical research. Some of the information is timely and clinicians and institutions. or a stack of print- ered over the network back to the clinicians so outs. is gaining wide attention now that and bring with them when they seek care from we can envision an interconnected community of clinicians. of care.3 which has published a series of research databases that can support retrospec. unlocking the experience that and other care providers must be prepared to deal has traditionally been stored in unusable form in with information that patients discover on the net paper . This new knowledge and encountered a patient who comes to an appoint- decision-support functionality can then be deliv. 26). and a rare North American physician who has not educational materials.8 The ultimate goal is to create a cycle of informa. reports on the topic (IOM 2007. As the penetration of the Internet continues to Researchers can draw information either directly grow. Treatment Decision-Support.S. ment armed with a question. using a variety of computer-supported decision- tion whereby data from distributed electronic health support delivery mechanisms.ketabpezeshki.

and there are specialized settings sections of biomedicine. decision science. Many of the trainees are life limited for much penetration of the idea beyond science researchers. however. 1. physicians. cognitive and cost-effective (e. workers (Accessed 3/3/2013).9). and we now take this kind of inter. the demand for such individuals far outstrips the supply. (021) 66485438 66485457 www. survey-strong-demand-health-information-technology- tions for the future of health care delivery in gen. (Shortliffe 2010). security. There is a difference between computer literacy People who lack medical training can be misled (familiarity with computers and their routine uses by such information. workers.4) that provide custom-tailored educa- tury (see Fig. As medicine.1 Biomedical Informatics: The Science and the Pragmatics 15 about innovations from their patients and will availability of the Internet.4. nurses. tion to increasing speed. Fortunately. and video-based care of has been clear for over two decades (Greenes and patients in prisons). the new communications (Accessed 4/21/13).2. Shortliffe 1990). there are many other need to be increasingly open to the kinds of Internet technology all have significant implica. both for aca- demic and industrial career pathways. and should play in our health care system. This notion of “medicine become known as biomedical informatics (see at a distance” arose early in the twentieth cen. some generally doing a poor job of training future cli- sites provide personalized advice. We are ments from anecdotal sources. the notion of telemedicine creation of formal training programs in what has emerges (see Chap. sometimes for nicians in the latter area and are thereby leaving a fee.2. In addition. 18). we have begun to see the nication about/press/press/803-ehealth-initiative-survey-reveals- eral and of EHRs and their integration in high-demand-for-health-information-technology-workers particular (Shortliffe 1998b.g. deep understanding of the biomedical milieu and action with patients for granted. adequate.1 Education and Training the Web lacks peer review or is purely anecdotal. In a positive light. and communications technologies. with all the attendant concerns about the them poorly equipped for the challenges and quality of the suggestions and the ability to give opportunities they will face in the rapidly chang- valid advice based on an electronic mail or Web. Computer science training alone is not typically relying on the Internet as our commu. and other health professionals who see the years.5 We need 1. 1. rural care. On the other hand. but the technology was too tional opportunities. international science. (Accessed 9/11/2013). 2000). and 5 http://www. pharma- telephone conversations until the last 30–40 cists. ing practice environments that surround them based interaction. information science.ketabpezeshki. much of the health information on 1. But in addi. just as they have been in our society) and knowledge of the role that poorly served in the past by printed information computing and communications technology can in books and magazines dealing with fad treat. Years ago medicine adopted the individuals who not only comprehend computing telephone as a standard vehicle for facilitating and communications technology but also have a patient care.4 Requirements for Achieving the Vision Efforts that continue to push the state of the art in 4 http://www..4. The use of telemedicine has subsequently career opportunities and challenges at the inter- grown rapidly.ehidc. in which it is already proving to be successful computer science. teleradiology. Sect. much of the future vision we technologies offer clinicians creative ways to have proposed here can be achieved only if edu- interact with their patients and to provide higher cational institutions produce a cadre of talented quality care. areas that need attention if the vision of a learn- tions that this enhanced access to information ing health care system is to be achieved. If we extend the of the needs of practitioners and other health audio channel to include our visual sense as well. will generate from patients in their practices. reliability.

S. there needs to be a a new institution without major analysis. and cooperative joint-development efforts. with technical leadership and planning. of software implementation. This 1924 example is from the cover of a popular magazine and envisions video enhancements to radio (Source: “Radio News” 1924) more training programs.amia. as implied above. rede- greater understanding among health care 1.9 “The Radio Doctor”: long before television was invented.6 expansion of those that ers regarding the role of specialized multi- already exist. plus support for junior faculty in disciplinary expertise in successful clinical health science schools who may wish to pursue systems implementation. Shortliffe and M. and it is simplistic to assume that off-the- Management Change shelf products will be smoothly introduced into Second. account (021) 66485438 66485457 www. . Blois Fig. provides some of the most complex organizational structures in society (Begun and Zimmerman 1.2 Organizational and 2003).H. creative observers were suggesting how doctors and patients could communicate using advanced technologies. The health care system additional training in this area. as well as problems courses (Accessed 3/3/2013). Underinvestment and a failure to understand the 6 requirements for process reengineering as part A directory of some existing training programs is available at http://www.2.16 E.ketabpezeshki.4.

It was accordingly remarkable institutions. and a learning health nomic “Stimulus Bill”. funding to provide fiscal incentives for health Second. and data sharing implied by a learn. made available. and 27). in the President’s State of the Union address however.ketabpezeshki. and public health. that first provided major care system. although our presentation of the systems. in 2004 (and in each of the following years of his the cyclical creation of new knowledge in a administration). which must be considered in the this volume. it was required for successful EHR implementation the American Recovery and Reinvestment Act are sensitive to the need for data integration. The topic of EHRs subsequently Chap. The criteria that are with strong bipartisan support. 27. when. There is also a discussion of HIT notion of patient-centered health care that is now policy and the federal government in Chap. larly supportive and. direction (see Sect. and providers to implement learning health care notion has focused on the EHRs in their practices.3 The US Government Steps In health care organizations report in their efforts to use computers more effectively in support of During the early decades of the evolution of clinical patient care and provider productivity. or administrator can EHRs—the Office of the National Coordinator mandate the standards for connectivity. Tommy Thompson. No individual intended to support the expansion of the use of system developer. although for computing and communications resources the organization served as a convening body for within and among institutions and clinics is EHR-related planning efforts and the National required before practitioners will have routine Health Information Infrastructure (see Chaps. referred to by the full acronym ONCHIT. the major role of govern- described previously is meant to motivate your ment was in supporting the research enterprise as enthusiasm for what lies ahead and to suggest new methods were developed. and national coordinating bodies work Human Services. only when eligible orga- other workers in the field have similar needs that nizations or individual practitioners implemented can be addressed in similar ways. in May 2004. There is also the and meaningful use criteria in many chapters in patient’s view. infrastructure. patient The notion of a learning health care system care. Before embarking on these topics. created an entity and resources that are necessary. The Secretary of Health and centers. You will see references to such certification access to data and information. but later ing health care system.1 Biomedical Informatics: The Science and the Pragmatics 17 for many of the frustrating experiences that 1. 13). The topic was seldom mentioned by the such as this one. Essentially all of the follow. access to the information that they need (see 12. still ongoing at present. information systems for use in hospitals. First. tested. The academic EHRs that were “certified” as meeting minimal research community has already developed and standards and when they could document that made use of much of the technology that needs to they were making “meaningful use” of those sys- be coalesced if the clinical user is to have similar tems. However. broadly accepted and encouraged (Ozkaynak Although the process of EHR implementation is et al. however. data for Health Information Technology (initially pooling. (ARRA) in early 2009.3). A recent federal incentive program became a talking point for both major candi- for EHR implementation is a first step in this dates during the Presidential election in 2008. hospitals. the trend is clear: because (021) 66485438 66485457 www. 13. A national initiative shortened simply to ONC). also known as the eco- public-health support. Such payments were clinician’s view of integrated information access. academic medical within 10 years. however. President Bush called for univer- learning health care system will become reality sal implementation of electronic health records only if individual hospitals. we must emphasize two points. . vendor. and formally the topics that need to be addressed in a book evaluated. of cooperative planning and implementation There was limited budget for ONC. nation’s leaders. was simi- together to provide the standards. even during the 1990s ing chapters touch on some aspect of the vision when the White House was viewed as being espe- of integrated systems that extend beyond single cially tech savvy. 2013).

tionalities that need to be brought together in the ing in EHRs and incorporating them into their integrated bio medical-computing environment practices. The remainder of this chapter deals ers skilled in health information technology has with biomedical informatics as a field and with grown much more rapidly than has the general biomedical and health information as a subject of job market.healthit. it might suggest the maintenance of clinical-care records using com- 1. where many Reference to the use of computers in bio- observers had previously felt that progress had medicine evokes different images depending been unacceptably slow (Shortliffe 2005). Shortliffe and M. Many physicians background. let us now consider the scientific immediately think of office-practice tools for discipline that is the subject of this volume and tasks such as patient billing or appointment (021) 66485438 66485457 www.10 Percent change in online health IT job postings Brief. it might mean the use of com- puters for maintaining. relative to health care jobs and all jobs: normal. to a decision per month. ONC Data of the federal stimulus package.10) . default/files/pdf/0512_ONCDataBrief2_JobPostings. Furthermore.H. this book. 2. it might mean the Informatics and Related assistance by computers in disease diagnosis. Blois Health IT jobs Healthcare jobs All jobs Percent change in Health IT job Positings per Month (normalized to Feb 2009) 250 199 200 HITECH Act February 2009 150 100 57 50 52 0 –50 May-07 May-08 May-09 May-10 May-11 Sep-07 Sep-08 Sep-09 Sep-10 Sep-11 Nov-07 Nov-08 Nov-09 Nov-10 Nov-11 Mar-07 Mar-08 Mar-09 Mar-10 Mar-11 Jan-07 Jan-08 Jan-09 Jan-10 Jan-11 Jan-12 Jul-07 Jul-08 Jul-09 Jul-10 Jul-11 Fig. To a hospital administrator. study.4 Defining Biomedical puters.18 E. large numbers of has led to the development of many of the func- hospitals. 1. systems. and practitioners are invest. No. on the nature of one’s involvement in the field. the demand for work. May 2012 [http://www.S. It provides additional background needed It is a remarkable example of how government to understand many of the subsequent chapters in policy and investment can stimulate major transi.pdf ized to February 2009 when ARRA passed (Source: ONC (Accessed 4/10/13)] analysis of data from O’Reilly Job Data Mart. to Disciplines a basic scientist. tions in systems such as health care. 1. even within health care (Fig. of the future. and analyzing With the previous sections of this chapter as gene-sequencing information. retrieving.

and other scholars increasingly con. including all those subtopics to which computers might be 1. book and occurs in various forms in essentially In the 1970s we began to use the phrase med- every chapter. tion to it as such—at the expense of considering The term information science. people have been uncertain the scientific issues that underlie such efforts.4. Even . we might want to use computers to process it). planning. These investigations have been interest under the name cognitive science.1 Terminology applied. (3) the importance of evidence-based knowledge of biomedical and health topics. The cogni- decision-support tools that assist in applying the tive activities of clinicians in practice probably most current patient-care guidelines. Since the 1960s. we must understand more clearly can help us to tie together the diverse aspects of such human processes as diagnosis. therapy health care and its delivery. paper-based and electronically stored informa- sider such matters as the nature of information tion. and problem solving To achieve a unified perspective. and cognitive underpinnings of ence. and (4) the applications of computers or clinical practice had access to some kind of and communication methods in biomedicine and computer system. The fourth topic. decision making. The third topic. by which time a growing number including its derivation and proper management of individuals doing serious biomedical research and use. however. is intrinsic to this scientific content. Much of what information science origi- and knowledge and how human beings process nally set out to be is now drawing evolving such concepts. is the ical computer science to refer to the subdivision principal subject of this book. Unless you keep tive powers have motivated many of these stud- in mind these surrounding factors. (2) the structural features of medicine. and was only vaguely defined. originated in the field of library science and fields such as medicine. many others too. (021) 66485438 66485457 www. medicine has provided a rich area individuals have always lived in a world in which for computer science research. sionally used in conjunction with computer sci- concepts. ies. term computer science is used more as a matter knowledge to support effective decision making of convention than as an explanation of the field’s in support of human health. ideas. of computer science that applies the methods of Computers have captured the imagination the larger field to medical topics. Computer scientists. health. To develop computer-based tools to assist ficult to understand how biomedical computing with decisions. which is occa- what the computer can do given the numbers. people been derived from applied medical-computing may pay a disproportionate amount of atten. The field have received more attention over the past three includes study of all these activities and a great decades than in all previous history (see Chap. it includes 4). or the simple existence of the computer. to the broad research. Today’s younger see. psy.1 Biomedical Informatics: The Science and the Pragmatics 19 scheduling. the existence of the computer and the the consideration of various external factors that possibilities of its extending a clinician’s cogni- affect the biomedical setting. range of issues related to the management of both chologists. it may be dif. somewhat generally. As you will (and attention) of our society. More importantly. philosophers. the dents who wish to learn more. and several basic computers are ubiquitous and useful. Again. and we provide references name computer science was itself new in 1960 in the Suggested Readings section for those stu. Nurses often think of computer-based given a sense of timeliness (if not urgency) by tools for charting the care that they deliver. and biomedical is used to refer. Because computing insights and methodologies have the computer as a machine is exciting. however.ketabpezeshki. The and the next chapter. research. we might in medicine. what name they should use for the biomedical We mention the first two topics briefly in this application of computer science concepts. We must also understand how per- consider four related topics: (1) the concept of sonal and cultural beliefs affect the way in which biomedical information (why it is important in information is interpreted and decisions are ulti- biological research and clinical practice and why mately made.

as we mation theory have illuminated many processes will argue shortly. or health care informatics. As a result. Thus.bisti. sive and to embrace the biological applications and it does not capture the broader implications with which many medical informatics groups of the European They are often was no longer possible without computational associated with bioengineering applications support and analysis (see Chaps. In addition. and 7 academic settings.ketabpezeshki. expressed concern that the adjective “medical” is 8 Available at http://www. Those in the field were attracted report8 recommending that the NIH undertake by the word’s emphasis on information.20 E.nih. imply. organization called the Bioinformatics Working ing the 1980s. (Accessed 4/8/2013). unclear. lic health. Several academic groups United States. in an effort to be more inclu- however. pub- little effect on our understanding of human infor. it tends to focus the field’s in communications technology. and it gained momentum as a With the subsequent creation of another NIH term for the discipline. However. many observers http://www. tage of tending to exclude applications to bio- The results scientists have obtained with infor. group called the Working Group on Biomedical speaking community began to use the term medi. the use of informatics meth- are viewed more as tools for a bioengineering ods in such work had become widely known as application than as a primary focus of research. may be confused with library science. too focused on physicians and fails to appreciate html(Accessed 4/8/2013).ornl. Computing. bioinformatics and the director of the National In the 1970s. (021) 66485438 66485457 www. of computers. Because the term informatics became regarding the naming conventions we have been widely accepted in the United States only in the discussing.S. this is the name of the field have changed their names. had already been involved. The terms biomedical computing or biocom. industrial. and it is still some- times used in professional. gained physics of home. and prevention) rather than the basic mation processing. it has evolved into some popularity. in contrast. however. the visibility of informatics applications computing (it includes such topics as medical in biology was greatly enhanced. to the human genome project7 and the growing ing only that computers are employed for some recognition that modern life-science research purpose in biology or medicine. The term is broader than medical Group. inspired by the French term for Institutes of Health (NIH) appointed an advisory computer science (informatique). Blois Information theory. 24 and 25). record keeping. in which the devices By the late 1990s. the name medical cal informatics appeared by the late 1980s to informatics gradually gave way to biomedical have become the preferred term. the English. Applications of informatics methods in biol- putation have been used for a number of years. Shortliffe and M. Today bioinfor- statistics. the relevance of this discipline to other health and 9 See http://www. The explosive growth of the nature of the field to which computations are this field. in many universities and biotechnology compa- phasizes the computer while focusing instead on nies around the world. Indeed. but they have had name on application domains (clinical care. As a result.nih. was first life-science professionals. has added to the confusion applied. dur. computer itself. and the study of the matics is a major area of activity at the NIH9and nature of medical information itself) and deem. the term health developed by scientists concerned about the informatics. medical research (Chaps. however. which an initiative called the Biomedical Information they saw as more central to the field than the Science and Technology Initiative (BISTI). the name medi. even though it has the disadvan- what may be viewed as a branch of mathematics. 24 and 25) and. and a major medical that we used in the first two editions of this text. the relationship between late 1980s. especially in Europe.shtml (Accessed 4/8/2013). even in the informatics (BMI).H. medical information science was medical informatics and bioinformatics became also used earlier in North America. ogy and genetics exploded during the 1990s due They are nondescriptive and neutral. discipline and its broad range of . informatics journal (Computers and Biomedical book (from 1990 to 2000). the group provided a cal informatics. In June 1999. this

and the field and its definition. the investigates and supports reasoning.10 competencies that need to be acquired by stu- Despite this convoluted naming history. and decision making. that biomedical informatics has many other based methods. (021) 66485438 66485457 www.journals. The mation. . When acknowledges that the emergence of biomedical we speak specifically about computers and their informatics as a new discipline is due in large use within biomedical informatics activities. The resulting definition. and knowledge for scientific inquiry. information science. and biomedical research. beginning with the third organization. and to a growing conviction that component sciences in addition to computer sci. It clinical practice. a simple sentence and then adds four clarify- medical informatics for this purpose. translation across the spectrum from mole- inition. the process of informed decision making is as ence.11 They accordingly processes for the generation. most distinctive feature of the modern computer problem solving.1 Biomedical Informatics: The Science and the Pragmatics 21 Research) was reborn in 2001 as The Journal of definition of the field and to specify the core Biomedical Informatics. appointed a working group to develop a formal use. Technological approach: BMI builds on 10 http://www.amia. methods. is the generality of its application. Note.elsevier. As a result. infor- and in defining or restricting its contents. statis. do just this for biomedical informatics as well. and knowledge. and lection of facts on which clinical decisions or even management sciences. identifies the focus of the field in edition of this book (2006). and contributes to computer. information. and sharing of biomedical data. We return to this research plans are made. we part to rapid advances in computing and com- use the terms biomedical computer science (for munications technology. eling. We adopt core discipline and should be viewed as encom. to an increasing aware- the methodologic issues) or biomedical comput. important to modern biomedicine as is the col- tics. The nearly driven by efforts to improve human health. which was founded in the the health care enterprise.ketabpezeshki. point shortly when we discuss the basic versus applied nature of the field when it is viewed as a basic research discipline. cognitive science.1: Definition of Biomedical Although labels such as these are arbitrary.1). storage. retrieval. These include the decision sciences. and trated by examples than by attempts at formal def. we used the term bio. we dents seeking graduate training in the discipline. believe that the broad range of issues in biomedi. Informatics they are by no means insignificant. telecom- biomedical-informatics (Accessed 4/8/13). and applies theories. Theory and methodology: BMI published in AMIA’s cal information management does require an journal and approved by the full board of the appropriate name and. unlimited range of computer uses complicates Scope and breadth of discipline: BMI the business of naming the field. disciplinary field that studies and pursues they are important both in designating the field the effective uses of biomedical data. essentially unmanageable by traditional paper- ever. 11 http://www. has recognized the confusion regarding studies. bridging The American Medical Informatics basic and clinical research and practice and Association (AMIA). this definition. ness that the knowledge base of biomedicine is ing (to describe the activity itself). which is very similar to the one passing broadly all areas of application in health. from biological to social systems. pline. mod- nature of computer science is perhaps better illus. experimentation. Box 1. munication. It has ing corollaries that refine the definition and the become the most widely accepted term for the field’s scope and content (Box 1. we offered in previous editions of this text. late 1980s under the former name for the disci. simulation. In the case Biomedical informatics (BMI) is the inter- of new fields of endeavor or branches of science. and information sciences and (Accessed 4/8/13). Much of this book presents examples that cules to individuals and to populations.

which began to appear in The modern digital computer grew out of devel. puting relevant to medicine was Herman Hollerith’s tions. These techniques were the 1. however. draws upon the Babbage in the mid nineteenth century. 1.H. develop a calculus that could be used to simu- Human and social context: BMI.22 E. social and behavioral sciences to inform The first practical application of automatic com- the design and evaluation of technical solu. and the evolution of eco.ketabpezeshki. begun much ear. recog. and general-purpose computers ing was the attempt to construct systems that began to appear in the marketplace by the mid. methods were soon adapted to epidemiologic and Reproduced with permission from public health surveys. 22). educational. census (Fig. late human reasoning. of the early ones instead investigated the notion lier. a seventeenth-century technologies. such as the ENIAC. at least as far back as the Middle of a total hospital information system ( . 2012) mechanical punched-card data-processing technol- ogy. opments in the United States and abroad during One early activity in biomedical comput- World War II.S. development of a punched-card data-processing nomic. The notion of a “logic nizing that people are the ultimate users of engine” was subsequently worked out by Charles biomedical information. social. Gottfried with practical applications in the short term.4. the late 1940s (Collen 1995). Not all biomedical-computing be done with such machines (if they should ever programs pursued this course. tried to in biomedicine. Scholars. Shortliffe and M. however. Early computers. These projects were perhaps less human reasoning might be explained in terms ambitious in that they were more concerned of formal or algorithmic processes. 1. Ages.11 The ENIAC. policies. Fig. His organizational systems. were the precursors of today’s personal computers (PCs) and handheld calculators (Photograph courtesy of Unisys Corporation) (021) 66485438 66485457 www.11). and system for the 1890 U.12). which matured and was widely adopted during the 1920s and 1930s. 1. often had raised the question of whether see Chap. ethical. 14). Many become reliable) had.S. Blois Wilhelm von Leibnitz. Speculation about what might (see Chap. initiating the era of electro- (Kulikowski et al. would assist a physician in decision making 1950s (Fig. emphasizing their application German philosopher and mathematician.2 Historical Perspective precursors of the stored program and wholly elec- tronic digital computers.

large. of the minicomputer in the early 1970s. by heterogeneous types of machines.ketabpezeshki. however.1 Biomedical Informatics: The Science and the Pragmatics 23 Fig. Inc in 2010). Massachusetts.) quently used microcomputers to support administrative and clinical functions (Copyright 2013 Hewlett-Packard Development Company. One approach was based on the con. 1983). Fig.12 Tabulating machines. Utah. The earliest work on HISs in the United States was probably that associated with the MEDINET project at General Electric. 1. ing (such as the UNIX operating system and (which in turn was acquired by Allscripts. and then at the computers—minicomputers—thereby permitting Massachusetts General Hospital (MGH) in the independent evolution of systems in the Boston. Such distrib- Weiderhold at Stanford University in Stanford. Beranek. Reproduced from ~1985 the difficulties they encountered. An alternative was a distributed their own application systems (Fig. California. A number of hospital application pro. Sunnyvale. respective application areas. design that favored the separate implementation followed by work at Bolt.13 Departmental system. Hospital departments. programming environment). the minicomputer (021) 66485438 66485457 www. 1.12 Biomedical-computing activity broadened The course of HIS applications bifurcated in in scope and accelerated with the appearance the 1970s. 1. time-shared computer ments or small organizational units to acquire would be used to support an entire collection of their own dedicated computers and to develop applications. were able to implement Machine was an early data-processing system that per. LP.13). In tandem with the introduction of general-purpose 12 The latter system was later taken over and further devel. . These cept of an integrated or monolithic design in machines made it possible for individual depart- which a single. however. The multi-machine model the early 1960s. The Hollerith Tabulating such as the clinical laboratory. Work on similar systems was was not practical. by Collen at cation among distributed and (sometimes) Kaiser Permanente in Oakland. were original with permission) still formidable. These departments subse- (Photograph courtesy of the Library of Congress. Later the system was part of the suite of products available from Eclipsys. and by scientists at Lockheed in et al. their own custom-tailored systems when affordable mini- formed automatic computation using punched cards computers became available. A common assump- grams were developed at MGH by Barnett and tion was the existence of a single shared database his associates over three decades beginning in of patient information. California. uted HISs began to appear in the 1980s (Simborg California. until network tech- undertaken by Warner at Latter Day Saints (LDS) nologies permitted rapid and reliable communi- Hospital in Salt Lake City. Newman of specific applications on smaller individual in Cambridge. software tools that provided standardized facili- oped by the Technicon Corporation (subsequently TDS ties to individuals with limited computer train- Healthcare Systems Corporation).

1. Blois Fig. especially with support for basic for drugs and other medical products. The trend—reductions in 1970s and early 1980s. The development of the personal computer (PC)—an innovation that made it compact. 080112083626.15).ketabpezeshki. (021) 66485438 66485457 www.24 E. computer circuits. As increasing vidual physicians find it practical to employ PCs in a variety of settings. when the microprocessor size and cost of computers with simultaneous and the personal computer (PC) or microcom. An early advertisement in the 1970s. Not only could hospi. The first articles on computers funding from either government or commercial in medicine had appeared in clinical journals in sources. prices. last four decades.14 Miniature computer. but it was not until the late 1970s research is exploratory and is far from ready for that the first use of computers in advertisements commercial application. a central processing unit (CPU) hardware of various sizes. Within a few years. Because most biomedical-computing the late 1950s.13 This change enormously broadened the base of Progress in biomedical-computing research computing in our society and gave rise to a new will continue to be tied to the availability of software . including for applications 13 http://www. 1. Today indi.14). research in the field (Fig. types. By installing chips in small boxes and con. or “computer on a chip.S. The microprocessor.15 Medical advertising.sciencedaily. 1. the Agency for Health Care Research and Quality management tools were available as commercial (AHRQ). the federal government dealing with computers and aimed at physicians has played a key role in funding the work of the began to appear (Fig. increases in power (Fig.16)—shows no sign of puter became available.” revolutionized the computer industry Fig. The National Library of Medicine products. mainly through the NIH and a wide range of computer-based information. Shortliffe and M. capabilities. although scientists foresee the ultimate tal departments afford minicomputers but now physical limitations to the miniaturization of individuals also could afford microcomputers. 1. inexpensive peripheral devices and personal possible for individual users to purchase their own systems computers (PCs) inspired future experiments in marketing directly to clinicians (Reprinted by permission of copy- put more computing power in the hands of more right holder Texas Instruments Incorporated © 1985) biomedical investigators than did any other sin- gle development until the introduction of the The stage is now set with a wide range of microprocessor.17). for a portable computer terminal that appeared in general necting them to a computer in patient care or clinical investigation. and contained on one or a few chips (Fig. 1. slowing. their descriptions began to appear in (NLM) has assumed a primary role for biomedi- journals alongside the traditional advertisements cal informatics.htm (Accessed 4/8/13). all of which will continue to evolve Everything changed radically in the late in the decades ahead. 1.H. engineers produced medical journals in the late 1970s.

must be viewed in the con. The short-term commercialization.17 The National Library of Medicine (NLM). and the implied potential for future Will both patients and health workers eventually benefits to medicine. 102 ers (Source: San Jose Mercury News. on the campus of the National Institutes of Health (NIH) in Bethesda. subsequent enormous growth in computing activ- ity has been met with some trepidation by health professionals. It is also a major source of support for research in biomedical informatics (Photograph courtesy of the National Library of Medicine) numbers of applications prove to be cost. system. This graph shows the i486TM Pentium® 105 64K i386TM exponential growth in the 16K 4K 80286 number of transistors that 104 1K 8086 can be integrated on a single 103 8080 microprocessor by two of 4004 the major chip manufactur.ketabpezeshki. Will 1. 1.1 Biomedical Informatics: The Science and the Pragmatics 25 Fig.3 Relationship to Biomedical health workers gradually be replaced by comput- Science and Clinical Practice ers? Will nurses and physicians need to be highly trained in computer science or informatics before The exciting accomplishments of biomedical they can practice their professions effectively? informatics. The NLM. Former Intel chairman 4G 1010 Microprocessor 2G 1G Gordon Moore is credited 512M with popularizing the “law” 109 Memory 256M 128M that the size and cost of 108 16M 64M ItaniumTM microprocessor chips will Transistors Per Chip 107 4M Pentium®4 half every 18 months while Chip Memory 1M Pentium®III they double in computing 106 256K Pentium®II power. As early as . 101 used with permission) 100 1970 1975 1980 1985 1990 1995 2000 2005 2010 Fig. an eminent clinician effective. They ask where it will all end.16 Moore’s Law. revolt rather than accept a trend toward automa- text of our society and of the existing health care tion that they believe may threaten the traditional (021) 66485438 66485457 www. on med- programs will focus increasingly on fundamental ical education. Maryland. Dec 2007. it is likely that more development work suggested that computers might in time have a will shift to industrial settings and that university revolutionary influence on medical care. and even on the selection criteria research problems viewed as too speculative for for health-science trainees (Schwartz 1970). 21). 1. is the principal biomedical library for the nation (see Chap.

including patient-care foci such as cine (see Sect. percentage of basic research discoveries that ture. 1. of these historically has been clinical care ment and biomedical science. has shown that there information management and decision making can be a very long period of time between the (Fig. Blois of potential areas of application (Fig. Inc) informatics. To illustrate what (including medicine. called basic advertisements in medical journals (such as this one for an antihypertensive agent) began to use computer equipment research. Research.19. This approach sup- ports the integration of the findings and their anal- yses. with a wide variety former name of the discipline. 1. The analogy with other basic sciences is that biomedi- cal informatics uses the results of past experience to understand. 10) (Shortliffe 1993)? Will clinicians be and the identification of basic research tasks that viewed as outmoded and backward if they do not characterize the scientific underpinnings of the turn to computational tools for assistance with field. ply to search for new knowledge. the as a basic biomedical science. and Biomedical informatics is perhaps best viewed even veterinary informatics. erinary care). we can contrast patient-oriented informatics applications. Division of ICI Americas.H. information. however.20). health planning. 1. there is an especially tight coupling between the application areas. 1. designing experiments. 1. 1.19). One goal is sim- Fig. knowledge. and encode objective and subjective biomedical findings and thus to make them suitable for processing.18)? development of new concepts and methods in Biomedical informatics is intrinsically basic research and their eventual application in entwined with the substance of biomedical sci. leaving only a small biomedical science is constrained by that struc. The for practical ends. We the properties of the information and knowledge refer to this area as clinical informatics. thereby inherently motivated by problems encountered in addressing specifically the interface between the a set of applied domains in biomedicine. by its nature. and supporting technologies with analyses of Work in biomedical informatics (BMI) is biomedical information and knowledge. perform- ing analyses. It typical of such fields as physics or engineering includes several subtopics and areas of special- with the properties of those typical of biomedi. an experimental science. structure. ized expertise. and using the information gained to design new experiments. In turn.26 E.5). and vet- we mean by the “structural” features of biomedi. medical (021) 66485438 66485457 www. dental informatics. research. It determines and analyzes the structure of Furthermore (see Fig. A second goal is to use this knowledge as props and even portrayed them in a positive light. There is a continuity between these date physician feels comfortable using computer-based two endeavors (see Fig. nursing. many discoveries biomedical information and knowledge. Shortliffe and M. Furthermore. Chap. 1. an area of activity that demands cal information and knowledge. the selective distribution of newly created knowledge can aid patient care. In biomedical tools in his practice (Photograph courtesy of ICI Pharma.S. called applications (applied) suggestion in this photograph seems to be that an up-to. of patients. whereas are discarded along the way. characterized by pos- ing questions. The first science of information and knowledge manage. . the biomedical world (Balas and Boren 2000). Biomedical informatics melds the study have a practical influence on the health and care data. broad categories humanistic values in health care delivery (see of which are indicated at the bottom of Fig. nursing informatics.19). By the early 1980s. and basic biomedical research.18 Doctor of the future. dentistry. Biomedical informatics is. decision making.

0 . . A synthesis of studies focusing on various National utilization rates of specific. Note that work in biomedical informatics is basic science discipline in which the development and motivated totally by the application domains that the field evaluation of new methods and theories are a primary is intended to serve (thus the two-headed arrows in the focus of activity. sometimes within a few (021) 66485438 66485457 www. The informatics subfields indicated by the terms in the real world of health or biomedicine for which an across the bottom of this figure are accordingly best informatics solution is sought (see text) viewed as application domains for a common set of Original research (100%) Negative results :18% Variable Negative results: 46% Submission 0.6 year Publication Lack of numbers: 35% 0.5 year Acceptance 0. guidelines. These core concepts and methods in turn diagram). practice. procedures also suggests a delay of two decades in reach- age of 17 years to make innovation part of routine care ing the majority of eligible patients (Courtesy of Dr. well-substantiated phases of this transfer has indicated that it takes an aver. We concepts and techniques from the field of biomedical view the term biomedical informatics as referring to the informatics. 1. and Theories Imaging Clinical Public Health Applied Research Bioinformatics Informatics Informatics Informatics And Practice Fig. Techniques.13. but nationwide introduction is usually slow. Pioneering institutions often Andrew Balas) apply innovations much sooner. textbook 5.8 years Inplementation (14%) Fig. field generally result from the identification of a problem ences.3 year Bibliographic Inconsistent indexing: 50% databases 6.20 Phases in the transfer of research into clinical weeks.1 Biomedical Informatics: The Science and the Pragmatics 27 Basic Research Biomedical Informatics Methods. Therefore the basic research activities in the have broad applicability in the health and biomedical sci.19 Biomedical informatics as basic science.ketabpezeshki.0 years Reviews. (Balas and Boren 2000).

1. 1. Shortliffe and M.H. the term health informatics. Blois informatics. It is not a syn- informatics as a core scientific discipline and its diverse onym for the underlying discipline. rather. 9 and 20). work (also called structural informatics at some Closely tied to clinical informatics is public institutions).21. informatics and public health informatics. 16).28 E. At the next and the role of physicians. As was previously dis. 1. 1. The core sci- als (see Chap. methods and techniques. since same informatics methods as well (see Chap. BMI is applicable to basic human biology as well As is shown in Fig. informatics” ence. health informatics (Fig.S.19. there is a spectrum as as to health.19). and others (021) 66485438 66485457 www.19 have “fuzzy” Basic Research Biomedical Informatics Methods. tend to be the emphasis of imaging informatics ger used to refer to the discipline as a whole.19 that biomedical informat- include imaging informatics (and the set of ics and bioinformatics are not synonyms and it is issues developed around both radiology and other incorrect to refer to the scientific discipline as image management and image analysis domains bioinformatics. These Note from Fig. workers focus on tissues and organs. workers application shown in Fig. cepts. clinical practice. In bioinformatics. level. an important such as pathology. and Theories Health Informatics Applied Research Imaging Clinical Public Health And Practice Bioinformatics Informatics Informatics Informatics Molecular and Cellular Tissues and Individuals Populations Organs (Patients) And Society Processes . imaging. the term “medical informatics” is no lon. dermatology. is now reserved for those applied deal with molecular and cellular processes in the research and practice topics that focus on disease application of informatics methods. which is. where similar the focus is on individual patients. is also not offering challenges that draw on many of the an appropriate name for the core discipline. Finally.21 Building on the concepts of Fig. and finally to methods are generalized for application to popu. Thus clinical informatics and ence of biomedical informatics has important public health informatics share many of the same contributions to make across that entire spectrum. which cussed. Progressing to clinical informatics. there is the burgeoning area of bioinformatics. public health. which is “biomedical array of application domains that span biological sci. Two other large areas of and many informatics methods are broadly appli- application overlap in some ways with clinical cable across the same range of domains. Note that “health informatics” diagram demonstrates the breadth of the biomedical is the term used to refer to applied research and practice informatics field. 1. The relationship between biomedical in clinical and public health informatics. public health.ketabpezeshki. lems of populations and of society. Techniques. one moves from left to right across these BMI We acknowledge that the four major areas of application domains. where researchers address prob- lations of patients rather than to single individu. and molecular area of application of BMI methods and con- visualization—see Chaps. 24). 1. Similarly. this not illustrated (see text). which refers to applied research and practice in which at the molecular and cellular levels is clinical and public-health informatics.

cusses these issues in greater detail. in actual therapy for a patient or proceeding to other mod- systems that are used in clinical or biomedical ular tasks. the various cognitive research involve more than one of the categories. One important implication of this model is oversimplified and that such a this viewpoint is that the core discipline is identi. and biologic ric of medical care is a continuum in which these applications are often trained together and are required elements are tightly interwoven. decomposition of cognitive tasks may be quite cal. ceived and taught as though it were a free-standing ration from one of the application areas. Elective courses and internships in results. or a science. . biomolecular imaging involves to be considered separately and in isolation— both bioinformatics and imaging informatics they have been largely treated as though they are concepts. activities of physicians traditionally have tended For example.) is areas of specific interest are of course important a process that never really terminates. a technology. 2007). consumer health informat. Physicians seem to deal with several tasks some BMI methods have greater relevance to at the same time. sion making is diagnosis. managing therapy. databases. 17) includes elements of both clin. which is often con- In general.14 evidence supporting the diagnosis or suggests a The scientific contributions of BMI also can second and concurrent disorder. physicians carry out in isolation before choosing totypes or. ing the education of health professionals When we speak of making a diagnosis. We can assume computers will be used (021) 66485438 66485457 www. separating trainees it turns out that treatment based on it is based on the application areas that may interest unsuccessful or if new information weakens the them would be counterproductive and wasteful. choos- (Shortliffe 2010). For example. 25). identify. Patel and Groen given individual is motivated to address. Chap. Medical students may ing fundamental methodologic issues that need thus be led to view diagnosis as a process that to be addressed and testing them in system pro. The fab- were designed with this perspective in mind. is that of formal medical decision making (see Another important area of BMI research activi. 3). analysis of treatment tions interests. Regardless of to learn something about each of the other application whether we view computer and information sci- areas. for more mature methods. icine. (diagnosis. mance. monitoring a patient. and independent activity. ones that bring when they see a new patient. patient assessment together students with a wide variety of applica. given the need for teamwork erally appropriate to alter the original diagnosis if and understanding in the field. Such work requires the knowledge and information reveal that it is one analysis of linked genotypic and phenotypic aspect of biomedical informatics. Students with interests in clinical. an entity that has overall unity.1 Biomedical Informatics: The Science and the Pragmatics 29 boundaries.ketabpezeshki. although 1986). One activity attracting increasing interest ical informatics and public-health informatics. in the education ing a treatment. 14 we are using labels for different aspects of medi- Many current biomedical informatics training programs cal care. and therefore lies at the intersection of A particular topic in the study of medical deci- bioinformatics and clinical informatics. 1978. even while specializing in one subarea for their ence as a profession. BMI researchers derive their inspi. A number of studies have shown that research settings. but. or preventing disease. and many areas of applied informatics of medical students. but the discipline’s is the effort to infer genetic determinants of dependence on formal methods and its use of human drug response. public health. making deci- sions. Chapter 4 dis- be appreciated through their potential for benefit. It is gen- should receive. which to be defined completely. Similarly. etc. management. own research. Several such programs were described in a series of articles in the Journal of Biomedical Informatics there is no doubt about its importance to biomed- in 2007 (Tarczy-Hornoch et al. monitoring of disease progression. A physi- complements to the core exposures that students cian must be flexible and open-minded. one of the first things physicians think about fied BMI educational programs. The specific content of this area remains ties is pharmacogenomics (see Chap. This argues for uni. regardless of the area of application that a misleading (Elstein et al. Although a diagnosis may be some domains than to others. independent and distinct modules of perfor- ics (see Chap.

A program for sci- programs were located in departments of electri. however. medicine.21. 1970. design and the development of practical hard. developed specifically for biomedical applica. ability theory. MUMPS was the most widely used language for medical record processing. than simply the biomedical application of com- Biomedical informatics draws from all of puter science. the study of computer science with a “biomedical guages and software. or the psychol- minals) have people assumed that biomedical ogy of human problem solving).4 Relationship to Computer its new name. Science New implementations have been developed for each generation of computers. any programming language. like During its evolution as an academic entity in uni. contribute to. known basic researcher or doctoral student will typically as the MUMPS language (Greenes et al. which was specially research. or the manage- designed for general use. The issues that it addresses not these activities—development of hardware. computer computer science? Is the new discipline simply scientists were interested in programming lan. be motivated by one of the application areas. medical informatics researcher will be tightly tions would require specialized programming linked to some underlying problem from the real languages might have been answered affirma. Furthermore. M. M. but the underlying sciences on computing generally has not had a large enough which BMI professionals draw are inherently market to influence the course of major hardware interdisciplinary as well (and are not limited to developments. 1. for example. developed special medical ter. successful BMI research will often draw on. because major concerns of their whether it is designed for chemical engineering researchers were computer architecture and or for pharmacokinetic calculations. For several years.. undertakings not particu.ketabpezeshki. and tions. ware. computer science followed an unsettled for all computing tasks.15 and other theoretical topics seemed begin to see why biomedical informatics is more more related to mathematics. but a dissertation worthy of a PhD in the field will 15 Many interesting problems cannot be computed in a rea. 1. such as those shown at the bottom of Fig. information sciences. soft. Under 1. does BMI differ from biomedical ware components. only have broad relevance to health. (021) 66485438 66485457 www.1.22 tively in the 1970s by anyone examining the illustrates. i. Biomedical and biology. and health science education. a bio- The question of whether biomedical applica. cognitive applications would use hardware other than that science. is not equally useful versities. 1. ment sciences (Fig. scientists can build in the future. ical informatics that we provided in Sect. biomedical Bowie and Barnett 1976). computer science. As Fig.22).19. 1. usually be identified by a generalizable scientific sonable time and require heuristics.4. How. a biomedical informatics MGH Utility Multi-Programming System. developed for use in medical applications. entific computation looks pretty much the same cal engineering. Computability theory result that also contributes to one of the compo- is the foundation for assessing the feasibility and cost of computation to provide the complete and correct results to nent disciplines (Fig. world of health or biomedicine. Shortliffe and M.S.30 E. Thus. we believe you will theory.4. At the same time. computers have not been computer science topics). Not since the early 1960s (when health. then. they organizational place. it is still in widespread use. but it may also computing experts occasionally talked about and.H. their work with algorithm design. be closely related to the decision sciences (prob- in a few instances.22) and on which other a formally stated problem. 1. flavor”? If you return to the definition of biomed- larly characteristic of engineering. rather. Blois increasingly in clinical practice. for example.1. Many computer science are specific to the kind of task.e. computability and then refer to Fig. the software course as involved faculty attempted to identify requirements of medicine are better understood key topics in the field and to find the discipline’s and no longer appear to be unique. In . decision analysis. Furthermore. and computer science theory.

1. Cognitive Science Information Sciences Management Clinical or Sciences Biomedical Domain of Other Applied Interest Component Information Sciences Draws upon...4. begun to be true of biomedical informatics aca. and Theories Computer Science Decision Science Draw upon.22 Component sciences in biomedical informat. and image-enhancement techniques for engineering is an older and more well-established use in radiology). At the methodologic level.16 The emphasis in such departments and hourly summaries are examples of such has tended to be research on. when technology began ogy. ics. For example.g. As Figs. advanced monitoring systems. How does biomedical informatics been used both in the design and construction relate to biomedical engineering. Thus any science is only one. especially in an of medical devices and in the medical devices era when engineering and computer science are themselves... the “smart” devices increasingly intertwined? increasingly found in most medical specialties Biomedical engineering departments are all dependent on computational technol- emerged 40 years ago. of which computer attempts to contribute solutions to problems. computing techniques have demic units. Techniques. both in its areas of application and in the delineation of basic research challenges in biomedical component sciences on which it draws informatics that must be tackled if the applied biomedical 1.. Only in the last 2 decades or so has this ing and related wet-bench research . pros- formal academic programs in the latter subject. An informatics application area is motivated by the biomedical informatics draws on. biomedical use. There often with departmental status and full-time is also a major emphasis on tissue engineer- faculty.. a needs of its associated biomedical domain. theses.. instrumentation (e. In recent years. and in turn often leads to the plinary. biomedical informatics is inherently multidisci- domain for its inspiration.. Statistics Contributes to.5 Relationship to Biomedical of. Many engineering and medical schools have the development of medical devices... as discussed in Chaps. (021) 66485438 66485457 www. Engineering 19 and 20.. domain is ultimately to benefit.21 show explic- applied informatics work draws upon a biomedical itly. and contributes to.19 and 1. spe- cialized transducers for clinical or laboratory If BMI is a relatively young discipline. Fig. with an orientation toward one. and specialized research tools. Intensive care monitors that generate blood to play an increasingly prominent role in medi. to which it wide variety of component disciplines.1 Biomedical Informatics: The Science and the Pragmatics 31 Contribute to. 1972) accredited by the Engineering Council for cal engineering was the first department (September Professional Development.. and development “intelligent” devices. 1.ketabpezeshki. 16 The Duke University undergraduate major in biomedi. pressure records while calculating mean values cal practice. Biomedical Informatics Methods.

We need to know something use computing technology. If we are talk- clude that biomedical applications do not raise ing about the solvent properties of water. no difference whether we happen to be work- trary. it makes any unique problems or concerns. Physics characteristically has a where they meet. so using applied computing. in interesting ways.. Clinical information seems computers for these applications requires only to be systematically different from the informa. physics will be placed tems and . including medicine. you can consult Blois’ book on this subject (see One might object to this line of argument by Suggested Readings). the biomedical environment raises several ing in geology. 20 on radiology systems. we are using the descriptions of als. Blois The overlap between biomedical engineering Let us examine an instance of what we will and BMI suggests that it would be unwise for us call a low-level (or readily formalized) science. however. The con- ences in emphasis that can help you to under. The puter science differs from conventional computer person using computers to analyze the descrip- science in fundamental ways. is independent of the application. topics—e. the computer is secondary. after all. We shall explore tions of these high-level objects and processes these differences only briefly here.S. The cal devices. librarian Melvil Dewey). engineering. computers are used (021) 66485438 66485457 www. 19 on patient-monitoring sys. in all applied fields. the behavior of whom has some investigators suggest that biomedical com. remarking that. tion used in physics.H. there are other chemistry (which more closely resembles chemi. we talk spend much time examining biomedical engi. higher-level processes carried out in more com- cal applications generally than it does medical plex objects such as organisms (one type of which ones). certain kind of simplicity. Shortliffe and M. in any discus- the two fields. you might con. Aspects of biomedical information include is patients). in BMI. In both senting or explaining functions that we regard as fields. and there are chapters in this book sciences (from the fourth-century BC Greek phi- that clearly overlap with biomedical engineering losopher Aristotle to the twentieth-century U.S. cepts and descriptions of the objects and pro- stand their different evolutionary histories.5 The Nature of Medical tory or for an engineering application. Even near the bottom. human beings. the fields have differ. to draw compulsively strict boundaries between Physics is a natural starting point. Applying a computer (or any formal computa- tion) to a physical problem in a medical context is no different from doing so in a physics labora- 1. When we discuss. all about a physiological process—and this describe. Chap. The use of Information computers in various low-level processes (such as those of physics or chemistry) is similar and From the previous discussion.ketabpezeshki. although both medical in nature. There are ample opportunities for sion of the hierarchical relationships among the interaction. the emphasis is on biomedi. encounters serious difficulties (Blois 1984). These differences raise special problems and very high-level objects.g. to explain biomedical computer science.32 E. for details. are quite distinct low-level processes of physics are particularly from those encountered in most other domains of receptive to mathematical treatment. no counterpart in physics or in engineering. Many of the important informational an essence of uncertainty—we can never know processes are of this kind. In cesses of physics. so we shall not how nutrient molecules are metabolized. about the role of electron-transfer reactions. or record the properties or behavior of results in inevitable variability among individu. for example. laws of physics and the descriptions of certain cal information and knowledge and on their kinds of physical processes are essential in repre- management with the use of computers. however. or generality. engineering. or even clinical In biomedicine. neering topics. however. Such issues that. to under- book is on the informatics end of the spectrum of stand why water is such a good solvent. The emphasis in this about molecular physics. the emphasis is on medi. are necessarily used biomedical engineering. On the con. or medicine. conventional numerical programming.

change the delivery of health care. In gen. however. 1. In bio. In formal highly abstracted that the events or processes logic. will be determined in part by exter- as human cognition. It is difficult or impos- For example.g.. the gen. whereas those of high-level ones tend will determine the extent to which computers (021) 66485438 66485457 www. patients. to assume that all propositions in the banking business is the customer. the customer can Woodrow Wilson a good president?” cannot be remain an abstract entity. inten. answered with a “yes” or “no” (unless we limit however. we could not begin to deal with a the question to specific criteria for determin- patient represented with such skimpy abstrac. To describe commercial activities such as these.ketabpezeshki. Such questions as “Was we need only a few properties. we emphasize the myriad ways in very low-level descriptions.g. we begin with the assumption that a given have been reduced to low-level objects. In this chapter and through- puter applications that range from processing of out the book. Many common tions. and complex issues to which conventional logic and the health care system to accrue the potential mathematics are less readily applicable. the last five decades. proposition must be either true or false. We must deal with the rich descrip. Practice cess this information using the tools of mathe- matics and computer science that work so well It should be clear from the previous discussion at low levels. involved and that relevant computations are car. and inexact (e. In clinical medicine. and unambiguous (e. and their rate of beings in their entirety (including such aspects occurrence. or engineering. which are completely and sys. In light of these remarks. the burdens of information processing and the istry.6 Integrating Biomedical tions occurring at high levels in the hierarchy. When we study human which such changes are realized.” “good”). We will find that they raise the ability of scientists.1 Biomedical Informatics: The Science and the Pragmatics 33 routinely in commercial applications in which to be soft.” are affecting biomedical computing and that “mass”). which computers are used in biomedicine to ease ferent from their counterparts in physics. abstractions carried to this degree ture is essential because logic is concerned with would be worthless from either a clinical or the preservation of truth value under various research perspective. Informatics and Clinical and we may be hard pressed to encode and pro. formal transformations. We have argued that (AI) can be aptly described as the application of information management is intrinsic to clinical computer science to high-level. and implementing biomedical applications and level descriptions. who have truth values when we deal with the many may deposit. clinicians. We must be prepared to analyze most of questions in biomedicine have the same property. indeed. high-level descriptions in medicine or. This fea- medicine. real-world practice and that interest in using computers to problems. The degree to tematically . chem. the inference these commercial applications. borrow. ing the goodness of presidents). self-consciousness. fuzzy. which are little dif. aid in information management has grown over Biomedical informatics thus includes com. the complex behaviors that human beings dis- play and to describe patients as completely as possible. eral. benefits. nal forces that influence the costs of developing tionality. in everyday situations. one instance of a human being sible. The explanation is that. “unpleasant human beings and situations concerning them are scent. crisp. or invest money.. to processing of extremely means by which new technology promises to high-level ones. Just as we need to develop different meth- ried out successfully. we must use these high. that biomedical informatics is a remarkably eral enterprise known as artificial intelligence broad and complex topic. the descriptions methods we use with such objects may differ of human beings and their activities have been so from those we use with low-level ones. withdraw. “length. in ods to describe high-level objects. and behavior). the attributes of low-level objects appear We can summarize several global forces that sharp.

As they become the historical development of biomedical com. com/2012/08/19/harvard-stores-704tb-in-a-gram-of-dna/ technologies replace less expensive. We touched on the first of ing. computers does not. Some new of information in a gram of DNA. whether as an ATM. tasks is a frequently cited phenomenon in mod- phone that takes photographs and shares them ern medical practice. 4 and 6). tests. prices. and as the programs they develop are of data-intensive applications. (2) a gradual increase in the number of a specially trained librarian. Standardization of hardware and advances able when they turn to the computer for help with in network technology are making it easier to information management tasks. based systems that are poorly designed. likely to have useful and usable systems avail- 20. Physicians trained in longer are considered adequate for making recent years may have used computer programs diagnoses and planning treatments. ness to embrace the machine. and even to individual physicians.H. or as a tele. introduced. unduly time-consuming. Sect. As more profes- are facilitating the inexpensive storage of large sionals are trained in the special aspects of both amounts of data. Clinical personnel and (3) ongoing changes in health care financing will continue to be unwilling to use computer- designed to control the rate of growth of health. Consider. confus- related expenditures.ketabpezeshki. guarantee an eager- cine or another health profession and in BMI. more sophisticated in the use of computers in puting and the trend from mainframe computers. 1. health care professionals are more digital radiology department discussed in Chap.1 similarly builds on the influence that the The second factor is the increase in the num- Internet has provided throughout society during ber of professionals who are being trained to the last decade. or lacking in clear these factors in Sect. They may have physicians to find subtle diagnostic signs by examining various parts of the body nonetheless 17 Technological progress in this area is occurring at a often choose to bypass or deemphasize physical dizzying rate. In fact. Shortliffe and M. for example.34 E. they do so without paying grown” storage and can store as much as 704 terabytes sufficient attention to the ensuing cost. as the dency to apply technology to all patient-care microprocessor in a microwave oven. Mere physical findings no wirelessly with others. and capabilities makes computer applications Health professionals who receive formal training both attractive and accessible. Technological in BMI are likely to build systems using well- advances in information storage devices. thus improving the feasibility fields.S. and to the mobile of clinical software will become only more devices of today.4. the announcement examinations in favor of ordering one test after that scientists are advancing the notion of “organically. however. another. Computer scien- thus available to hospitals. The escalating ten- aspects of our lives. either directly or with the assistance of ware. The future view outlined in demanding. The new hardware technologies understand the biomedical issues as well as the have made powerful computers inexpensive and technical and engineering ones. literature. mistakes of other developers. is managed care and the increasing pressure to Computers are increasingly prevalent in all control medical spending. medi- to learn diagnostic techniques or to manage the cal students who are taught by more experienced therapy of simulated patients. other aspects of their lives. share data and to integrate related information. The to design systems responsive to actual needs and broad selection of computers of all sizes. http://www. their expectations to microcomputers and PCs.17 established techniques while avoiding the past including the movement of files to the “cloud”. the use of (021) 66485438 66485457 www. 1. The third factor affecting the integration of management functions within a hospital or other computing technologies into health care settings health care organization. Blois are assimilated into clinical practice: (1) new learned to use a computer to search the medical developments in computer hardware and soft. such as the all. sensitive to workflow and the clinical . to departments within tists who understand biomedicine are better able hospitals. Sometimes. Simple exposure to individuals who have been trained in both medi. logically inferior.engadget. In such cases. but techno- (Accessed 4/21/13).2. when we described benefit (see Chaps.

that the nature of medical descriptions accounts for difficulties in applying computing technology to ing the quality are less clear. analyzes the structure of medical knowledge in terms viding that care clearly will be favored. A. Shortliffe. Berkeley: University of California Press. J. Ozbolt. In this classic volume. Occasionally. will be embraced only if Coiera E. Medical technologies.1 Biomedical Informatics: The Science and the Pragmatics 35 the more expensive approach is generally pressures produce a growing reluctance to justified. (1984). out excessive delays? Is the system reliable This classic collection of papers describes detailed enough to avoid loss of data? Can users interact studies that have illuminated several aspects of the ways in which expert and novice physicians solve easily and intuitively with the computer? Are medical problems. Elstein... MA: Harvard University Press. S. H. Hartman Publishing. I. Bethesda: American Medical Informatics Association. & Sprafka. S. the opposing trends will in large part determine the scans produced with computed tomography or degree to which computers continue to be inte- magnetic resonance imaging (see Chaps. providers have a greater better. rapid advances in computer sectional slices of the body for the first time. and the belief that more technology is health care industry. tice. (021) 66485438 66485457 www. (1995). (1978). ever. P. Boca Raton. (2013).. the ideas of high. S. cost-control K. are computers conveniently located? Should in the English-language literature in 1974). S. helped to fuel the rapidly escalating health need for the information management capabili- care costs of the 1970s and 1980s. M. from prying eyes? In addition. including computers. FL: CRC either reducing costs or providing benefits that Press. to evaluate the quality of care provided. Designers of medical systems must. For exam. ing the domain accessible and understandable to the Improvements in hardware and software make non-specialist. Sintchenko V. R. leading to the ties supplied by computer systems. and identifies the fully integrated into medical practice. computers more suitable for biomedical applica. Cambridge. In summary... to analyze the relation. F. Shulman. 11). and life sciences research. C. public health. Information and medicine: The tify areas of inefficiency. P. aimed at mak- clearly exceed their costs. Integrated computer systems potentially provide the means to capture data for detailed cost accounting.. A. consciousness. Systems that improve nature of medical descriptions. This introductory text is a readable summary of clinical and public health informatics. Kohane. Magrabi F. M. mobile devices replace the tethered workstations Medical problem solving: An analysis of clinical rea- of the past? Can users complete their tasks with. Guide to they improve the delivery of clinical care while health informatics (3rd ed).. how. Altman. He explores effect of cost containment pressures on technolo. favor the implementation of functions that previously could be checked only effective computer applications in clinical prac- episodically (see Chap. G.... L. 19). address satisfactorily many logistical and This comprehensive book traces the history of the field engineering questions before computers can be of medical informatics through 1990. Yet the development of expensive new tech. 9 and grated into the health care environment.and low-level sciences and suggests gies that increase the cost of care while improv. The of a hierarchical model of information. McCormick. in the increasingly competitive nologies. coupled with an increas- medical instruments in intensive care units per. Suggested Readings ship of costs of care to the benefits of that care. Furthermore. L. ing computer literacy of health care profession- form continuous monitoring of patients’ body als and researchers. embrace expensive technologies that add to the nologies have allowed us to perform tasks that high cost of health care. Collen. A history of medical informatics in . A.. and to iden. Miller. medicine. soning. For example. origins of the discipline’s name (which first appeared ple. The net effect of these previously were not possible. 20) have allowed physicians to visualize cross. S. patient data secure and appropriately protected Friedman. and hardware and software. B.ketabpezeshki. computer-related tech. the United States: 1950 to 1990. The challenge introduction of managed care and capitation— is to demonstrate in persuasive and rigorous ways changes in financing and delivery that were the financial and clinical advantages of these sys- designed to curb spending in the new era of cost tems (see Chap. Blois. the author the quality of care while reducing the cost of pro. E.

health professionals in management roles or IT pro. (2012). nical quality of health care entails the nology among policy makers. Major programs in patient safety have resulted from 5. and they have provided motivation for a heightened interest in health care information tech. the relationship between clinicians and their patients. Raleigh. Institute of Medicine (2001). patients. Discuss two ways of information technology and the enhancement of in which such a system could change quality and reduction in errors in clinical practice. Do you believe that improving the tech- these reports. the author examines the frequently expressed concern that the (021) 66485438 66485457 . risk of dehumanization? If so. National Research Council (2000). Doctors. Tuck. How do you interpret the phrase “logi- informatics: practical guide for healthcare and infor. Academy Press. Principles of biomedical informatics. retrieval. because they share the same core meth- Shortliffe. He argues. 1. This volume provides a techni- cal introduction to the core methods in BMI..H. Health 1. which shows that New York: Academic. Yoshihashi A. Szczepaniak. computer program is “effective”? Make a Institute of Medicine (1991 [revised 1997]). DC: National that will determine the extent to which Academy Press.S. with storage. To Err is Human: Building a 3. how do you interpret it? Institute of Medicine (2000). Training the next generation of settings will disrupt the development of rapport informaticians: The impact of BISTI and bioinformat. 390–398 In this paper. rather. C. logically? Do people behave logically? NC: Lulu. The list of a dozen computer applications with Computer-Based Patient Record: An Essential Technology for Health Care. as well as organization of academic groups and curricu- lum development. 11. dealing bioinformatics. Information.1. clinical informatics. (2008).. cal behavior”? Do computers behave mation technology professionals (5th ed). E.. This set of three reports the same tasks (this will require that you from branches of the US National Academies of determine what it means for a human Science has had a major influence on health informa- being to be effective). For The Record: Protecting Electronic Health as you have explained this concept.ketabpezeshki. that Association. Application examples are drawn from of the biomedical informatics discipline bioinformatics. indicate your esti- Press.36 E. Reread the future vision presented in Safety: Achieving a New Standard for Care. and use of biomedical clinical informatics. (1993). informatics are all application domains sion making. Discuss three society-wide factors Safer Health System. Washington. Sect. 1. J. 137(3). DC: National practice. Bailey N. D. What do you think it means to say that a fessionals who are entering the clinical world. This important analysis computers may have precisely the opposite effect on addresses the changing nature of biomedical infor. (2004). clinical practice. This introductory volume provides a broad view of informatics and is aimed especially at Explain your answers. are discussed. computers are assimilated into clinical Crossing the Quality Chasm: A New Health Systems for the 21st Century. D. Then.. and computers: ods and theories: Will information technology dehumanize health care (a) Briefly describe two examples of delivery? Proceedings of the American Philosophical core biomedical informatics methods Society. Journal of American Medical Informatics ize the therapeutic process. DC: National Academy Press. Washington. Describe the characteristics of Washington. Blois Stormo. mate of how well human beings perform DC: National Academy Press. Patient 4.. DC: which you are familiar. and public health data for biological problem solving and medical deci. the risk? Explain your reasoning. J. M. 167–172. pattern? If so. imaging informatics. Shortliffe and M. DC: National Academy for each application. This an integrated environment for managing series of three reports from the Institute of Medicine has outlined the crucial link between heightened use clinical G. (1997). Networking Health: Prescriptions for the Internet. 6.19. & introduction of computing technology into health care Williamson. display. between clinicians and patients and thereby dehuman- ics. E. Do you discern any tion technology education and policy over the last 25 years. National Research Council tions in decreasing order of effectiveness. Washington. Consider Fig. Institute of Medicine (2004). List the applica- National Academy Press. Washington. 2. matics due to the revolution in bioinformatics and computational biology. I. and public health informatics. Kalet. is it worth and even patients. Questions for Discussion Hoyt R. Implications for training. provider organizations. Washington.

In 2000. computer decreases the incidence of prevent- science.19 In other words.ketabpezeshki. address this problem (other than based systems could increase the “designing a better user interface incidence of medical errors? Explain. the study design to a mathematics faculty member. informatics? [Hint: See Fig. a major report by the Institute ment? Do you believe that it is of Medicine entitled “To Err is Human: important to the quality and/or effi- Building a Safer Health System” (see ciency of care for clinicians to be Suggested Readings) stated that up to able to record their observations. 8. ture “nuance” in the description of what a medical informatics? clinician has seen when examining or (c) Why is biomedical informatics not interviewing a patient may not be as cru- simply computer science applied to cial as some people think. against the use of structured data-entry vance of psychology and cognitive methods using a controlled vocabulary and science to the field of biomedical picking descriptors from lists. using free ventable medical errors in American text/natural language? hospitals each year. data entry because of the increased What are three specific ways in time required for documentation at which they could reduce the num. should address whether the com- who responds that “in that case. I’d puter-based system increases or also argue that . 2.22] (a) What is your own view of this argu- 7. the case of bio. (b) Many clinicians may be unwilling (a) It has been suggested that electronic to use an electronic health record health record (EHR) systems should (EHR) system requiring structured be used to address this problem.” In your opin. What are two ber of adverse events in hospitals? strategies that could be used to (b) Are there ways in which computer. The desire to be biomedicine. 1. the point of care. (b) Imagine that you describe Fig. share the same core mathematical (d) What are the limitations of the methods and theories. for the system”)? (021) 66485438 66485457 www. It has been argued that the ability to cap- and different from. or the practice of able to express one’s thoughts in an unfet- medicine using computers? tered way (free text) is often used to argue (d) How would you describe the rele. at 98.1 Biomedical Informatics: The Science and the Pragmatics 37 and theories that can be applied (c) Describe a practical experiment that both to bioinformatics and clinical could be used to examine the impact informatics. and physics are all applica. experimental design you proposed ion. is this a legitimate argument? In in (c)? what ways is this situation similar to. able adverse events in hospitals – tion domains of math because they and by how much. least part of the time.000 patient deaths are caused by pre. of an EHR system on patient safety.

1 What Are Clinical Data? answers to these questions: • What are clinical data? From earliest times. and interpretation? the modern physician’s use of complex labora- • What distinguishes a database from a knowl. vidual patients (their risks.H. Storage. simple reflection 272 W 107th St #5B. New York 10025. making (as described in detail in Chaps. and again in Chap.H. or using data. knowledge. Data provide G. public health. because they are crucial to the process of decision Weill Cornell College of Medicine. © Springer-Verlag London 2014 (021) 66485438 66485457 www. storage. Barnett. They also help a Boston 02114. USA will reveal that all health care activities involve e-mail: ted@shortliffe. Shortliffe. USA physician to decide what additional information E.O. MD. and specificity? responses to therapy).1007/978-1-4471-4474-8_2. retrieval. the sheer amounts of data • How are the terms related? that may be used in patient care have become • What are the alternatives for entry of data into huge. Cimino (eds. 39 DOI 10.ketabpezeshki. and use. Whether paper medical record? we consider the disease descriptions and guide- • What is the potential role of the computer in lines for management in early Greek literature or data storage. MD. Shortliffe. and Use 2 Edward H. prognosis. and likely predictive value. both in the clini- cal world and in applications related to. analyzing. tory and X-ray studies. the ideas of ill health and its • How are clinical data used? treatment have been wedded to those of the • What are the drawbacks of the traditional observation and interpretation of data. J.).net . you should know the 2. the health care process. sensitivity. biology and human genetics. FACP. FACMI the basis for categorizing the problems a patient Laboratory of Computer Science (Harvard Medical School and Massachusetts may be having or for identifying subgroups General Hospital). Octo Barnett After reading this chapter. Biomedical Informatics. it is Departments of Biomedical Informatics at Columbia University and Arizona State University. MA. Shortliffe and G. within a population of patients. it is clear that gathering edge base? data and interpreting their meaning are central to • How are data collection and hypothesis gen. E. and computers in biomedicine.J. Biomedical Data: Their Acquisition. In fact. With the move toward eration intimately linked in clinical diagnosis? the use of genomic information in assessing indi- • What are the meanings of the terms prevalence. 3 and 4 and the New York Academy of Medicine. PhD (*) If data are central to all health care. A textbook on informatics will accordingly a clinical database? refer time and again to issues in data collection. 22). This chapter lays the foundation for this recurring set of issues that is pertinent to all aspects of the use of information. NY.

they often turn to the chart or electronic dipstick. Human beings shortsighted to think that it was adequate to encode the can glance at a written blood pressure value year of an event with only two digits. urine another. clinicians need on 14FEB20131) to share descriptive information with others.. history of rheumatic patient assessment and management.g. a red blood cell count. pro- is difficult to record them in ways that convey fession is steel worker) their full meaning. ketoacidosis (acid production due to poorly con- ence between diastolic (while the heart cavities trolled blood sugar levels) or the continuous mea- are beginning to fill) and systolic (while they are surements of mean arterial blood pressure for a contracting) blood pressures is important for deci- sion making or for analysis. however. Time can particularly complicate the assess- We consider a clinical datum to be any sin. information about the details of a observations made . even from one clinician to 4. Such data may involve several different cal interview. a past history of adequate—e. or a blood pressure reading. method for data storage and analysis. the date of the observation is reading. In others. tary view as a single data point and the decom- lem or most effectively to treat the problem that posed information about systolic and diastolic has been diagnosed.6 °F. No clinician disputes the importance of 2. The time of the observation (e. fever. Barnett is needed and what actions should be taken to and easily make the transition between its uni- gain a greater understanding of a patient’s prob... If the differ.g. a variety of more subtle types of data design of medical data systems. The method by which the observation was When they cannot interact directly with one made (e. quently and the data collected need to be iden- spective whether a single observation is in fact tified in time with no greater accuracy than a more than one datum. heart size on chest X-ray film) cise role of these data and the corresponding 3. The parameter being observed (e. patient report. Such dual views can be much more It is overly simplistic to view data as the col. gle observation of a patient—e. the frequent blood setting where knowledge that a patient’s blood sugar readings obtained for a patient in diabetic pressure is normal is all that matters. health record for communication purposes. liver such observations in decision making during size. laboratory instrument). in outpatient clinics or private rubella. 5. Although laboratory a data model for computer-stored medical data test results and other numeric data are often accordingly becomes an important issue in the invaluable.M. another.g.O. pressures. yet the pre.ketabpezeshki. decision criteria are so poorly understood that it weight is 70 kg. Thus. The patient in question room. the blood 1 pressure reading is best viewed as two pieces Note that it was the tendency to record such dates in computers as “14FEB12” that led to the end-of-century of information (systolic pressure = 120 mmHg. clinical data are multiple obser- seems to be avoiding a question during the medi.g. It was diastolic pressure = 80 mmHg)..40 E. however. thermometer.H. As the blood offices where a patient generally is seen infre- pressure example shows.g. (021) 66485438 66485457 www. unless they umns of numbers or the monitored waveforms are specifically allowed for in the design of the that are a product of our increasingly technologi. a single datum gen- severity that an experienced clinician will often erally can be viewed as defined by five elements: have within a few moments of entering a patient’s 1. the observation patient’s symptoms or about his family or eco. Shortliffe and G. vations. 2:30 A.. Despite these limitations. urine sugar value. it is a matter of per. difficult for computers. complexities that we called the Year 2K problem. or the subjective sense of disease points in time. ment and computer-based management of data. or both. of the same patient parameter made at several nomic setting. temperature is 98. The idea of cal health care environment. The value of the parameter in question (e... a temperature In some settings. minute-to-minute varia- might well be recorded as a single element in a tions may be important—e. may be just as important to the delivery of opti.g.g. A blood pressure of 120/80 calendar date. If a clinical datum is a single observation mal care: the awkward glance by a patient who about a patient.

Such narrative data were child and thinks that she hears a heart mur. the general review of her pediatrician called the disease nor whether systems that is part of most evaluations of new anyone thought that he or she had scarlet fever. 2 As is discussed in Chap. taken in the arm or leg? Was the patient lying or standing? Was the pressure obtained just after exercise? During sleep? What kind of recording 2. It is rare that an observation—even Narrative data account for a large component one by a skilled clinician—can be accepted of the information that is gathered in the care of with absolute certainty. or modifiers.1. also other narrative descriptions such as reports tional data that will either confirm or eliminate of specialty consultations.2 Electronically stored transcriptions variety of possible responses to deal with incom. the patient’s description of examples: his or her present illness. The idea of trade-offs in data collec- the circumstances under which a data point was tion thus becomes extremely important in guid- obtained. For example. 12. health care organizations are ever. For example. ues of data. Increasingly. ies for inclusion in paper or electronic medical • A confused patient is able to respond to simple records. (021) 66485438 66485457 www. because the costs of data collection must increasingly relying on electronic health records to the be considered. Could he or she have had scarlet medical record. there is a broad range of data types in the practice Two patients with the same basic problem or of medicine and the allied health sciences. the narrative summaries • A radiologist looking at a shadow on a chest are dictated and then transcribed by typists who X-ray film is not sure whether it represents produce printed summaries or electronic cop- overlapping blood vessels or a lung tumor. Storage. This solution is not always appropriate. including responses to • An adult patient reports a childhood illness focused questions from the physician. textual data to numerical tions for their problem. images. The electronic versions of such reports questions about his or her illness. drawings. The same is true of the patient’s fever? The patient does not know what his or social and family history.1). how. traditionally handwritten by clinicians and then mur—but is not certain because of the patient’s placed in the patient’s medical record (Fig. Consider the following patients. can be of crucial The examples in the previous section suggest that importance in the proper interpretation of data.2 Biomedical Data: Their Acquisition. sometimes called contexts. revealed by careful measurements. 2. but under can easily be integrated into electronic health the circumstances the physician is uncertain records (EHRs) and clinical data repositories how much of the history being reported is so that clinicians can access important clinical reliable. They symptom often have markedly different explana. 3 and 4. or wasteful blood pressure due to failure of the heart muscle). and in their patient histories and physical examinations but interpretation. patients. loud wheezing. and Use 41 patient in cardiogenic shock (dangerously low be expensive. genetic information. The additional observation might exclusion of printed records. was the blood pressure ing health care decision . and the clinician’s report of physical • A physician listens to the heart of an asthmatic examination findings. generally with fevers and a red rash in addition to joint is gathered verbally and is recorded as text in the swelling. and even photographs or other A related issue is the uncertainty in the val.1 What Are the Types of Clinical device was used? Was the observer reliable? Data? Such additional information. there are a available. One technique is to collect addi. information even when the paper record is not As described in Chaps. recorded assessment of the modifiers of that problem. surgical procedures. risky for the patient. of dictated information often include not only plete data. range from narrative. the concern raised by the initial observation. of time during which treatment could have been It often also is important to keep a record of instituted. methods. the uncertainty in them. signals.ketabpezeshki. however.

in which recorded observations Some narrative data are loosely coded with reflect the stereotypic examination process taught shorthand conventions known to health .ketabpezeshki. and hospital.1 Much of the information gathered during a physician–patient encounter is written in the medical record pathologic examinations of tissues. (021) 66485438 66485457 www. for example. particularly data collected during the physical ization summaries when a patient is discharged.O.42 E. 2. Barnett Fig. to all practitioners. Shortliffe and G. It is common.H. examination.

however. This range of data-recording next significant? Was the patient weighed on the conventions presents significant challenges to same scale both times (i. or photo- sodium level to two-decimal-place accuracy? Is graphic images (of the patient or of radiologic a 1-kg fluctuation in weight from 1 week to the or other studies).com . vations. Can physi. 19). instruments rather than changes in the patient)? In some fields of medicine. and Reactive to Light systems. abbreviations with meanings.e.” and “failure to thrive. also is common for physicians to draw simple ings depending on the context in which they are pictures to represent abnormalities that they have used. a sketch is a concise what is commonly called a heart attack). tant category of data. of data recording.. When ized training know how to interpret the data- such numerical data are interpreted. a conventional ideas of the health care team evoke graphical tracing frequently is included. Although When such data are stored in medical records. For example. with a images of coworkers treating ill patients. “MI” can mean “mitral insuf. and Accommodation (the process of focusing on Visual images—acquired from machines or near objects). and certain measurements can understand. erwise heterogeneous conditions that together A glance at a medical record will quickly reveal characterize a simple concept about a patient. For example. The notations may be highly cians distinguish reliably between a 9-cm and a structured records with brief text or numerical 10-cm liver span when they examine a patient’s information. It are not standard and can have different mean. the of such standardization is often unsuccessful.1. 2.” sketched by the physician—are another impor- Note that there are significant problems asso. The range goes from hand- numeric values. Many tographs of skin lesions are obvious examples.” Such standard. a tracing Health data on patients and populations are gath- of the electrical activity from a patient’s heart. Radiologic images or pho- ciated with the use of such abbreviations.” “pain relieved by antacids caring for patients. Round. for issue of precision becomes important. This clear challenges in determining how such data encoded form indicates that the patient’s “Pupils are best managed in computer-based storage are Equal (in size).2 Biomedical Data: Their Acquisition. notation to cryptic symbols that only specialists ture and pulse rate). Many way of conveying the location and size of a nod- hospitals try to establish a set of “acceptable” ule in the prostate gland (Fig. These include such parameters written text to commonly understood shorthand as laboratory tests. machine- abdomen? Does it make sense to report a serum generated tracings of analog signals. observed. idea of data is inextricably bound to the idea Complete phrases have become loose stan. and Use 43 to find the notation “PERRLA” under the eye written interpretation of its meaning. communication to themselves and other people.ketabpezeshki. analog data in the form of continuous signals are particularly 2. Perhaps the best-known example is an electrocardiogram (ECG). Storage. hand-drawn sketches. the team (021) 66485438 66485457 www. such drawings may serve as a basis for ficiency” (leakage in one of the heart’s valves) comparison when they or another physician next or “myocardial infarction” (the medical term for see the patient. few physicians without special- taken during the physical examination. but the enforcement As should be clear from these examples. vital signs (such as tempera.3). They must record their obser- or milk. for later text notation as a form of summarization for oth. care personnel are taught from the outset that it is nel. the recording conventions of an ophthalmologist. the wide variety of data-recording techniques Many data used in medicine take on discrete that have evolved. Examples include “mild dyspnea (shortness crucial that they do not trust their memory when of breath) on exertion.2 Who Collects the Data? important (see Chap. ered by a variety of health professionals. There are examination in a patient’s medical record. could the different the person implementing electronic health record values reflect variation between measurement systems. Physicians and other health dards of communication among medical person. example (Fig. 2.2). as well as the actions they have taken ized expressions are attempts to use conventional and the rationales for those actions.

2. and the system review.O. nurses play a central role in making observations and recording them for future reference. 15. In addition. The data that they gather contribute to nursing care plans as well as to the assessment of patients by physi- cians and by other health care staff. cially in the hospital setting. 2. to understanding of pertinent psychosocial issues. espe- Fig. Thus. family and admissions personnel gather demographic and social information. tion and insights that contribute to proper diag- mologist has used nosis. In both outpatient and hospital settings. they generally decide what additional data to collect by ordering laboratory or radiologic studies and by observing the patient’s response to therapeutic interventions (yet another form of data that contributes to patient assessment). data collection and recording are a central mation systems in contributing to patient care part of its task. Office staff and on the chief complaint. Physical or respiratory (021) 66485438 66485457 www. and examination of the patient. The role of infor- per . data collection and interpretation. tasks such as care planning by nurses is the sub- Physicians are key players in the process of ject of Chap. Most physicians trained in other specialties would relationships with patients that uncover informa- have difficulty deciphering the symbols that the ophthal. Shortliffe and G.44 E.H. history taking. or to proper planning of therapy or dis- has much broader responsibilities than treatment charge management (Fig. nurses often build tion. Barnett Fig.2 A physician’s hand-drawn sketch of a prostate nodule. nurses’ training includes instruction in careful and accu- rate observation.ketabpezeshki. Because nurses typically spend more time with patients than physicians do. A drawing may convey precise information more easily and compactly than a textual description examine the patient.3 An ophthalmologist’s report of an eye examina. past illnesses. collecting pertinent data and recording them during or at the end of the visit. They converse Various other health care workers contribute with a patient to gather narrative descriptive data to the data-collection process.4). 2. They financial information.

such as blood or urine. Traditional data-recording blood pressure. With the advent of compre- record the results for later use by physicians and hensive EHRs (see Chap. or even smart phones. as we move to “paperless” records whereby all ing data and report their findings to the patients’ access to information is through computer work- physicians. ECG machines. the broadly trained team of individuals who be read by an operator and then recorded in the are involved in a patient’s care. sphygmomanometers for taking see Chaps. 12). Sometimes a trained specialist multiple providers to access a patient’s data and must interpret the output.2 Uses of Health Data ent individuals employed in health care settings gather. about their medications or about drug allergies and then monitor the patients’ use of prescription drugs. Pharmacists may interview patients . and make use of patient data in Health data are recorded for a variety of purposes. This ability is just one of the ways in which nurses play a key role in data collection and recording (Photograph courtesy of Janice Anne Rohn) therapists record the results of their treatments the devices feed their results directly into com- and often make suggestions for further manage. there are the technological devices care of the patient from whom they were obtained. 16 and 26). Sometimes such a device produces a reasonably well when care was given by a single paper report suitable for inclusion in a traditional physician over the life of a patient. However. to communicate effectively with one another (021) 66485438 66485457 www. record. Clinical data may be needed to support the proper Finally. however. Radiology technicians perform X-ray such data summaries may no longer be required examinations. monitoring equipment in ety through the aggregation and analysis of data intensive care units. Laboratory personnel perform tests on or formatted for electronic storage as well as biological samples. but they also may contribute to the good of soci- imaging machines. These relationships may allow the nurse to make observations that are missed by other staff. and reported on paper.4 Nurses often develop close relationships with patients. and the need for patient’s chart. their work. that generate data—laboratory instruments. clinical research or public health assessments. and Use 45 Fig. Increasingly. puter equipment so that the data can be analyzed ment.2 Biomedical Data: Their Acquisition. 2. and spirometers for testing lung techniques and a paper record may have worked function). Storage. hand-held tablets. medical record.ketabpezeshki. Sometimes the device indicates given the increased complexity of modern health a result on a gauge or traces a result that must care. radiologists interpret the result. and measurement devices regarding populations of individuals (supporting that take a single reading (such as thermometers. many differ. the printing of nurses. 2. As these examples suggest.

and record for individual patients is of inestimable demographic information)? importance to clinical research.2. Another problem occurs because tra.ketabpezeshki. • What symptoms has the patient reported? When did they begin. pertinent family.O. What was the nature of the disease or adequately supports optimal care of individual symptom? patients. and what has provided relief? 2. Shortliffe and G. other diseases that coexist or large numbers of patients. . medical records are intended to pro. especially over the last half cen- observed.g. the physician may learn how a specific patient tends to respond to pain or how family interactions tend to affect the patient’s response 2. the historical have resolved. On the other hand. Just as a labo.1 Create the Basis to disease). skilled health care personnel that has always vide a detailed compilation of information about been of fundamental importance in the genera- individual patients: tion of new knowledge about patient care. disorders—possibly over many years.2 Support Communication • What physical signs have been noted on Among Providers examination? • How have signs and symptoms changed over A central function of structured data collection and time? recording in health care settings is to assist person- • What laboratory results have been. Although laboratory research has contributed ratory notebook provides a record of precisely dramatically to our knowledge of human disease what a scientist has done.2. or are now. Most patients who have significant medical • What radiologic and other special studies have conditions are seen over months or years on sev- been performed? eral occasions for one or more problems that • What medications are being taken and are require ongoing evaluation and treatment. ized (e. the value of some for the Historical Record experiments may be derived only by pooling of data from many patients who have similar prob- Any student of science learns the importance lems and through the analysis of the results of of collecting and recording data meticulously various treatment options to determine efficacy. What was the outcome of that treatment? have made clinical research across populations As is true for all experiments. it is careful observation and recording by sion points. 12 and 16. the experimental data and treatment. when carrying out an experiment. 2. Barnett through the chart. The lessons learned regard. with the goal of It was once common for patients to receive answering three questions when the experiment essentially all their care from a single provider: is over: the family doctor who tended both children and (021) 66485438 66485457 www. what has seemed to aggravate them.. inher. Electronic to learn from experience through careful obser- record keeping offers major advantages in this vation and recording of data. thus. the care given to • What is the reasoning behind the management a patient is less oriented to diagnosis and treatment decisions? of a single disease episode and increasingly Each new patient problem and its management focused on management of one or more chronic can be viewed as a therapeutic experiment. ently confounded by uncertainty. the paper record no longer 1. one purpose is of patients extremely cumbersome. as we discuss in more detail later in this in a given encounter may be highly individual- chapter and in Chaps. What was the treatment decision? ditional paper-based data-recording techniques 3. nel in providing coordinated care to a patient over available? time. tury.46 E. Given there any allergies? the increasing numbers of elderly patients in many • What other interventions have been undertaken? cultures and health care settings.H. We • What is the patient’s history (development of learn from the aggregation of information from a current illness. and the rationale for intermediate deci.

Shared access to a paper chart (Fig.2 Biomedical Data: Their Acquisition. Now the record not only contains observations by a physician for reference on the next visit but also serves as a communication mechanism among physicians and other medical personnel. especially in smaller communities. or discharge plan- ners. the emer- events.6 Today similar communication sessions occur around a computer screen rather than a paper chart (see Fig. sometimes conferring as they look at the same computer screen (Fig.5) (Photograph courtesy of James J.ketabpezeshki. nursing staff. social workers. Such gence of subspecialization and the increasing provision of care by teams of health profes- sionals have placed new emphasis on the cen- tral role of the medical record.5) is now increasingly being replaced by clinicians accessing electronic records. patients receive care over time from a variety of physicians—col- leagues covering for the primary physician. Fig. Storage.5 One role of the medical record: a communica- tion mechanism among health professionals who work or a managed care organization’s case manager. and Use 47 adults. often seeing the patient over many or all doctors nonetheless kept records of all encoun- the years of that person’s life. 2. family and sharing in many of the patient’s life In the world of modern medicine. or specialists to whom the patient has been referred. We tend to pic. Cimino) (021) 66485438 66485457 www. 2. ters so that they could refer to data about past ture such physicians as having especially close illnesses and treatments as a guide to evaluating relationships with their patients—knowing the future care issues. radiology technicians.6).com . 2. 2. such as physical or respiratory therapists. 2. In many outpatient settings. together to plan patient care (Photograph courtesy of It is not uncommon to hear complaints from Janice Anne Rohn) patients who remember the days when it was Fig.

Many physicians will ask a patient 2. slight increase in temperature that accompanies ovulation and thus may discern the days of maxi- mum fertility. Women who want to get pregnant often In these cases. injections of gamma globu- diseases rather than care for trauma or acute lin to protect people from hepatitis. When eas- Problems ily accessible in the record (or from the patient). glance. future development of disease.. Such graphs are increasingly created and data on interventions that have been performed to displayed for viewing by clinicians as a feature of prevent common or serious disorders.48 E. or reduce 2.2. and dietary changes to lower tance of the medical record in ensuring quality cholesterol). This idea is of par. keep similar records of body temperature. safe sex practices. Shortliffe and G. 2. such data can prevent unnecessary treatments (in Providing high-quality health care involves more this case. interactions between patients and their doctors.ketabpezeshki.4 Record Standard Preventive to keep such graphical records so that they can Measures later discuss the data with the patient and include the record in the medical chart for ongoing refer- The medical record also serves as a source of ence. Physicians are sensitive ing cessation. Barnett possible to receive essentially all their care from the interventions involve counseling or educa- a single physician whom they had come to trust tional programs (for example.g. a basis for specific patient education or preven.H. Sometimes a patient’s medical record. in which chronic larly high risk (e. When a patient comes to his local hospital emer- gency room with a laceration. By tive of later symptomatic disease. regarding smok- and who knew them well. women can identify the tions have an opportunity to develop fully. the physicians routinely check for an indication of when he most 2. ventions include immunizations: the vaccina- ing of the details and logic of past interventions tions that begin in early childhood and continue and ongoing treatment plans. and providing and weight may have limited use by themselves. medication. It also requires educating patients about the ways in which their environ- ment and lifestyles can contribute to. such as diet. it is the trend in such data points observed over tive interventions.2.5 Identify Deviations the risk of. such administered when a person will be at particu- as the one in the United States.3 Anticipate Future Health recently had a tetanus immunization.O. and elevated serum cholesterol levels. including special treatments ticular importance in a health care system. It is accordingly common for such ongoing risk assessment in our society are such parameters to be recorded on special charts routine monitoring for excess weight. from Expected Trends Similarly. throughout life. health problems. high blood or forms that make the trends easy to discern at a pressure.2. values on special charts. administered infections increasingly dominate the basis for before travel to areas where hepatitis is endemic). Perhaps the most common examples of a medical problem. data gathered routinely in the ongoing care of a patient may suggest that he or she is at Data often are useful in medical care only when high risk of developing a specific problem even viewed as part of a continuum over time. optimal care measuring temperature daily and recording the requires early intervention before the complica. a repeat injection) that may be associ- than responding to patients’ acute or chronic ated with risk or significant cost. or months or years that may provide the first clue to exercise. Clinical data therefore are normal growth and development by pediatricians important in screening for risk factors. to this issue and therefore recognize the impor. measures for stopping drug abuse. Single data points regarding height patients’ risk profiles over . Other important preventive inter- and continuity of care through adequate record. following (Fig.7). (021) 66485438 66485457 www. abnormal data may be predic. An though he or she may feel well and be without example is the routine monitoring of children for symptoms at present. normally (Source: National Center for Health Statistics (021) 66485438 66485457 www.ketabpezeshki. 2. http:// time that indicate whether development is progressing www.2 Biomedical Data: Their Acquisition. it is the changes in values over Disease Prevention and Health Promotion (2000).com .cdc. Single data points in collaboration with the National Center for Chronic would not be useful. and Use 49 Fig.7 A pediatric growth chart.

11 and 26) is a common method by which specific clinical questions are addressed experi. By pooling such data. Barnett 2. • Can I find the medical record in which they vide a record of how each patient fared under are recorded? (021) 66485438 66485457 www. 2. for its intended uses. data do not exist in a gen. The delivery of health care unless they are recorded. Thus. 2. Such results then help investiga- the presumed diagnosis for a patient and the tors to define the standard of care for future choice of management. type. the analysis of large patient data sets even when erally useful form unless they are recorded. and heart disease. the entered the study or on the details of how patients chart generally should describe and justify both were managed.g. record to which the courts can refer if necessary.6 Provide a Legal Record treatment and precisely how the treatment was administered. gender. for example. As the study • Can I find the data I need when I need them? progresses. RCTs typically involve the random Issues assignment of matched groups of patients to alter- nate treatments when there is uncertainty about Recall. Thus.2. however.7 Support Clinical Research 2. Their optimal use depends on positive responses age. other pulmonary problems. it is only by The preceding description of medical data and formally analyzing data collected from large their uses emphasizes the positive aspects of numbers of patients that researchers can develop information storage and retrieval in the record. to the following questions: lems) are measured and recorded. first.50 E. is based on irre- delivered. data are gathered meticulously to pro. and validate new clinical knowledge of general All medical personnel. 1).com .3 Weaknesses of the Traditional Medical Although experience caring for individual Record System patients provides physicians with special skills and enhanced judgment over time. coexisting medical prob. quickly learn applicability.ketabpezeshki. We emphasized earlier the importance of Medical knowledge also can be derived from recording data. ties that greatly limit the record’s effectiveness ered from populations of patients (see Chap.2. The medical record is the founda. patients with the same or similar problems. unsubstantiated memories of what they observed Much of the research in the field of epidemiology or why they took some action are of little value in involves analysis of population-based data of this the courtroom. variables that might affect a patient’s course (e.. The the patients were not specifically enrolled in an legal system stresses this point as well. sometimes Another use of health data. a well-maintained record is a futable statistics derived from large populations source of protection for both patients and their of individuals with and without lung cancer. that data cannot effectively serve the how best to manage the patients’ problem. the cal difference among the study groups depending responsible individual must sign most of the clin. Providers’ RCT. once they are charted after years of experimentation (depending on the and analyzed. physicians. Our knowledge of the risks associated with tion for determining whether proper care was cigarette smoking. is as the foundation for a legal time course of the disease under consideration).H. often referred to as retrospective studies.O. In addition. another use of clinical data is that use of the traditional paper record is compli- to support research through the aggregation and cated by a bevy of logistical and practical reali- statistical or other analysis of observations gath. researchers may be able to demonstrate a statisti- The medical record is a legal document.1 Pragmatic and Logistical mentally. Shortliffe and G.3. on precise characteristics present when patients ical information that is recorded. weight. A randomized clinical trial (RCT) (see also Chaps. in fact.

it is not surprising that buried on a desk. For example: professional still must provide patient care. creates situations in which people answer such • When a chart is unavailable.ketabpezeshki. • The patient’s paper chart may be unavailable Thus. typically is provided. or it may have been located. the traditional paper record to effective use of the chart (Fig. their organization in the chart may lead the user records grow so large that the charts must to spend an inordinate time searching for the be broken up into multiple volumes. or office (Fig.8 A typical storage room for medical records. the health care questions in the negative.8). 2. It is not uncommon to hear one physi- them? cian asking another as they peer together into • Can I update the data reliably with new obser. When a data. it may have been misplaced despite nation reveals. 2. It basing their decisions instead on what the may be in use by someone else at another patient can tell them and on what their exami- location. and similarly clear why such paper repositories are being replaced as EHRs increas- ingly become the standard (Photograph courtesy of Janice Anne Rohn) (021) 66485438 66485457 www. In a large institution with thousands taken by someone unintentionally and is now of medical records. Storage.2 Biomedical Data: Their Acquisition. only the most recent volume histories. It is not surprising that charts sometimes were mislaid. especially in the massive paper charts hospital clinic or emergency room orders the of patients who have long and complicated patient’s chart. when the health care professional needs it. that the actual chronology of management is The data may have been known previously disrupted in the record. Poor diseases are seen over months or years. such loose notes often fail to make it to the • Once the chart is in hand. but never recorded due to an oversight by a • When patients who have chronic or frequent physician or other health professional. inclusion in the chart—when the chart is clinic. Old but pertinent data Fig. it might still be patient’s chart or are filed out of sequence so difficult to find the information required. They then write a note for the record-tracking system of the hospital. and Use 51 • Can I find the data within the record? • Once the health care professional has located • Can I find what I need quickly? the data. providers make do without past data. a five?” “Whose signature is that?” Illegible ments for future access by me or other people? and sloppy entries can be a major obstruction All too frequently. a chart: “What is that word?” “Is that a two or vations in a form consistent with the require. he or she may find them difficult to • Can I read and interpret the data once I find read. 2.9).com .

famil- iar paper form.3.9 Written entries are standard in paper records.H. and consulting physicians and by the nursing an early volume may be mistaken for the staff.” For ward quickly in the laboratory section of the complicated procedures. often with color-coding. 2. the same data often are chart to find the most recent urinalysis sheet and summarized in brief notes by radiologists in the to check at a glance the bacterial count. Notes that cannot be interpreted by other people due to illegibility may cause delays in treatment or inappropriate care—an issue that is largely eliminated when EHRs are used may be in early volumes that are stored offsite tioned in notes written by the patient’s admitting or are otherwise unavailable. The prob- narrative part of the chart. A similar inefficiency occurs because of a tension between opposing goals in the design of reporting forms used by many laboratories. 2. yet handwritten notes may be illegible. Shortliffe and . paper that report only a single data element. chart. 2. Barnett Fig. In addition. it becomes increas- for this reason that more and more hospitals. This standard radiology reporting form. record systems offer solutions to all these practi. accordingly.2 Redundancy and Inefficiency Most health personnel prefer a consistent. Alternatively. which they enter at the lem is that such forms typically store only sparse time of studies because they know that the formal information. complicates the chart’s logistical cal problems in the use of the paper record. misleading its users and recording such information multiple times in dif- resulting in documents being inserted out of ferent ways and in different locations within the sequence. ingly difficult to locate specific patient data as the health systems. Furthermore. For example. which is filed knowledge allows the physician to work back- in the portion of the chart labeled “X-ray.52 E. (021) 66485438 66485457 www.ketabpezeshki. the combined bulk of these notes acceler- As described in Chap.O. For example.10). 1 and 27). health it helps them to find information more quickly professionals have developed a variety of tech. hours to track down. nance (see Chaps. The predictable implementing EHRs–further encouraged in the result is that someone writes yet another redun- US by Federal incentive programs that help to dant entry. 12. the result low paper and records the bacteria count half- of a radiologic study typically is entered on a way down the middle column of the form. because To be able to find data quickly in the chart. the study results often are men. It is clearly suboptimal if a rapidly report will not make it back to the chart for 1 or 2 growing physical chart is filled with sheets of days. a physician may know niques that provide redundant recording to match that a urinalysis report form is printed on yel- alternate modes of access. summarizing information that it took cover the costs of EHR acquisition and mainte. electronic health ates the physical growth of the document and. (Fig. It is management. Although there may be good reasons for most recent volume. and individual practitioners are chart succumbs to “obesity”.

ketabpezeshki. medical consultation service notice that patients Note the distinction between retrospective receiving a certain common oral medication for chart review to investigate a question that was diabetes (call it drug X) seem to be more likely to not a subject of study at the time the data were have significant postoperative hypotension (low collected and prospective studies in which the blood pressure) than do surgical patients receiv. be discussed shortly. familiar format 2.3 Influence on Clinical Research One efficient way to follow up on their theory from existing medical data would be to exam- Anyone who has participated in a clinical research ine the hospital records of all patients who have project based on chart review can attest to the diabetes and also have been admitted for surgery. For all the reasons described in Chap. for example. 11 and 26).10 Laboratory reporting forms record medical data in a consistent. ments. that physicians on a and data gathering) might well be appropriate. lect future data that are relevant to the question ences postoperative blood pressure—on only a under consideration (see also Chaps. Observers often when admitted and (2) whether they had postop- wonder how much medical knowledge is sitting erative hypotension. patients (1) whether they were taking drug X nerable to transcription errors. 1. esis—from unintentionally skewing the results (021) 66485438 66485457 www. If the statistics showed that untapped in paper medical records because there patients receiving drug X were more likely to is no easy way to analyze experience across large have low blood pressure after surgery than were populations of patients without first extracting similar diabetic patients receiving alternate treat- pertinent data from those charts. tediousness of flipping through myriad medical The task would then be to examine those records records. Storage. (difficult and arduous with paper charts as will it is arduous to sit with stacks of patients’ charts. a controlled trial (prospective observation Suppose. but subject to automated extracting data and formatting them for struc.2 Biomedical Data: Their Acquisition. and Use 53 Fig. few recent observations. analysis in the case of EHRs) and to note for all tured statistical analysis. bound to be biased. so they decide Subjects are assigned randomly to different study to look into existing hospital records to see groups to help prevent researchers—who are whether this correlation has occurred with .3. having developed the hypoth- cient frequency to warrant a formal investigation. clinical hypothesis is known in advance and the ing other medications for diabetes. and the process is vul. 2. The doctors research protocol is designed specifically to col- have based this hypothesis—that drug X influ. however.

Automated was also administered (as well as ordered) and record systems introduce new opportunities for the admission history to see whether a routine dynamic responses to the data that are recorded treatment for diabetes.4 The Passive Nature of Paper surgical procedures. or implications for patient time-consuming. and the discharge summary. Medical matting the information for statistical analyses). For the same reason. ress notes. They cannot take an active role in look for it? The admission drug orders might responding appropriately to those implications. time of the surgery and to determine whether They are insensitive to the characteristics of the the patient had postoperative . follow. EHR systems have changed control.. 2. might be possible to use such an index to find all charts in which the discharge diagnoses included diabetes and the procedure codes included major 2. been hypotensive until after leaving the recov- sicians what therapy is being given (such as surgi. researchers subset of medical records dealing with surgical can use computer-based data retrieval and anal- patients who are also diabetic. Information age. They obviate the need to would have to identify the charts of interest: the retrieve hard copy charts.ketabpezeshki. ing notes from the ward need to be checked too. and for- chart selection can be overwhelming. they chart review process. difficult to locate. accuracy. it in prospective clinical trials (Chap. and tedious process and that people performing Returning to our example. but it would also be wise to check the our perspective on what health professionals medication sheets to see whether the therapy can expect from the medical chart.54 E. The researchers might start ble. and then methods such as retro. Records pile a list of patient identification numbers and have the individual charts pulled from the file The traditional manual system has another limi- room for review. Where should the researcher management. to the extent possi. retrieval. neither the with nursing notes from the recovery room or researchers nor the subjects know which treatment with the anesthesiologist’s datasheets from the is being administered. taken right up until the in them. As described in many of the chapters to patient entered the hospital. locating pertinent data. Such blinding is of course operating room. the studies are double blind. consider the prob. the charts each patient was receiving for diabetes at the sit waiting for something to be done with them. 12) is their ability to facilitate the hypotension question retrospectively. the task of relevant patients. but the patient might not have impractical when it is obvious to patients or phy. data recorded within their pages. double-blind studies are considered as well as vital signs sheets. physicians’ prog- the best method for determining optimal manage. Thus. i. ery room and returning to the ward. 26). and analysis make it feasible to (021) 66485438 66485457 www. tation that would have been meaningless until The researchers’ next task is to examine the emergence of the computer age. One of the great appeals of would encounter in addressing the postoperative EHRs (Chap. show what the patient received for diabetes Increasingly. records departments generally do keep indexes of Researchers can use similar techniques to har- diagnostic and procedure codes cross-referenced ness computer assistance with data management to specific patients (see Sect. First. it are prone to make transcription errors and to lems in paper chart review that the researchers overlook key data.H. but it is often impractical to carry It should be clear from this example that out such studies. So the nurs- cal procedures versus drug therapy). retrospective paper chart review is a laborious spective chart review are used. ment of disease. The researcher might com. Prospective.5. A manual each chart serially to find out what treatment archival system is inherently passive. Shortliffe and G. was not adminis. instead. randomized. computational techniques for data stor- tered during the inpatient stay.e. In a hospital record ysis techniques to do most of the work (finding room filled with thousands of charts.1). Barnett by assigning a specific class of patients all to one about hypotensive episodes might be similarly group.O.3. such as leg- Finding such information may be extremely ibility.

intended by the observers or recorders and those The issues that arise are practical as well as intended by the individuals retrieving informa- scientifically interesting. What is an “upper ers of EHRs have begun to grapple with ques. Initiative 2012). and the The debate has been accentuated by the intro- management of large amounts of genomic/pro. have logicians. In cancer.ketabpezeshki. 7). 26) and public health (Chap. and Use 55 develop record systems that (1) monitor their and tactical questions need answering regarding. Storage. the predicate calculus used by (Chap. These issues will undoubtedly influence the evolution of data systems and EHRs. The vast amounts of data example. Medicine is Researchers are finding that the amount of data remarkable for its failure to develop a widely that they must manage and assess has become so accepted standardized vocabulary and nomen- large that they often find that they lack either the clature.5 The Structure of Clinical Data The revolution in human genetics that emerged with the Human Genome Project in the 1990s Scientific disciplines generally develop a precise is already having a profound effect on the diag. contents and generate warnings or advice for pro.2 Biomedical Data: Their Acquisition. This problem. or only those components (e. prognosis. duction of computers for data management. Some suggest that the genetic these people question whether it is possible to material itself will become our next-generation introduce too much standardization into a field method for storing large amounts of data that prides itself in humanism. (2) provide automated that are already understood (Masys et al. NSF-NIH Interagency tion between medicine and the “hard” sciences. whether to store an entire genome viders based on single observations or on logi. sometimes until this problem is addressed (see Chap. cells.4 New Kinds of Data years ahead. cuits used by electrical engineers.. as well as the growth of personalized medicine. or (3) provide feedback on lines can occur. dubbed the “big data” problem. and the Resulting Challenges 2. and many observers believe that a true capabilities or expertise to handle the analytics scientific basis for the field will be impossible that are required. in the 2. Consider. develop. respiratory infection”? Does it include infec- tions regarding how they might be store an tions of the trachea or of the main stem bronchi? individual’s personal genome with the electronic How large does the heart have to be before we health record. 16). teomic or clinical/public-health data. can refer to “cardiomegaly”? How should we (021) 66485438 66485457 www. quality control. or the conventions for describing cir- created new challenges as well as opportunities. the precise defi- Chaps. New standards will be . 2012). including the flagging of poten. 2012). has gathered the Other people argue that common references to attention of government funding agencies as the “art of medicine” reflect an important distinc- well (Mervis 2012. tion or doing data analysis. for (Palotie et al. (Church et al. the universal language of chemistry that are generated in biomedical research (see embodied in chemical formulae. have because such machines tend to demand confor- accordingly become major research topics and mity to data standards and definitions. for example. 24 and 25).g. issues of data retrieval and analysis are con- opment by biomedical informatics scientists founded by discrepancies between the meanings (Ohno-Machado 2012). and treatment of disease accepted by all workers in the field. 2013). key opportunities for new methodology devel. terminology or notation that is standardized and nosis. for example. For example. genetic markers) cal combinations of data. where mutations in cell tially erroneous data. Data analytics. an individual may actually have patient-specific or population-based deviations many genomes represented among his or her from desirable standards. and that can be pooled from nitions and mathematical equations used by patient datasets to support clinical research physicists. Otherwise.

ization in terminology. 9th dards. Shortliffe and G.ketabpezeshki. case-mix analysis (determining the relative fre- Given the lack of formal definitions for many quencies of various disorders in the hospitalized medical terms. predefined vocabu. 16). both in acute care settings and when patients are Another kind of reporting involves the coding seen over long periods. sion of this standard. and the automatic summarization of data ics in their early stages. patients.O. This point is discussed in greater detail in in much of the world. an automated program will fail to public-health organizations. Because of the needs to know about health lary. The 10th revi- terms and data elements entered by the observ. 7. for Disease Control and Prevention in Atlanta Regardless of arguments regarding the “artis. Alzheimer’s disease. and aries of the abdomen well agreed on? What are similar health statistics that we tend to take for the time constraints that correspond to “acute.56 E. etc. Only each disease category) and for research. If EHRs are to as well as uniformly applied and accepted.g. Without a controlled. ICD10. For example. Hodgkin’s disease) that are not descriptive of the illness and We are used to seeing figures regarding the grow- may not be familiar to all practitioners? What do ing incidences of certain types of tumors. plus coding of certain procedures (e. one physician reporting requirements for hospitals (as well as might note that a patient has “shortness of breath. we would know much less about the wish to aggregate data recorded by multiple health status of the populations in various com- health professionals or to analyze trends over munities (see Chap. granted. cases of gonorrhea. another physician might note that she has example. their encoded logic must agnostic coding scheme called the International be able to presume a specific meaning for the Classification of Disease (ICD). it is remarkable that medical population and the average length of stay for workers communicate as well as they do.g. controlled terminology for biomedicine. Otherwise. 3 http://www.3 is currently in use ers. however. including a shared. personnel to communicate effectively is clear bacterial-resistance patterns. become dynamic and responsive manipulators of The World health Organization publishes adi- patient data. (021) 66485438 66485457 www. time. the efforts to develop health care-computing stan. Such codes are reported to state and to miscommunication. trends for populations and to recognize epidem- .. but if their accumulation required chart Imprecision and the lack of a standardized review through the process described earlier in vocabulary are particularly problematic when we this chapter. differences in type of surgery) that were performed during the intended meaning or in defining criteria will lead hospital stay. How are such data accumulated? Their ness” of abdominal pain? Is an “ache” a pain? role in health planning and health care financing What about “occasional” cramping? is clear. data to be useful. For Later. deaths we mean by an “acute abdomen”? Are the bound.icd10data. which code the data indicate that the patient had the same problem on to allow trend analyses over time. which deals in part with the multiple States a derivative of the previous version. data interpretation is inherently compli. there are various health- may be impossible.1 Coding Systems eponyms (e. and tuber- “dyspnea. federal health-planning and analysis agencies and potentially negative consequences for the and also are used internally at the institution for patients involved.” Unless these terms are designated as culosis generally must be reported to local synonyms. the codes must be well defined promised by miscommunication. (CDC) then pool regional data and report national tic” elements in medicine. For such occasionally is the care for a patient clearly com.5.” other public organizations) and practitioners. Both high-quality care of all discharge diagnoses for hospitalized and scientific progress depend on some standard. Barnett deal with the plethora of disease names based on 2. syphilis.. The Centers both occasions. the need for health as well as local trends in disease incidence. International Classification of Diseases. although in the United (Accessed 12/2/2012). improper interpretation. from influenza during the winter months.H.

902 Unspecified asthma with status asthmaticus J45. Storage. extrinsic allergic asthma.8).40 Moderate persistent asthma. exposure to tobacco smoke in the perinatal period (P96. tobacco use (Z72. and must be reported on the bills sub. allergic rhinitis with asthma. chronic obstructive asthma (J44. 7). nonallergic asthma Use additional code to identify: exposure to environmental tobacco smoke (Z77. American College of Patho- mitted to most insurance companies (Fig. asthma with chronic obstructive pulmonary disease (J44.891). 2.-).9). intrinsic nonallergic asthma. uncomplicated Moderate persistent asthma NOS J45. ICD9-CM is used by all nonmili.html.30 Mild persistent asthma.2 Mild intermittent asthma J45. accessed Volumes 1 and 2 (Source: Centers for Medicare and September 11. uncomplicated Severe persistent asthma NOS J45. the new diagnosis coding system Human Services.9) J45.11).22 Mild intermittent asthma with status asthmaticus J45. tobacco dependence (F17. 2.42 Moderate persistent asthma with status asthmaticus J45.2). and Use 57 Edition – Clinical Modifications (ICD9-CM). uncomplicated Mild intermittent asthma NOS J45. US Department of Health and taken from ICD-10-CM.990 Exercise induced bronchospasm J45. chronic asthmatic (obstructive) bronchitis (J44. allergic bronchitis NOS.21 Mild intermittent asthma with (acute) exacerbation J45.9). history of tobacco use (Z87.5 Severe persistent asthma J45.909 Unspecified asthma. logists 1982) and then merged with the Read Pathologists have developed another widely used Clinical Terms from the Great Britain to become J45 Asthma Includes: allergic (predominantly) asthma. ICD10/2013-ICD-10-CM-and-GEMs.3 Mild persistent asthma J45. uncomplicated Asthma NOS J45. atopic asthma.32 Mild persistent asthma with status asthmaticus J45. wood asthma (J67. uncomplicated Mild persistent asthma NOS J45.31 Mild persistent asthma with (acute) exacerbation J45. 2013) (021) 66485438 66485457 www.4 Moderate persistent asthma J45.901 Unspecified asthma with (acute) exacerbation J45.2 Biomedical Data: Their Acquisition. originally known is still transitioning to the new version (see as Systematized Nomenclature of Pathology Chap.ketabpezeshki.8). (SNOP).998 Other asthma Fig.51 Severe persistent asthma with (acute) exacerbation J45.90 Unspecified asthma Asthmatic bronchitis NOS Childhood asthma NOS Late onset asthma J45. lung diseases due to external agents (J60-J70).99 Other asthma J45.9 Other and unspecified asthma J45.50 Severe persistent asthma.0) Excludes: detergent asthma (J69. it was expanded to the Systematized tary hospitals in the United States for discharge Nomenclature of Medicine (SNOMED) (Côté coding. diagnostic coding scheme. http://www.cms.81). wheezing NOS (R06.20 Mild intermittent asthma. and Rothwell . hay fever with asthma.41 Moderate persistent asthma with (acute) exacerbation J45.22).gov/Medicare/Coding/ that is being developed as a replacement for ICD-9-CM.31).52 Severe persistent asthma with status asthmaticus J45. occupational exposure to environmental tobacco smoke (Z57. eosinophilic asthma (J82).991 Cough variant asthma J45. miner's asthma (J60).11 The subset of disease categories for asthma Medicaid Services. idiosyncratic asthma.

. 2. 7. In recent to describe what they observe.5 one hand. another practitioner may prefer to aggre. (021) 66485438 66485457 www. On the other suitably curated and organized so that they have hand. a hematologist (person who studies It generally can be regarded as the value of a spe- blood diseases) may want to distinguish among a cific parameter for a particular object (e. collective (singular) noun. for active tuberculosis—into a single category to analysis or display.ketabpezeshki. common structure that ties together the various rants emphasis here. however. 2010). is the (Accessed 12/2/2012). At the same for the codes’ development: health care personnel time. More details on a unified medical language system (UMLS). Shortliffe and G. Yet if physicians years. assumptions. It is similarly widely used in producing bills have worked for over two decades to develop for services rendered to patients.58 E. Bernstam et al. There is an from data. For example. of medicine.H. on the tional point that characterizes a relationship. be completely satisfactory. In some cases. and health plan- oped by the American Medical Association. Barnett SNOMED-CT (Stearns et al. The challenge is to learn how to meet all Current Procedural Terminology (CPT) (Finkel these needs. In this volume. those for they have been organized in some way..4 Another coding scheme. clinical research. meaning.g. system that is general enough to cover many dif. 2001). has been view the EHR as a blank sheet of paper on which assumed by the International Health Terminology any unstructured information can be written. Thus. a such schemes are provided in Chap.g. is derived simplify the coding and retrieval of data. a variety of hemoglobinopathies (disorders of the patient) at a given point in time. devel.ihtsdo. 1998) (see Chap. The term infor- structure and function of hemoglobin) lumped mation refers to analyzed data that have been under a single code in ICD8-CM. Consequently.2 The Data-to-Knowledge tings. Researchers at many institutions 1977). the Standards Development Organization. heuristics (strategic rules of thumb). information. through the formal or informal analysis (or inter- Such schemes cannot be effective unless pretation) of information that was in turn derived health care providers accept them. SNOMED versions have long permit. and knowledge functional status. ferent patients and the need for precise and unique and models—any of which may reflect the expe- terms that accurately apply to a specific patient and do not unduly constrain physicians’ attempts 5 Note that data is a plural term. 7). based in data they record will be unsuitable for dynamic Copenhagen. Data do not constitute information until gate many individual codes—e. then.5. These terms none of the common coding schemes is likely to are often used interchangeably. is the motivation vocabularies that have been created. the we shall refer to a datum as a single observa- granularity of the code will be too coarse.” Workers in the field have tried to ted coding of pathologic findings in exquisite clarify the distinctions among three terms fre- detail but only in later years began to introduce quently used to describe the content of computer- codes for expressing the dimensions of a patient’s based systems: data. although it is often erro- neously used in speech and writing as though it were a 4 http://www. In a particular clinical setting. ICD9-CM was derived from Spectrum a classification scheme developed for epidemio- logic reporting.. processing. . The historical roots of a coding system reveal themselves as limitations or idiosyncrasies when the system is applied in more general clinical set. determining charges for individual patients.O. Knowledge. What war. support for SNOMED-CT. the developers of specific terminologies are need standardized terms that can support pooling continually working to refine and expand their of data for analysis and can provide criteria for independent coding schemes (Humphreys et al. it has more than A central focus in bio medical informatics is the 500 separate codes for describing tuberculosis information base that constitutes the “substance infections. (Blum 1986b.g. knowledge includes the results inherent tension between the need for a coding of formal studies and also common sense facts.

and how to decide whether they war- appropriate interpretation of these definitions rant formal recording. This analysis of differences in both style and problem solving that organized data (information) has produced a piece account for variations in the way practitioners of knowledge about the world. and summarizing report developed piece of data. inefficient. Selection and Use The observation that patient Brown has a blood pressure of 180/110 is a datum. physical els. A physician’s collect and record data for the same patient under belief that prescribing dietary restriction of salt is the same circumstances. and performing a physical examination may take vations without any summarizing analysis. as is the report that It is illusory to conceive of a “complete clinical the patient has had a myocardial infarction (heart data set. There can be marked interpersonal normal or low blood pressure. Thus. that knowledge as an aid to case-based problem What do we mean by selectivity in data col- solving. the statement that the patient has by a medical student and the similar process under- hypertension is an interpretation of several such taken by a seasoned clinician examining the same data and thus represents a higher level of informa. 22). these elements in the EHR provide informa. to record only those data that will be pertinent in including semantic links among knowledge justifying the ongoing diagnostic approach and in items. selectivity. to perform only analysis of organized data (information). Medical students tend to work from com- tion. subsequent analysis may determine reflect the selective collection and recording of that patients with high blood pressure are more data by the health care personnel responsible for likely to have heart attacks than are patients with the patient. A blood pressure of 180/110 mmHg is a raw examination. As input to a diagnostic decision aid. is a collection of . patient.2 Biomedical Data: Their Acquisition. and knowledge base provides sufficient structure. A knowledge base. on the assessing every new patient.6 Strategies of Clinical Data mary data and the resulting information. physicians in practice to spend this amount of time tion about the patient. the presence or absence of hypertension may physical tests to perform. An more than 1 h.” All medical databases. It clearly would When properly collated and pooled with other be impractical. and inappropriate for data. heuristics. creating records. ever. however. in which case the presence of hyper. Because they have not developed skills of tension is treated as a data item. collect. When researchers pool such data. and lenge for the neophyte is to learn how to ask only models that can be used for problem solving and the questions that are necessary. part of the chal- other hand. Such variations do not unlikely to be effective in controlling high blood necessarily reflect good practices. how. and additional data to be requested. how to inter- physicians in their decision making. and Use 59 rience or biases of people who interpret the pri. reflecting cess that often is viewed as a central part of the this distinction between knowledge bases and “art of medicine. and medical attack). are necessarily incomplete because they information. If the the examination components that are required. the computer itself may be able to apply guiding the future management of the patient. 2. Storage. after which students develop exten- EHR system is thus primarily viewed as a data. and pressure in patients of low economic standing much of medical education is directed at helping (because the latter are less likely to be able to physicians and other health professionals to learn afford special low-salt foods) is an additional per.ketabpezeshki. the process of taking a medical history A database is a collection of individual obser. ence between the first medical history. have interpreted their observations. sive reports of what they observed and how they base—the place where patient data are stored. Many decision-support systems have lection and recording? It is precisely this pro- been called knowledge-based systems. what observations to make. Knowledge at one level of An example of this phenomenon is the differ- abstraction may be considered data at higher lev. individual styles and the sometimes marked (021) 66485438 66485457 www. Note that the pret them.” an element that accounts for databases (see Chap. prehensive mental outlines of questions to ask. how to make them sonal piece of knowledge—a heuristic that guides (generally an issue of technique). depends on the context.

These known as the hypothetico-deductive approach hypotheses then serve as the basis for selecting (Elstein et al.6. FH. HPI history of results present illness. their data col. Social social history. At that point. the data collected with these initial ques- tions typically are recorded as the patient identi- 2. studies ROS review of systems. CC. 2. hearing the patient’s response to the first six or tation may be imbedded in an iterative process seven questions (Elstein et al. chief complaint. FH family disease Radiologic history.12 A schematic view with a problem Initial hypotheses of the hypothetico-deductive ID. Barnett distinctions among clinicians. comprises the set of possible diagnoses among directed selection of the next most appropriate which the physician must distinguish to deter- data to be collected. cal informatics research. at least the uncertainty is reduced to a satisfactory and to formalize the ideas so that they can better level).ketabpezeshki. derived from research As is shown.60 E. set of active hypotheses as the differential lection. or perhaps the need for (Shortliffe 2010) but also are providing insights routine care). idly on the nature of the problem.g. PE Select most accordingly likely diagnosis physical examination (021) 66485438 66485457 www. The of the patient’s problem. As is discussed each stage. 1978. PMH past Chronic medical history. not only are patient presents to the physician with some issue enhancing the teaching and practice of medicine (a symptom or disease. ROS hypothesis generation and no further care refinement. See text for full required Patient discussion. and physician to refine hypotheses about the source their medical records become more compact. ID patient dies Laboratory Refine hypotheses tests identification. HPI PE approach. with a few questions that allow one to focus rap- based decision-support tools. Shortliffe and G. and initial portion of Approach the history of the present illness. . medical students learn this process. a management. data collection begins when the activities in biomedical informatics. Studies have shown that an experienced physician will have an Studies of clinical decision makers have shown initial set of hypotheses (theories) in mind after that strategies for data collection and interpre. or therapeutic decision can be made.H..12. Physicians refer to the central idea is one of sequential.1 The Hypothetico-Deductive fication. PMH.12 clarifies this process. followed by data interpretation and the diagnosis for a patient. 1978). Patient presents Fig. Improved guidelines The diagram in Fig. an ongoing process of Patient is better. for such decision making. are central in biomedi. or how expert clinicians internalize this process. they are added to the growing data- with numerous clinical examples in Chaps. Treat patient ECG etc. be taught and explained. As shown in Fig. CC chief Observe complaint. the idea of selectivity implies an ongoing or refine the active hypotheses. disposition. it is proved to be true. answers to these additional questions allow the lection becomes more focused and efficient. Attempts to understand level of certainty (e. the differential diagnosis generation of hypotheses. The physician generally responds that suggest methods for developing computer. In the written report. iterated until one hypothesis reaches a threshold tion and interpretation. 2. leading to hypothesis. 3 base of observations and are used to reformulate and 4. The process of Ask questions Examine medical data collection and More questions patient treatment is intimately tied to HPI. This process is decision-making process that guides data collec. staged data col. As data are collected at mine how best to administer treatment.O. 2. Kassirer and Gorry 1978). As additional questions.

electroencephalograms.. partially integrated with the examination process. and Use 61 Note that the question selection process is tests that occasionally turn up new abnormalities inherently heuristic. computed tomog- observed expertise. ease is resistant to that therapy and that their phy- tions. or it may already be narrowed to a single diagnosis) then mean that the initial diagnosis was incorrect and serves as the basis for a focused physical exami. as taking the history of a patient’s present illness well as returning to the patient to ask further (Pauker et al. Treatments are administered. to help them to avoid missing important modalities). e. family their hypothesis list. as in the ical care that is accounting for an increasing pro- question-asking process. If patients do not drug reaction or allergy). social history. that physicians should consider alternate expla- nation. As the results of such studies become that follow the collection of information about available. expectations of what they will find on exami. it may mean that their dis- When physicians have finished asking ques. (021) 66485438 66485457 www.. Note data collected to measure finds important information that modifies the response to treatment may themselves be used to hypothesis list or modulates the treatment options synthesize information that affects the hypothe- available (e. and many others). physicians constantly revise and refine the chief complaint: past medical history. specific treatment. 2. solving approach. When physicians have completed the physi- tic problem solving is the playing of a complex cal examination. The patient may remain in a cycle of treatment nation or may have specific tests in mind that and observation for a long time.g. 1976). other body fluids. and other specialized tests (electro- issues that they might not discover when collect. cardiograms (ECGs). as shown in will help them to distinguish among still active Fig. raphy (CT) studies. the refined hypothesis list (which may sicians should try an alternate approach. Occasionally. once again guided by the current .12. expert chess play.g.ketabpezeshki. physicians are sufficiently certain tems in which the physician asks some general about the source of a patient’s problem to be able questions about the state of health of each of the to develop a specific management plan. Because it would require an be narrowed sufficiently for them to undertake enormous amount of time to define all the pos. Such testing is from a given board position. or any of a number of other imaging however. focused hypothesis. respond to treatment.2 Biomedical Data: Their Acquisition. In addition. Additional data gathering sible moves and countermoves that could ensue may still be necessary. however. ing data in a hypothesis-directed fashion when nerve conduction studies. if the patient reports a serious past ses about a patient’s illness. and a brief review of sys. but it is not guaranteed to collect every expect on the basis of the medical history alone. major organ systems in the body. portion of the health care community’s work (and directed examination is augmented with general an increasing proportion of health care cost). it is personalized and and generate hypotheses that the physician did not efficient. nuclear-imaging scans. A common example of heuris. or biopsy for how best to proceed. urine. Physicians have developed safety measures. physicians may well have nations for the patient’s problem. Differences among specimens). Thus. By this time. unexplained findings on examination Human beings use heuristics all the time in their may raise issues that require additional history decision making because it often is impractical taking. and the patient is the physician discovers entirely new problems or observed. This long cycle reflects the nature of hypotheses about diseases based on the ques. their refined hypothesis list may game such as chess. chronic-disease management—an aspect of med- tions that they have asked. Storage. sonograms. the asking of questions generally is or impossible to use an exhaustive problem. radiologic studies (X-ray examina- such heuristics account in part for variations in tions. ers develop personal heuristics for assessing the The options available include laboratory tests game at any point and then selecting a strategy (of blood. history. These measures tend to be questions or perform additional physical exami- focused in four general categories of questions nation. magnetic resonance scans. Once again. piece of information that might be pertinent. Ultimately.

or a CT scan of the head (all given disease or condition—is an important one. patient has a specific . and then examined under Data and Hypotheses the microscope) with grossly abnormal cells (called class IV findings) is never seen unless We wrote rather glibly in Sect. What might be the characteristics of ment in most areas of medicine. often account for differences of opinion Suppose a given datum is never seen unless a among collaborating physicians. diagnosis. Differences in Perhaps the clinical manifestation seldom the assessment of cost-benefit trade-offs in data occurs unless the hypothesis turns out to be true. are female. Although (e. patient is female. and variations among individuals in is that enough to explain hypothesis generation? their willingness to make decisions under uncer. they ease or disease category more frequently than it are discussed in greater detail in Chaps. (021) 66485438 66485457 www. makes the diagnosis. when a clinical finding makes a specific the process outlined in Fig. It is dependent not only eration? A simple example will show that such a on a significant fact base that permits proper simple relationship is not enough to explain the interpretation of data and selection of appropriate evocation process. When an observation evokes a hypothesis ger need therapy. The idea of sensitivity—the likelihood that a ered require.O.ketabpezeshki. however. diagnosis.1 about the the woman has cancer of the cervix or uterus. but if the data being consid. although having a particular test result Elstein et al. a feature is seen in one dis- only a brief introduction to these ideas here. female. collection. It costs nothing but time nant patient is female). be true.H. at the opening to the uterus. and treatment is inher. is the need for physicians to balance financial is pregnant is not immediately evoked. 1976.62 E. Shortliffe and G. Consider the hypothesis that follow-up questions and tests but also on the a patient is pregnant and the observation that the effective use of heuristic techniques that charac.2 The Relationship Between Papanicolaou’s stain. then but it will not alone account for the process of it may be preferable to proceed with treatment hypothesis generation in medical diagnosis. is in others. but the association is not absolute. a Pap smear (a smear of cells swabbed from the cervix. When a new patient is observed to be Another important issue. Thus. 2. Is that enough to explain hypothesis gen- ently knowledge-based. For example. or he or she may die.g.12 is oversimplified diagnosis come to mind). given datum will be observed in a patient with a nary angiography.. and treat. now we Such tests are called pathognomonic. treated with 2. Barnett Alternatively. the patient may recover and no lon. This idea seems to be a little closer to the mark. Clearly.6. tainty. the possibility that the patient 3. in the absence of full information. to say about that process as well. coro. Unfortunately. Pople 1982) and the tradi. addressed in Chap. 4. 3 and 4. for example. costs and health risks of data collection against female gender is a highly sensitive indicator of the perceived benefits to be gained when those pregnancy (there is a 100 % certainty that a preg- data become available. additional question. but it is not a good predic- to examine the patient at the bedside or to ask an tor of pregnancy (most females are not pregnant).6. all pregnant patients terize individual expertise. Not only need to ask: What precisely is the nature of that do they evoke a specific diagnosis but they also process? As is discussed in Chap. X-ray exposure. researchers immediately prove it to be true. it is generally applicable to the sumably has some close association with the process of data collection. few patients with the condi- tional probabilistic decision sciences have much tion actually have that finding). with a psychological orientation have spent much there are few pathognomonic tests in medicine time trying to understand how expert problem and they are often of relatively low sensitivity solvers evoke hypotheses (Pauker et al. that association? Perhaps the finding is almost Note that the hypothesis-directed process of always observed when the hypothesis turns out to data collection. We provide More commonly. 2. 1978. hypothesis. the observation pre- in many regards. of which have associated risks and costs). “generation of hypotheses from data”. (that is.

2 Biomedical Data: Their Acquisition. it population composed of cigarette smokers in the tends to evoke that disease during the diagnostic United States. the prevalence of lung or data-gathering process.. For any test diseases that can cause the observed abnormality. ease hypothesis. cal diagnosis thus involves both the evocation of The prevalence of a disease is simply the per. the patient has other diseases (and does not have the PV will be greater than the prevalence (also the suspected one) unless (1) the finding is called the pretest risk). we must introduce negative test). the prevalence of ease if it is generally not seen in patients who do the disease is low). To explain the basis for (the PV of a positive test) and PV− (the PV of a this confusion in more detail. but the chance increases if not have that disease. has lung cancer is low (i. updated probability of lung cancer than it would ing test results that evoke a disease hypothesis had the patient been selected from the population and that physician being willing to act on the dis. although they the post-test (updated) probability that a disease have made an observation that is highly specific is present based on the results of a test. The reflects the test . For example. for example. A pathognomonic observa. The task of diagnosis therefore patients who do not have infections will have nor. two additional terms: prevalence and predictive The process of hypothesis generation in medi- value. you may have realized that there is a chest X-ray report would result in an even higher substantial difference between a physician view. the prob- word used to describe this relationship is speci. ability that any given person in the United States ficity. cancer would be higher. then. but it does tend to evoke or support was selected) to a post-test probability that the hypothesis that an infection is present. and Use 63 For example. An observation is highly specific for a dis. who develop black lung from inhal- as can the use of the drug prednisone. but most ing coal dust. there is one PV if the test result This mistake is one of the most common errors of is positive and another PV if the test result is neg- intuition that has been identified in the medical ative. If the patient were a member of the an observation is highly specific for a disease. that lation but has a much higher prevalence among leukemia can raise the white blood cell count. it may still be more likely that observation supports the presence of a disease. however. the identical By now. Certainly it is true. An elevated white has a disease from the baseline rate (the preva- count therefore does not prove that a patient has lence in the population from which the patient an infection. as well as the prevalence of the disease. Yet even experienced physicians The predictive value (PV) of a test is simply sometimes fail to recognize that. If the observation tends to pathognomonic or (2) the suspected disease is argue against the presence of a disease. The PV of a positive test may have a prevalence of only 5 % in the gen. In this case. his or her chest X-ray examination shows a pos- tion is 100% specific for a given disease. coal miners. involves updating the probability that a patient mal white blood cell counts. These values are typically abbreviated PV+ decision-making process. black-lung than infections that elevate a patient’s white blood disease has a low prevalence in the general popu- cell count. If an for a given disease. For example. there are few disease entities other selected subpopulation.e. eral population (1 person in 20 will have the dis. Storage. the PV considerably more common than are the other will be lower than the prevalence. When sible tumor. depends on the test’s sensitivity and specificity. The for- ease) but have a higher prevalence in a specially mula that describes the relationship precisely is: ( sensitivity )( prevalence ) PV + = ( sensitivity )( prevalence ) + (1 − specificity ) (1 − prevalence ) (021) 66485438 66485457 www. hypotheses and the assignment of a likelihood centage of a population of interest that has the (probability) to the presence of a specific disease disease at any given time. of all people in the United States. A particular disease or disease category.ketabpezeshki. and disease.

6. Because understanding issues of transcriptionists.O. Shortliffe and G. and many other We have described the process of hypothesis. and preva. The PV+ formula is one of One question is pertinent to all such applications: many forms of Bayes’ theorem. Barnett There is a similar formula for defining PV− of medical decision making. specificity. refine the physician’s hypotheses about what A variety of approaches have been used to abnormalities account for the patient’s illness. that PV+ be 100 %). Much of the rest of this ignored only when a test is pathognomonic (i. Physicians may at length (Elstein et al. which mandates the computer’s primary role is data management. Many clinical research systems (see Chap. well. retrieval. How do you get the data into the computer in the bining probabilistic data that is generally attrib. it is possible for data to to understanding medical data and their uses. the potential role of the com- Note also (by substitution into the formula) that puter in medical data storage. since the earliest days of the field. a rule for com. In Sect.ketabpezeshki. try to finesse this problem. Doctors. health care staff. 3.3 Methods for Selecting confusing movement through multiple display Questions and Comparing screens by the physician—have probably done Tests more to inhibit the clinical use of computers than have any other factor. Pople 1982) and be asked to fill out structured paper datasheets.H. lected? This question also has been analyzed 26) have taken this approach. 3. book deals with specific applications in which when its specificity is 100 %. disease (PV+) if the prevalence of that disease is low. we discuss the use of decision- lence. It is this relationship that tends to be Collection of Medical Data poorly understood by practitioners and that often is viewed as counterintuitive (which Although this chapter has not directly discussed shows that your intuition can misguide you!). 1978. tors who review patient charts.64 E.. computer systems. in terms of sensitivity. is pertinent for computer programs that gather or such sheets may be filled out by data abstrac- data efficiently to assist clinicians with diag. but the actual nosis or with therapeutic decision making (see entry of data into the database is done by paid Chap. be entered automatically into the computer by the we devote Chap. test selection and data interpretation is crucial In some applications. One is to design sys- The complementary question is: Given a set tems such that clerical staff can do essentially all of current hypotheses. 22). Note that positive treat a patient on the basis of available infor- tests with high sensitivity and specificity may mation or to perform additional diagnostic still lead to a low post-test probability of the tests.7 The Computer and is true. Bayes’ theorem is discussed in posed a problem for medical-computing systems greater detail in . sometimes simply refuse to use directed sequential data collection and have computers because of the awkward interfaces asked how an observation might evoke or that are imposed. You should substitute values in the PV + formula to convince yourself that this assertion 2. how does the physician the data entry and much of the data retrieval as decide what additional data should be col. Both formulae can be derived from analytic techniques in deciding whether to simple probability theory. 3 to these and related issues device that measures or collects them. for example. Awkward or nonintuitive interactions at computing devices— particularly ones requiring keyboard typing or 2. and inter- test sensitivity and disease prevalence can be pretation should be clear. For exam- (021) 66485438 66485457 www. first place? uted to the work of Reverend Thomas Bayes in The need for data entry by physicians has the 1700s.6.

for example. technologies that manage data and apply knowledge to generate advice. pen-based or finger-based mechanisms for data Patel. & Kaufman. & Pratt. as you learn about the application units. T. sensitive computer screens. E.. Klasnja. 4). monitors in intensive care or coronary care Chap. cally relevant predictions. J. approaches that compute on massive amounts of data iar with computers at home. Shah. 25. The authors to generate electronic or hard copy reports for discuss the transformation of data into informa- tion and knowledge. pulmonary function or ECG machines. Translational bioinformatics a patient’s care. in which genomic and proteomic data domi- nate. Papers deal with both translational bioinfor- the user does not need to learn a set of specialized matics.. such devices are allowing clini- This paper illustrates the role of theory-driven psycho- cians to maintain normal mobility (in and out logical research and cognitive evaluation as they relate of examining rooms or inpatient rooms) while to medical decision making and the interpretation of accessing and entering data that are pertinent to clinical data. P. R. delineating the ways in which physicians and also may be stored directly in a this focus lies at the heart of the field of biomedical computer database for subsequent use in other informatics. H. What is biomedical informatics? Journal of The patient’s responses to the questions are used Biomedical Informatics. big data. 184– use the machine themselves. Healthcare in the pocket: When physicians or other health personnel do mapping the space of mobile-phone health interven- tions. V. . (2012).. 104–110.. choice questions that follow a branching logic. When conventional computer workstations This editorial introduces a special online issue of the Journal of the American Medical Informatics are used. in which commands to enter or review data. laboratory can interface directly with a computer in which a database is stored. As more physicians are becoming famil. mouse-pointing Ohno-Machado. 45(1).. Psychology of Learning and Motivation. entry were introduced. Yearbook of Medical Informatics. W. Certain data can be entered directly by patients. Suggested Readings senting on a computer screen or tablet multiple- Bernstam. D. 31. because they can be crucial to and major developments in translational bioinformat- the successful implementation and use of a sys. Storage.2 Biomedical Data: Their Acquisition. Journal of the on a mobile tablet or smart phone (see Chap. N. and the cognitive issues is where new medical breakthroughs will occur. ics. and increasingly the clinician’s finger big role for biomedical informatics. touch. (2012). We encourage you to consider human– gies with large amounts of electronic health care data computer interaction. that take the patient’s history by pre. 130–134. 187–252. and a devices. specialized keypads can be helpful. and clinical research informatics. See also Chap. 19. R. It discusses data-centric design. and Use 65 ple. (1994).. they will find the (“Big Data”) to discover patterns and to make clini- use of computers in their practice less of a hin. Smith. This article reviews the latest trends tion areas. (021) 66485438 66485457 www. embraces big data. and. & Johnson. Journal of Biomedical Informatics. Arocha. 43(1). and areas and the specific systems described in later measurement equipment in the clinical chemistry chapters. 5). There were large clinical and public health datasets are clear improvements when handheld tablets using prominent. arguing that the field is ready to revolutionize human health and health care using large-scale tem. Most of these devices use a variant ered through the introduction of affordable mobile of the “point-and-select” approach—e. L. Association in which the rapidly evolving world of Designers frequently permit logical selection of biomedical and clinical “big data” challenges are the items from menus displayed on the screen so that focus. data services.ketabpezeshki. L. These issues arise in essentially all applica. e1. With ubiquitous wireless Diagnostic reasoning and medical expertise. (2010). J. (2012). settings. there are systems. arguing that research that bridges the latest multimodal measurement technolo- drance.g. V. W. 4. 7(1). This review article describes the multiple ways in often allow rapid and intuitive operator–machine which both patients and providers are being empow- interaction. American Medical Informatics Association. specialized devices 198. Big science. they warrant particular attention in system measurements on individuals. See that arise in dealing with computer systems (see also Chap.

Consider what you know about the to change as the number of bacteria typical daily schedule of a busy clini.” (Senior Medical Review 1987. from one to five? less devices. Although context in which data are collected influ. organisms per field a 95 % specificity mation. achieve a sensitivity of 95 %? (021) 66485438 66485457 www. obviously unrealistic for the physician cal terms. physicians’ offices. laboratories can provide such quantifi- ence the interpretation of those data? cation with reasonable accuracy.H. What are the advantages of wire.O. liliter of the patient’s urine. As a result. Given the imprecision of many medi.ketabpezeshki. physi. ml. level of at least 10. You check your pulse and discover that number of bacterial organisms in a mil- your heart rate is 100 beats per minute.000 bacteria per milliliter. Physicians Is this rate normal or abnormal? What generally assume that a patient has a uri- additional information would you use nary tract infection if there are at least in making this judgment? How does the 10. p.66 E. article offers the following guideline to dardization of terminology necessary physicians: “When interpreting … if computers rather than humans are to microscopy of … stained centrifuged manipulate patient data? urine. 4) pitals. and research (a) Describe an experiment that would institutions. a threshold of one organism per 3. per microscopic field to obtain a 6. have allowed the researchers to 4. To decide whether a patient has a sig. it is 2. why do you think that seri. per microscopic field increases cian. medical data? (b) How would you expect specificity 5. Shortliffe and G. How would medical practice change determine the sensitivity and speci- if nonphysicians were to collect all ficity of the microscopy. Based on the discussion of coding field yields a 95 % sensitivity and five schemes for representing clinical infor. (c) How would you expect sensitivity as tools for such clinicians? Can you to change as the number of bacteria think of disadvantages as well? Be sure per microscopic field increases to consider the safety and protection of from one to five? information as well as workflow and (d) Why does it take more organisms clinical needs. connected to the . Barnett Questions for Discussion cians commonly use a calculation of the 1.000 organisms per dardized terminology to be used in hos. discuss three challenges you for bacteriuria [bacteria in the urine] at a foresee in attempting to construct a stan. explicitly to count large numbers of bac- ous instances of miscommunication teria by examining a milliliter of urine among health care professionals are not under the microscope. one more common? Why is greater stan. specificity of 95 % than it does to nificant urinary tract infection.

and of information-retrieval West Lebanon. Stanford. and we D. CA. Geisel School of Medicine at flag abnormal test results. our approach is • What are utilities. we introduce probabilistic medical rea- results accurately? soning. Dartmouth College. 21).1007/ .H. we provide simple • What is a sensitivity analysis? How can we clinical examples that illustrate a broad range of use it to examine the robustness of a decision problems for which probabilistic medical reason- and to identify the important variables in a ing does provide valuable insight. of patient monitoring Dartmouth. health professionals discriminate between disease and health? often are faced with difficult choices.ketabpezeshki.1  he Nature of Clinical T answers to these questions: Decisions: Uncertainty • How is the concept of probability useful for and the Process of Diagnosis understanding test results and for making medical decisions that involve uncertainty? Because clinical data are imperfect and outcomes • How can we characterize the ability of a test to of treatment are uncertain. MD. you should know the 3. 31 Faraway Lane. decision? As discussed in Chap. J. medical practice is • What are influence diagrams? How do they medical decision making. are more Stanford University. An understanding of what E. we discuss ways that computers can help clini- cians with the decision-making process. Owens and Harold C. many medical decisions. USA The material in this chapter is presented in the Henry J. The concepts.C. USA systems (Chap. 67 DOI 10. NH 03784. Owens. Biomedical Decision Making: Probabilistic Clinical Reasoning 3 Douglas broadly applicable.). Throughout the chapter. an approach that can help health care pro- • What is expected-value decision making? viders to deal with the uncertainty inherent in How can this methodology help us to under. In this chapter. needs and system design and implementation. clinician. Kaiser Center for Primary Care context of the decisions made by an individual and Outcomes Research/Center for Health Policy. Together. Sox.K. Medical decisions are stand particular medical problems? made by a variety of methods. Biomedical Informatics. Sensitivity and specificity are important parameters of laboratory systems that H. Chaps. systems (Chap. MS (*) emphasize the relationship between information VA Palo Alto Health Care System. Shortliffe. © Springer-Verlag London 2014 (021) 66485438 66485457 www.J. and how can we use them to neither necessary nor appropriate for all deci- represent patients’ preferences? sions. however. In the remaining chapters. Cimino (eds. In this • What information do we need to interpret test chapter. USA e-mail: owens@stanford. Sox After reading this chapter. MACP Dartmouth Institute. we look differ from decision trees? at the process of medical decision making. CA. MD. 2 and 3 lay the groundwork for the rest of the book. 19). 2. Palo Alto.

Mr. Sox probability is and of how to adjust probabilities are we to understand this result? Before we try to after the acquisition of new information is a foun.K. James will die. The importance of probabil- ity in medical decision making was noted as long Example 2 ago as 1922: Mr. James is a 59-year-old man with coro- [G]ood medicine does not consist in the indiscrim- nary artery disease (narrowing or block- inate application of laboratory examinations to a age of the blood vessels that supply the patient. HIV? On an intuitive level. than that for a first operation. the test now has higher sensitivity that surgery will relieve the chest pain is lower and specificity. would be useful to identify peo- ple who have HIV. James undergo a third operation? and negative 99 % of the time antibody is The medications are not working. On the other hand. the donated blood specimen is priate for the blood bank. All choices in Example 2 entail considerable uncertainty. are grafted onto the old ones such that potential blood donors are tested to ensure blood is shunted past the blocked region. How quality and length of life. Owens and H. a clinician who uncriti- likely to contain the HIV.000 donors actually is inform many people that they had the AIDS infected. which may be fatal. social consequences. if only one in 1. Not only is the surgical Example 1 are consistent with the reported values of the sensitivity and specificity of the PCR test for mortality rate for a third operation higher than diagnosis of HIV early in its development (Owens that for a first or second one but also the chance et al. chest pain. let us consider a related example. Thus. the risks are grave. The test intuition is either misleading or inadequate. despite medication. but rather in having so clear a comprehen. that Mr. because blood cannot reach it. heart tissue). the test is more often mistaken than it is virus—a mistake with profound emotional and correct. these questions do not These examples illustrate situations in which seem particularly difficult to answer. the patient often experiences chest pain (angina). James has twice undergone coronary artery bypass graft (CABG) surgery. Should . In Example 2. the 1 The test sensitivity and specificity used in surgery is hazardous. which a section of the muscle dies. When the heart muscle does sion of the probabilities and possibilities of a case not receive enough oxygen (hypoxia) as to know what tests may be expected to give information of value (Peabody 1922). a procedure Example 1 in which new vessels. we are surprised cally reports these results would erroneously to find that. what is the likelihood that decision may substantially increase the chance a donor actually has HIV? If the test is negative. he has again begun to have immunodeficiency virus (HIV).C. The decision will be dif- how sure can you be that the person does not have ficult even for experienced clinicians. Similar situations are (021) 66485438 66485457 www. appears accurate. There would be making skill of the clinician will affect a patient’s ten wrong answers for each correct result. You ask whether use muscle is deprived of oxygen. find an answer. the caus. an incorrect If the test is positive. that they are not infected with the human Unfortunately. The PCR test is positive 98 % of the time when antibody is present. he runs a high risk of suffering a heart attack. the result can of the polymerase chain reaction (PCR). dation for our study of clinical decision-support systems (Chap. the decision-­ fewer than 10 would be infected. often taken from the You are the director of a blood bank. Furthermore. 1996b). of 100 donors with a positive test. and we would expect that. 22). if the Although the test results in Example 1 are appro- test is positive. which becomes progressively ative agent of acquired immunodeficiency more severe.1 gery. In fact. All leg.68 D.ketabpezeshki. If the heart syndrome (AIDS). nose HIV. without sur- absent. in gene-amplification technique that can diag. a be a heart attack (myocardial infarction).

but the radiologist cannot tell words such as “probable” and “highly likely” to whether there is a new blood clot. 3. therefore. there is ample oppor- experience of the patient. The treatment for a blood clot is to fect associations will be. Many deci. As Smith nicians do not want to treat the patient unless they (1985. uncertainty is not unique to medicine. The distinction may be blurred if the patient’s experience also can be observed by the The problem of how to express degrees of clinician. The main diagnostic concern is the recurrence of sion analysis can help to make clear the best a blood clot in his leg. The blood flow tainty in medical decision making.ketabpezeshki. much confidence should be required before start- Even the most astute physician will occasionally ing treatment? We will learn that it is possible to be wrong. ally have a blood clot. Furthermore. presents with the complaint 3. A symptom is a subjective common descriptive terms.1. But how probabilities are necessary but also perilous. p. This example illustrates an important con- 3. 3) said: “Medical decisions based on are confident that a thrombus is present. The degree of Uncertainty imperfection varies. The language that clinicians use to describe a bosis. a potentially activities of the healthcare professional.2 A test (ultrasonography) is per.2 °F. These words have strikingly different mean- 2 In medicine. On physical examination.” answer this question by calculating the benefits and harms of treatment. are made on the basis of causes of a swollen leg. A clot in the veins of the course of action. knowledge that has been gained through collec. however. equivocal. but all clinical data—includ- ing the results of diagnostic tests. a sign is an objective physical finding ings to different individuals. uncer. the physical findings leave con- empirical knowledge of associations between siderable uncertainty. describe their beliefs about the likelihood of dis- ease. 3. leg can dislodge.1. patient’s condition often is ambiguous—a factor formed. Thus. Therefore. administer anticoagulants (drugs that inhibit tain. which pose the risk of decisions made under uncertainty and present an excessive bleeding to the . a 33-year-old man with a history of a previous blood clot (thrombus) in a vein in his left leg. and cause a Decision making is one of the quintessential blockage in the vessels of the lungs. Kirk be treated for blood clots? ter is to show how the use of probability and deci. to some degree. in Example symptoms and disease to evaluate a problem.1. Kirk’s leg is evaluated. Should Mr. Given a swollen leg. In Sects. the leg is Uncertainty tender and swollen to midcalf—signs that suggest the possibility of deep vein throm. there are numerous other sions. about one-half actu- soning or of physiological principles. and the flow of blood in the veins of that further complicates the problem of uncer- Mr. a clinician cannot be sure that a clot is tive experience: the clinician often must rely on the cause.2 and 3. cli- overview of the diagnostic process.1.1 D  ecision Making Under cept: Clinical data are imperfect. Some fatal event called a pulmonary embolus. Because of the (something observed by the clinician) such as a widespread disagreement about the meaning of temperature of 101.3. Kirk. 3. How is it handled in other contexts? Horse racing has its (021) 66485438 66485457 www.1. patients with a swollen leg. we examine blood clot formation).2 P  robability: An Alternative of pain and swelling in that leg for the past Method of Expressing 5 days. the history given by the patient. Of decisions are made on the basis of deductive rea. flow with the blood.1. Clinicians use is abnormal. Mr.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 69 commonplace in medicine. and the findings on physical Example 3 ­examination—are uncertain. the results of the available diagnostic test are A decision that is based on these usually imper. feverish. such as feeling hot or tunity for miscommunication. Our goal in this chap.

nation. the second stage of the diagnos- may be based on previous experience or on tic process involves gathering more information. Smith exercises. explicit estimation of the probability of disease. If experienced gamblers are deciding whether to place bets. heart must pump more blood per stroke and The odds are simply an alternate way to express must beat faster (and thus requires more a probability. might assess the tic process into three stages. Sox share of uncertainty.2. data collection.5 The first stage involves making an initial judg.ketabpezeshki. physically stressed.K. expanding the con- chance that he has heart disease to warrant cepts of test sensitivity and specificity first intro- duced in Chap. (50 % chance or 1:1 odds. Smith. We ment about whether a patient is likely to have a explore methods used to estimate pretest proba- disease. they will find it ordering an exercise stress test. an electrocardiogram (ECG) is taken “high chance” of winning. 3. a 60-year-old man. The positive test result supports the diagnosis of This estimated probability. Owens and H. and severity of his or her prising successive iterations of hypothesis pain.  3. Because the know the odds. a clinician intuitively develops a belief After the pretest probability of disease has about the likelihood of disease. the quality.70 D. Here. many heart condi- expression of uncertainty avoids the ambiguities tions are evident only when the patient is inherent in common descriptive terms.3. The clinician in briefly in light of a specific example. 3. 3. In the stress unsatisfactory to be told that a given horse has a test. see Sect. decide what to do next based on his or her intu- We discussed how observations may evoke a ition about the etiology (cause) of the chest pain. Some tests Mr. For the pur. complains to reduce uncertainty more than do others (see his clinician that he has pressure-like chest Fig. from talking with the patient. pretest or prior probability of heart disease as 0.2). made before further heart disease. based on what he or she knows pose of our discussion. 2. This judgment been estimated. we review this process pretest probability of disease. Example 4 chose the exercise stress test. 3. Sect. duration. Mr. we described the hypothetico-­ The clinician would first talk to the patient about deductive approach.C. The more a test pain that occurs when he walks quickly.1. . hypothesis and how new information subse. this example. and this reduction in uncertainty is information is obtained.1a. The use of probability or odds as an oxygen) during exercise. the clinician would then generation. (021) 66485438 66485457 www. we separate the diagnos. knowledge of the medical literature. Our approach is to ask the clinician to make his quently may increase or decrease our belief in or her initial intuition explicit by estimating the that hypothesis. bility accurately in Sect. The clini- belief about the likelihood of disease usually is cian in Example 4 ordered a test to reduce the implicit. is the prior probability shown in Fig. 2. 3. and interpretation. In After taking his history and examining him. and the clinician would like to know which Example 4 test he or she should order next. They will demand to while Mr. Although the clinician in or pretest probability of disease. he or she can refine it by making an uncertainty about the diagnosis of heart disease. reduces uncertainty.3 O  verview of the Diagnostic Process How would the clinician evaluate this patient? In Chap.1b). the more useful it is. After an interview and physical exami. but may cost more. A clinician’s often by performing a diagnostic test. we explore ways to measure how well his clinician believes there is a high enough a test reduces uncertainty. a diagnostic strategy com. Smith’s results show abnormal changes in the ECG during exercise—a sign of heart disease. there are many tests available to diagnose heart dis- ease.

2 because the discussion was about the use of tests for screening populations of patients. of the test. there are other statistical interpretations of probability. In this framework.3 tain not to occur has a probability of 0. In bility of disease before they order tests. Methods to Assess Pretest mate. the 3.1a). 3 Note that pretest and post-test probabilities correspond to collectively exhaustive outcomes of a chance ­ the concepts of prevalence and predictive value. in a coin flip. Probability Sect.4. we a number between 0 and 1. Thus. The clinician in Example 4 must ask: Probability “What is the probability of disease given the abnormal stress test?” The clinician wants to In this section. we explore the methods that clini- know the posterior probability. the pretest probability of disease is simply 4 We assume a Bayesian interpretation of probability. is our preferred means of expressing uncertainty. (b) Test 2 reduces uncertainty about presence of disease (increases the probability of disease) more than test 1 does b Given new information provided by a test. As nician’s opinion about the likelihood of an event as we noted. terms were used in Chap. to calculate post-test probability. (021) 66485438 66485457 www.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 71 Fig. intro. or post-test cians can use to make judgments about the proba- probability. as well as the to occur has a probability of 1. we reexamine Bayes’ theorem. duced in Chap. (a) A positive test result increases the probability of disease. in a population. that disease’s prevalence in that population. an event that is cer- sensitivity and specificity. 3.ketabpezeshki. of disease (see Fig. 3.2 Probability Assessment: third step is to update the initial probability esti.  3. probability (p) expresses a cli- culating the post-test probability of disease.4 The probability of event A is written p[A]. An event that is certain must know the pretest . The latter event must be equal to 1. 2.1  The effect of test a results on the probability of disease. and we discuss its use for cal. The sum of the probabilities of all possible.

2. cal. we say the probability of mate probability is to ask themselves: What is a blood clot given a swollen leg is 0. One way that people esti- len leg have a blood clot.B] = p [ A ] ´ p [ B].0. findings (signs and symptoms) associated with a blood clot. thus giving an incorrect mental represen- The clinician may compare the current problem tation. To make these subjective assessments of prob- ring together is denoted by p[A&B] or by p[A. The (Tversky and Kahneman 1974). Owens and H.1 Subjective Probability Difficulties occur with the use of this heuristic Assessment when the disease is rare (very low prior prob- ability. we can think of heuristics as rules of p [ A. thumb.B]. what is the probability that p [ blood clot | swollen leg ] = 0. when the patient’s clinical profile is (021) 66485438 66485457 www. a cognitive heuristic is a men- vidual probabilities: tal process by which we learn. of our intuitive probability assessment.C.72 D. the pretest answer. If the patient has all the classic and limitations. B. Three heuristics have been identified as tional probability predicated on the test or ­finding. or similar to.” Thus a post-test probability is a condi. Representativeness. More specifically. if 30 % of patients who have a swol. this patient who has a swollen leg belongs to the class of patients who have blood clots? To Before the swollen leg is noted. 1. the clinician judges that the patient is highly likely to have a blood clot. we often rely on the representative- probability is simply the prevalence of blood ness heuristic in which probabilities are clots in the leg in the population from which the judged by the degree to which A is representa- patient was selected—a number likely to be much tive of. people rely on several discrete. recall.5 = 0. (Regardless and sophisticated decision makers (including clini- of the outcome of the first toss. denoted: the probability that object A belongs to class B? For instance.K. or p­ rocess information. often Events A and B are considered independent unconscious mental processes that have been if the occurrence of one does not influence the described and studied by cognitive psychologists probability of the occurrence of the other.25. These processes probability of two independent events A and B are termed cognitive heuristics. judge the probability of the development of a Now that we have decided to use probability blood clot (thrombosis) by the degree to which to express uncertainty. we may make mistakes in denoted by p[A|B] and read as “the probability of estimating probability in deceptive clinical situa- A given B. just as we may underes- that event B is known to occur is called the condi. Knowledge of heuristics is important because it helps us to understand the underpinnings Thus.ketabpezeshki. the probability of heads on two consec.3. when the clinician’s Most assessments that clinicians make about previous experience with the disease is atypi- probability are based on personal experience. and Kahneman 1974). or prevalence). each approach has advantages blood clot. So.3. 3.3.5 × 0. both occurring is given by the product of the indi. important in estimation of probability: For example. then ask: “What was the frequency of disease in similar patients whom I have seen?” The probability of event A and event B occur. how can we estimate the patient with a swollen leg resembles the probability? We can do so by either subjective or clinician’s mental image of patients with a objective methods. Sox to similar problems encountered previously and p [ heads ] + p [ tails ] = 1. The clinician will smaller than 0. timate distances on a particularly clear day (Tversky tional probability of event A given event B. ability.) fore make systematic—often serious—errors when The probability that event A will occur given estimating probability.5. Both naive utive coin tosses is 0. tions. the probability of cians and statisticians) misuse heuristics and there- heads on the second toss is .

8. A clinician who had cared for a useful starting point for estimating probability. atypical.2. the clini. Events more can use the prevalence of disease in the popula- easily remembered are judged more probable. patient who had a swollen leg and who then For example.6. A clinician or signs. can makes an initial probability estimate (the be used to place patients into a clinical subgroup anchor) and then adjusts the estimate based in which the probability of disease is known. it is a probability. places a patient in a subgroup with a known prob- lished research results to estimate probabilities. For on further information. disease.ketabpezeshki. The clinician can ical contribution to a total score. A trend has patient’s strong family history of heart dis. If 50 %. Another com- often are available in the medical literature. known as clinical ease increases the probability that he or she prediction rules. and he or she would from which you could increase or decrease the tend to overestimate the probability of a blood probability depending on your findings. such as difficulty with urination.2 Objective Probability findings depends on whether other findings Estimates are present. to help clinicians assign patients has heart disease.5. 2. such as a palpable prostate nodule. Instead of raising his or her esti. Our estimate of the probability of Published research results can serve as a guide an event is influenced by the ease with which for more objective estimates of probabilities. The total score avoid some of these difficulties by using pub. result is a list of symptoms and signs for an indi- An understanding of heuristics is thus important vidual patient. defined subgroup. The symptoms or signs that Heuristics often introduce error into our judg. This approach may be limited by difficulty he or she then learns that all the patient’s in placing a patient in the correct clinically brothers had died of heart disease. and it is prediction . Availability. take is to adjust the initial estimate (the Clinical prediction rules are developed from anchor) insufficiently in light of the new systematic study of patients who have a particu- information. (021) 66485438 66485457 www.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 73 atypical. the cians can use combinations of clinical findings to ­clinician might adjust it to only 0. Errors in our initial bility that a patient has a disease are identified estimates of probabilities will be reflected in the and assigned numerical weights based on statisti- posterior probabilities even if we use quantitative cal analysis of the finding’s contribution. mon heuristic used to judge probability is Symptoms. or clinical this rule is the availability heuristic. Anchoring and adjustment. 2. man. The usual mis. lar diagnostic problem. they define how clini- mate of prior probability to. ­probability of prostate cancer in a 50-year-old ber thrombosis as a cause of a swollen leg. estimate probability. to estimate the probability of often misleading. each with a corresponding numer- for medical decision making. been to develop guidelines. Estimates clot in patients with a swollen leg. 0. The methods to derive those posterior probabilities. ability of disease. patients referred to a urologist for evaluation of a nician in Example 4 makes an initial estimate prostate nodule. the prevalence is and therefore are likely to overestimate their the frequency of an event in a population. a fact the clinician could to well-defined subgroups in which the probabil- ascertain from the literature. if you wanted to estimate the died from a blood clot would vividly remem. and when the probability of certain 3. tion or in a subgroup of the population. ity of disease is known. of disease prevalence in a defined population 3. the prevalence of cancer is about of the probability of heart disease as 0. anchoring and adjustment. especially if the criteria for cian should raise the estimate because the classifying patients are ill-defined. the cli. make an independent contribution to the proba- ments about prior probability. For instance. or emotion-laden events more easily As we discussed in Chap. the prevalence of prostate cancer in men of The clinician would remember other causes of that age (5–14 %) would be a useful anchor point swollen legs less easily. We remember dramatic. say. We we remember similar events.

at considerable financial and per- a Fluid in the lungs due to reduced heart function sonal cost. a 65-year-old woman who had a Prevalence (%) of cardiac heart attack 4 months ago. pub. suggested that the Poor medical condition 5 probability of serious disease in asymptomatic Emergency surgery 10 patients with only microscopic evidence of blood Source: Modified from Palda et al. the prevalence of disease in the patient popula- the clinician can estimate that the patient has a tion in a specialist’s practice often is much higher 27 % chance of developing a severe cardiac than that in a primary care practice.74 D.1 Ms. ists seldom see patients with clinical findings ability of cardiac complications. and is about to undergo elec. The tests involve some risk.1  Diagnostic weights for assessing risk of car- diac complications from noncardiac surgery Clinical finding Diagnostic weight particular patient. Referral which the estimates are based. referred to urologists. as shown in that imply a low probability of disease.K. We add the diag. the special- places the patient in a group with a defined prob. Table  3. bias is common because many published studies lished prevalence data may not apply directly to a are performed on patients referred to specialists. 0–15 5 cal condition. This subject to error because of bias in the studies on example demonstrates referral bias. Because of this initial findings to obtain the total score.2  Clinical prediction rule for diagnostic weights Example 5 in Table 3. however. has abnormal Total score complicationsa heart rhythm (arrhythmia). many patients may have undergone unnec- tractions on preoperative electrocardiogram essary tests. studies per- complication. >30 60 Source: Modified from Palda et al. Nonetheless. A clinical illustration is that Age greater than 70 years 5 early studies indicated that a patient found to Recent documented heart attack have microscopic evidence of blood in the urine >6 months previously 5 (microhematuria) should undergo extensive tests <6 months previously 10 because a significant proportion of the patients Severe angina 20 would be found to have cancer or other serious Pulmonary edemaa diseases. heart attack. who are s­pecialists. PVCs premature ventricular con. (1997) a Cardiac complications defined as death. Sox Table 3. For instance. the Ever 5 approach of ordering tests for any patient with Arrhythmia on most recent ECG 5 >5 PVCs 5 microhematuria was widely practiced for some Critical aortic stenosis 20 years. Thus. (1997) was only about 2 % (Mohr et al. is in poor medi. Within 1 week 10 and expense to the patient. Ms. thus.ketabpezeshki. discomfort.1 lists clinical findings and their corre. In the ECG electrocardiogram. The primary care clinician refers patients whom sponding diagnostic weights. Troy will suf. A later study. Table 3. What explains the discrepancy in the estimates What is the probability that Ms. of disease prevalence? The initial studies that fer a cardiac complication? Clinical prediction showed a high prevalence of disease in patients rules have been developed to help clinicians to with microhematuria were performed on patients assess this risk (Palda and Detsky 1997). formed with the former patients therefore almost Objective estimates of pretest probability are always overestimate disease probabilities. Troy receives a score of 20. or congestive heart failure Table 3. (021) 66485438 66485457 www. past. he or she suspects have a disease in the special- nostic weights for each of the patient’s clinical ist’s sphere of expertise. Troy. 20–30 27 tive .C. The total score screening by primary care clinicians. 1986). Owens and H.2.

we present disease in a group that is heterogeneous with respect the issues that you need to consider when making to symptoms. 3. By Most clinical laboratories report an “upper limit combining subjective and objective methods for of normal.2). or bell-shaped) distribu- specific information about the patient. we can obtain a 3. objective data that are available. X-ray part of the chapter to illustrate how the clinician in studies. tion curve (Fig. more than two standard deviations from the mean The clinician should be careful. As an example. Smith. In this case. to avoid at each end of the distribution. The prevalence in a clinical subgroup. he or she should be aware of the ten. will predict the pretest probability determine criteria for deciding whether a result is more accurately than would the prevalence of heart normal or abnormal. result greater than two standard deviations above ability of heart disease. a pretest probability after assessing the specific clinical laboratory might choose as the upper (021) 66485438 66485457 www. usually in the form We now can use the techniques discussed in this of formal diagnostic tests (laboratory tests. if the mean choles- probability. such as men with symptoms typical of coronary The first challenge in assessing any test is to heart disease. dency to stay too close to the initial estimate How is a test result classified as abnormal? when adjusting for additional information. values for ill individuals may be normally distrib- tics to make subjective probability estimates. we summarized subjective test result below that cutoff is reported as normal and objective methods to determine the pretest (or negative). To help you to understand this step Example 4 might estimate the pretest probability of more clearly. We begin by using the use probability to interpret the results of the tests.1 C  lassification of Test Results lence is useful as an initial estimate that can be as Abnormal adjusted based on information specific to the patient.2). In this section. The distribution of the mistakes that can occur when one uses heuris. such as the population at large.). Although the prevalence of heart disease in Most biological measurements in a population men with typical symptoms is high. In uted as well. 10 % of patients of healthy people are continuous variables that with this history do not have heart disease. we discuss in the next two sections heart disease in his or her patient. to gather further information. About 2. however. and we learned how to adjust the terol concentration in the blood is 220 mg/dl. The next step in the diagnostic process is bility in other clinical settings. the clinician might adjust his or her initial tion will fall within two standard deviations of estimate of 0. 3. For exam. assume different values for different individuals. the clinician can dard deviations above the mean. Mr. etc. Thus.” which usually is defined as two stan- assessing pretest probability. however.95 or higher based on the mean.3  easurement of the M more refined estimate by placing the patient in a Operating Characteristics clinical subgroup in which the prevalence of disease of Diagnostic Tests is known. a In this section. one may need to adjust published estimates s­ ubpopulation of which the patient is representa- before one uses them to estimate pretest proba. the mean is reported as abnormal (or positive). The two distributions usually over- particular.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 75 Thus. lap (see Fig. who how to measure the accuracy of tests and how to has pressure-like chest pain. a test arrive at a reasonable estimate of the pretest prob. 3.5 % of the population will be information about family history of heart disease. We such a determination. 95 % of the popula- ple.9 upward to . assume that large studies have shown the prevalence of coronary heart disease in men with typical symp- toms of angina pectoris to be about 0. The clinician might use subjective methods to The distribution of values often is approximated adjust his or her estimate further based on other by the normal (gaussian.3.ketabpezeshki.9. Thus. tive. this preva. The prevalence of heart disease in 60-year-­old men could be our start- ing point.

Likewise. Note that a cut.ketabpezeshki. errors. FP. if the cutoff value were set TP true positive. A perfect test would have no FN or FP results. each group: TP.C. is • A false positive (FP) is a positive test result the total number of patients with a positive test obtained for a patient in whom the disease is result. If there is an overlap in the dis. The sum of the first row. Varying the cutoff between “normal” and “abnormal” across the continuous range of possible values changes the relative proportions of false positives (FPs) and false negatives (FNs) for the two populations limit of normal 280 mg/dl because it is two stan. The sum of the second col- absent (the test result correctly classifies the umn is the total number of nondiseased patients.2).76 D. If a Figure 3. . Disease Disease off that is based on an arbitrary statistical criterion Results of test present absent Total may not have biological significance. as shown in Table 3. TP + FN. present (the test result correctly classifies the The table summarizes the number of patients in patient as having the disease). and you obtained for a patient in whom the disease is can use a 2 × 2 contingency table to define the present (the test result incorrectly classifies measures of test performance that reflect these the patient as not having the disease). the number individuals. normal test (see Fig. patient as having the disease). we can con- with the terms used to denote these groups: veniently summarize test performance—the abil- • A true positive (TP) is a positive test result ity to discriminate disease from nondisease—in a obtained for a patient in whom the disease is 2 × 2 contingency table. Note that the sum • A true negative (TN) is a negative test result of the first column is the total number of diseased obtained for a patient in whom the disease is patients. Positive result TP FP TP + FP An ideal test would have no values at which Negative result FN TN FN + TN the distribution of diseased and nondiseased peo.2  Distribution of test results in healthy and diseased individuals. As the cutoff is moved further up tribution of test results in healthy and diseased from the mean of the normal values. the test would be normal in all false negative healthy individuals and abnormal in all individu- als with disease. FP + TN. some diseased patients will have a of FNs increases and the number of FPs decreases. Owens and H. Sox Fig.5 % of healthy individuals will have abnormal test will change the relative proportions an abnormal test.3. 3. of these groups. Table 3. and FN.2 shows that varying the cutoff point test result is defined as abnormal by the statistical (moving the vertical line in the figure) for an criterion.K. TP + FN FP + TN ple overlap. (021) 66485438 66485457 www. FP false positive. • A false negative (FN) is a negative test result Erroneous test results do occur. however. FN + TN gives the total number absent (the test result incorrectly classifies the of patients with a negative test result. FN appropriately.3  A 2 × 2 contingency table for test results dard deviations above the mean. That is. Few tests meet this standard. 3. You should be familiar Once we have chosen a cutoff point. TP + FP. TN true negative. TN. patient as not having the disease).

the sensitivity and speci- nondiseased patients with a negative test divided ficity of the HIV EIAs are greater than 99 %.2 Measures of Test Performance  Number of nondiseased patients   with negative test  Measures of test performance are of two types: TNR =  . So that the performance of the EIA can be The true-negative rate (TNR).3. presence or absence of antibody. test result. patient has a positive test result: The likelihood that a diseased patient has a posi-  Number of nondiseased patients  tive test is given by the ratio of diseased patients   with positive test with a positive test to all diseased patients: FPR =    Total number of nondiseased patients     number of diseased patients with positive test    TPR =  . and measures of disagree- ment or measures of discordance.  totalnumber of diseased patients  FP = . or . 2.  Total number of nondiseased patients  measures of agreement between tests or mea. FNR =   sensitivity is expressed as the probability of a  Total number of diiseased patients    positive test given that disease is present:   FN . positive rate (FPR) and the false-negative rate which we introduced in Chap. 3.3. TP + FN enzyme-linked immunoassay (EIA).3). In terms of conditional prob.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 77 3. the TNR is the number of calculations. The FNR is the measure in terms of the 2 × 2 table and in terms of likelihood that a diseased patient has a negative conditional probabilities. The true-positive rate (TPR).3). So.  Number of diseased patients  is the likelihood that a diseased patient has a pos.3: TPs and TNs. We discuss the idea of gold-standard tests in Sect. specificity is the probability of a negative test given that disease is absent: 5 This example assumes that we have a perfect method (different from EIA) for determining the p [ negative test | no disease ]. The total number of diseased patients is the Consider again the problem of screening sum of the first column. The number of diseased patients with a positive test is Example 6 TP. in the 2 × 2 table in Table 3. FP + TN We can determine these numbers for our example from the 2 × 2 table (see Table 3. The relative frequencies TNR = of these results form the basis of the measures of TN + FP concordance. = p [ positive test | disease ].     sures of concordance. We have chosen the numbers in the example to simplify the Viewed as a ratio.   with negative test itive test. As a ratio. Two types of From the 2 × 2 table (see Table 3. FN + TP The FPR is the likelihood that a nondiseased Another way to think of the TPR is as a ratio. In practice.4. the hypothetical results are shown negative test result. TP + FN. by the total number of nondiseased patients: (021) 66485438 66485457 www.4. or specificity.5 ability. We define each (FNR)—are defined similarly. These measures correspond to the The measures of discordance—the false-­ ideas of the sensitivity and specificity of a test. In conditional-probability notation. concordant test results occur in the 2 × 2 table in TN Table 3. One test used to TP screen blood donors for HIV antibody is an TPR = . measured. the test is performed on 400 is the likelihood that a nondiseased patient has a patients.ketabpezeshki. blood donors for HIV.

On the other Thus.C. we should set the cutoff value to the test were performed on 100 patients who truly minimize FP results.) characteristic (ROC) curve. The choice of cutoff antibody test. Any given point to Choose Among Tests along an ROC curve for a test corresponds to the test sensitivity and specificity for a given It may be clear to you already that the calculated threshold of “abnormality. as defined previously.78 D. had the antibody. Whenever a decision is made about EIA enzyme-linked immunoassay what cutoff to use in calling a test abnormal. If is dangerous. If the disease is serious and if TP 98 = = 0.4  A 2 × 2 contingency table for HIV anti. we ficity are characteristics not of a test per se but would expect two of the patients to receive incor. if the EIA test were performed on 100 specificity (i. In Fig. lifesaving therapy is available.ketabpezeshki.” Similar curves can values of sensitivity and specificity for a be drawn for any test used to associate observed continuous-­valued test depend on the particular clinical data with specific diseases or disease cutoff value chosen to distinguish normal and categories. is: depends on the disease in question and on the purpose of testing. or the specific values for any particular And the TNR is: patient). instead. the TPR against the FPR). Thus. 3. Similarly.3).99. known as a receiver-operating and FPR also must be 1: TNR + FPR = 1.99 is by the range of values of sensitivity and speci- TN + FP 297 + 3 ficity that it can take on over a range of possible The likelihood that a patient who has no HIV cutoffs. for an FNR of 2 %. EIA thereby increasing sensitivity while decreasing 100 300 specificity. an inherent philosophic decision is being made about whether it is better to tolerate FNs (missed To determine test performance.3. rather of the test and a criterion for when to call rect. result present absent Total Thus. it trades off specificity for sen- sitivity.98. we should try to TP + FN 98 + 2 minimize the number of FN . the cutoff level (moving it to the right) would body EIA decrease significantly the number of FP tests but EIA test Antibody Antibody also would increase the number of FN tests. 3. the likelihood that a patient with the HIV hand. (You that test abnormal.e. if the disease in not serious and the therapy antibody will have a positive EIA test is 0. the best way to characterize a test TN 297 = = 0. The typical way to show this relationship antibody will have a negative test is 0. Conversely. (Convince yourself that the sum of TNR resulting curve. (021) 66485438 66485457 www. note that increasing Table 3. as the individuals who had not been infected with HIV. 3.K.3 I mplications of Sensitivity applied to the field of psychology (Peterson and and Specificity: How Birdsall 1953. Owens and H. was originally described by researchers investigating methods of electromagnetic-signal detection and was later 3.) formed. cutoff is varied and the two test characteristics it would be negative in 99 and incorrectly posi. Swets 1973). The TPR. the test would have become more specific Positive 98 3 101 but less sensitive.. a lower cutoff value EIA Negative 2 297 299 would increase the FPs and decrease the FNs.2 should convince yourself that the sum of TPR and has no effect on the test itself (the way it is per- FNR by definition must be 1: TPR + FNR = 1. we calculate the cases) or FPs (nondiseased people inappropri- TPR (sensitivity) and TNR (specificity) of the EIA ately classified as diseased). is to plot the test’s sensitivity against 1 minus Therefore.2. we would expect the test to be We stress the point that sensitivity and speci- positive in 98 of the patients.98. Sox abnormal results. negative results. The tive in 1. are traded off against each other (Fig. Varying the cutoff in Fig.

When we classify a test result as TP.3. As shown in Fig.3  Receiver operating characteristic (ROC) curves the patient’s true disease state. The test may could be drawn. risk. for any level of specificity).population because diseased patients are more est sensitivity and specificity. provided that other likely to be included in studies than are nondis- factors. We are interested.3.ketabpezeshki.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 79 equal.g.3. study. Test B is more discriminative ­ diseased tissue or a surgical operation.perform than is the index test (otherwise. such as a biopsy of for two hypothetical tests. or more difficult to Suppose a new test were introduced that com. the specificity for any level of sensitivity (and when clinically relevant population comprises those its sensitivity is greater than test A’s sensitivity patients in whom a test typically is used. lation comprises those patients (usually a subset In other words. 3.4 D  esign of Studies of Test Performance In Sect. FP. and an ROC curve which it will be used in practice. Matters of cost. A at any value of true-positive rate (TPR)). are eased patients. Thus. published values of (021) 66485438 66485457 www. The more dis- criminative test may not always be preferred in clinical The test whose discrimination is being measured is practice. the more the results of that test will reduce uncertainty about probability of disease. discomfort. however. A gold- than test A because its curve is higher (e. the false-­ standard test is a procedure that is used to define positive rate (FPR) for test B is lower than the FPR for test unequivocally the presence or absence of disease. you should select the test that has the high. For example. We mentioned earlier the problem of referral and delay also are important in the choice about bias. Clinicians should not necessarily. a small. so we better discriminating power than a competing test make the following distinction: the study popu- if its ROC curve lies above that of the other test. the less peted with the current way of screening for the precise test would not be used at all). (derived from a study population) may differ When you must choose among several available from the prevalence in the clinically relevant tests. Similarly. in how the test This new test could be assessed for trade-offs in performs in the broader group of patients in sensitivity and specificity.5 Bias in the Measurement tion. we assume that we know with certainty whether a patient is diseased or healthy. test B is more discriminating than of the clinically relevant population) in whom test A when its specificity is greater than test A’s test discrimination is measured and reported. select group of patients enrolled in a ence or absence of pneumonia became available. a test has perform differently in the two groups. suppose a The performance of the index test is measured new radiologic procedure for assessing the pres. TN. of Test Characteristics always choose the test with the most discriminat- ing ROC curve. the validity of any test’s results must be ­measured against a gold standard: a test that reveals Fig. or FN. however. 3. riskier. Published estimates of disease prevalence what data to collect and what tests to perform. 3. presence of a disease.2. The gold-­standard test usu- ally is more ..3. The higher the sensitivity and specificity of a test. we discussed measures of test per- formance: a test’s ability to discriminate disease from no disease. however (see text) called the index test. such as cost and risk to the patient. Understanding ROC curves is important in understanding test selection and data interpreta.3.

proved the standard test than are patients with positive tests. (Remember. 1996). Early measures of test dis. the clinically relevant popula. Screening individuals with positive index–test results. so the reported val. patients with nega- the test had high sensitivity and specific. In clinical practice. per. healthy volunteers to have FP elevations of prostate-­ specific antigen (Meigs et al. the person notably spectrum bias. Subsequent work. the study population will have an relevant populations in terms of average level of artifactually low FNR. may affect test performance. Spectrum bias occurs when the study popula- tion includes only individuals who have advanced disease (“sickest of the sick”) and healthy volun. These differences factually high TPR (TPR = 1−FNR). (021) 66485438 66485457 www. the study population will have an artificially low FPR. artificial concordance between the tests (the crimination are overly optimistic. test referral bias.Measurement of prostate-specific antigen is often used to tion will contain more cases of early disease that detect prostate cancer. do not apply to the clinically relevant tion. CEA to be completely valueless as a screen. Owens and H. the results of the gold standard test. These problems usually are the result and TNs are the basis for measures of concor- of bias in the design of the initial studies—­ dance).ketabpezeshki.6Volunteers are often . a small portion of the colon. Sox sensitivity and specificity are derived from study are more likely to be missed by the index test populations that may differ from the clinically (FNs). however. This bias causes with numerous tests. patients with benign (rather than malignant) enlarge- easier to detect when it has spread throughout the ment of their prostate glands are more likely than are body (metastasized) than when it is localized to. Problems arise ously increases measures of concordance—the when the TPR and TNR. the relative frequencies of TPs population. In contrast to the blood that is elevated in men who have prostate cancer. higher percentage of patients with disease than and the differences between the study and does the clinically relevant population. the study population. index test is a criterion for ordering the gold stan- formed in selected patients. as measured in the study sensitivity and specificity—in the study popula- population. Reports of early investigations. Test-interpretation bias develops when the interpretation of the index test affects that of the The experience with CEA has been repeated gold standard test or vice versa. diseases that may cause FP results6. The result is overestimation of the CEA’s TPR and TNR (Ransohoff of the TPR and underestimation of the TNR in and Feinstein 1978). a blood test called the of the TPR and TNR of the CEA were partly due carcinoembryonic antigen (CEA) was to spectrum bias. as is often the case when a test is first being ­clinically relevant population often have several diseases in addition to the disease for which a test is designed. healthy volunteers are less likely than are patients ues may not apply to many patients in whom a in the clinically relevant population to have other test is used in clinical practice. For example. or test interpreting the index test should be unaware of interpretation bias. Advanced disease may be easier to These other diseases may cause FP test results. Test-referral bias occurs when a positive cer. cancer is ple. tive index tests are less likely to undergo the gold ity. For exam- detect than early disease. touted as a screening test for colon can. Therefore. study population. comprising ing blood test for colon cancer. a substance in the say. clinically relevant populations were partly both TN and FN tests will be underrepresented in responsible for the original miscalculations the study population. In other words.80 D. which produces an arti- health and disease prevalence. whereas patients in the teers. and subsequent results are more likely to be the same) and spuri- test performance is disappointing. Inaccuracies in early estimates In the early 1970s. and therefore the specificity will be overestimated Example 7 (TNR = 1−FPR). To avoid these problems.C. In addition. indicated that dard test. the study population. Thus. has a tests are used in unselected populations.K.

ketabpezeshki. 3. there are many studies that evaluate the sensitivity and specificity of the same diagnostic The third stage of the diagnostic process (see test.4. 1996a. result given the true state of the patient.1 Bayes’ Theorem b). Another lence in the general population) becomes the pre- approach is to perform a meta-analysis: a study test probability for the test that they perform. 1999.1a) is to adjust our probability estimate to about the sensitivity and specificity of the test.3. you should adjust the TPR (sensitiv. however. Thus. if these test with sensitivity (TPR) and specificity (TNR). For example. by 1995. Owens et al. As we noted earlier in this chapter. adjust the TPR and TNR when they are applied to Investigators develop a summary ROC curve by a new population. Summary ROC curves provide the best ity) downward when you apply it to a new available approach to synthesizing data from population. referral bias. 3. over 100 studies had assessed the sensitivity and specificity of the PCR for diagnosis of HIV (Owens et al. All the biases result in a TPR using estimates from many studies.4  ost-test Probability: Bayes’ P of Diagnostic Tests Theorem and Predictive Value Often. you should adjust the specific. biases are present. Test-referral bias. study. in contrast to that is higher in the study population than it is in the type of ROC curve discussed in Sect. 3. even the revised estimate (rather than the disease preva- high-quality studies did not agree. tion posed in the opening example: Given a posi- ulation that is lower than it will be in the clini. however. the clinically relevant population. however. take into account the new information gained you can have increased confidence in the results from diagnostic tests by calculating the post-test of the studies. But what if the studies disagree? probability. a clinician can and they assessed the specificity of PCR to be use the disease prevalence in the patient popula- between 40 and 100 %. If you suspect test patient has the disease? To answer this question. you should adjust the specificity we must learn methods to calculate the post-test upward when you apply it to a new population. (021) 66485438 66485457 www. what is the probability that this cally relevant population. not.6 Meta-Analysis 3. Adjustment of the TNR (specificity) depends Section 3. Leeflang et al. They do tion. The For evaluation of PCR. these studies estimated the sensitivity of PCR to be as low as 10 % and to be as high as 100 %.3. tive test result. that combines quantitatively the estimates from After they have gathered more information with individual studies to develop a summary ROC a diagnostic test. 2008). answer the clinically relevant ques- produces a measured specificity in the study pop.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 81 To counter these three biases. Once clinicians begin to accumulate assess the quality of the studies and to use the information about a patient. probability of disease with Bayes’ theorem.3. Thus. b. 1996a. on the other hand. their estimate of the probability of disease. Spectrum bias the diagnostic process: acquisition of further and test interpretation bias result in a TNR that is information with diagnostic tests. If the studies come to similar conclusions Fig. if you which is developed from the data in a single suspect bias. they revise estimates from the highest-quality studies. you may need to Hellmich et al. 1993. many studies. Which results should you tion as an initial estimate of the pretest risk of believe? One approach that you can use is to disease. probability of disease. We have higher in the study population than it will be in learned how to characterize the performance of a the clinically relevant . These measures reveal the probability of a test ity downward when you apply it to a new popula. 3.3 has dealt with the second step in on which type of bias is present. they can calculate the post-test curve (Moses et al.

the odds are 3:1.05 instead of 0.65 and 0.98 0. From Sect.95 changes to 0. (+). Substituting in Recall that a conditional probability is the Bayes’ formula for a positive test.75. by substituting FPR of the test to be 0. The change in probability is modest because p [D | R ] = p [ D ] × p [ R | D ] + p [ −D ] × p [ R | −D ] the pretest probability was high (0. Thus. 0. The probability of disease. on similar days.4. Assume that the erties of probability (see the Appendix to this TPR and FPR of the exercise stress test are chapter for the derivation).91. (021) 66485438 66485457 www.65 to be positive (event B).996. recall that p[+|D]  = TPR and p[+| − D] = FPR. At the end of Thus.2).2. important probability in Example 4: the 1 + odds post-test probability of heart disease after a positive exercise test. Example 8 1− p We are now able to calculate the clinically odds p= . 3. it is awkward for mental calcula- We can use a similar derivation to develop tions. given a test result.C. If we repeat the calculation with a pretest probability of 0.2 The Odds-Ratio Form of theorem for a positive test: Bayes’ Theorem and Likelihood Ratios p [ D ] × TPR p [D | +] = p [ D ] × TPR + (1 − p [ D ]) × FPR Although the formula for Bayes’ theorem is straightforward.20. a pre- p[D|+] for p[D|R].3.95.65 + 0. p[+|D] for p[R|D].ketabpezeshki.75. a positive test result: eral. We can develop a more convenient form of Bayes’ theorem for a negative test: Bayes’ theorem by expressing probability as odds and by using a different measure of test dis- p [ D ] × FNR p [ D | −] = crimination. we probability that event A will occur given that obtain the probability of heart disease given event B is known to occur (see Sect.20 of disease as D. We can reformulate this general equation in the post-test probability is 0. we estimated the pretest prob.05 ´ 0.98 from the pretest probability of p [ D] × p [ R | D] 0.2. Bayes’ theorem is: Thus. Owens and H. history of heart disease. If we assume the terms of a positive test.K. In gen. 3. Sox Bayes’ theorem is a quantitative method for calculating post-test probability using the pretest the prevalence of heart disease in men who probability and the sensitivity and specificity of have typical symptoms of heart disease and the test. if the probability of rain today is 0. The theorem is derived from the defini. respectively. Substitution provides Bayes’ 3. and the pretest probability of disease as p[D].95 ´ 0.95 ´ 0. based on expect rain to occur three times for each time it does not occur.20.82 D. p[+| − D] test probability of 0. given that the test is known 0. a test result as R. Sect. Probability and odds are related as p [ D ] × FNR + (1 − p [ D ]) × TNR follows: p odds = . is written p[D|R]. we should ability of heart disease as 0. and 1 − p[D] for p[− D]. the positive test raised the post-test probability to 0. we want to know the probability that disease is present (event A). on the prevalence in people with a family tion of conditional probability and from the prop. We denote the presence p [D | +] = = 0. for p[R| − D].20).95.95) and because the FPR also is high (0.  3. its absence as − .

respectively.75 p= = = 0.25 = 9. therefore.75 odds = = = = 3.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 83 A simple relationship exists between pretest In a test that discriminates well between dis- odds and post-test odds: ease and nondisease. odds 9. We describe the per. the test has no value. positive or negative) by two LRs: one corresponding to a positive test result and the other We can then convert the odds to a corresponding to a negative test. form of Bayes’ theorem: formance of a test that has only two possible out- post-test odds = 3 × 3. These ratios are probability: abbreviated LR + and LR−. An LR of 1 means that the proba- or bility of a test result is the same in diseased and nondiseased .ketabpezeshki.91 1 + odds 1 + 9. We can use We can calculate the post-test odds of a the LR to characterize clinical findings (such as a positive test result using the odds-ratio swollen leg) or a test result.. (021) 66485438 66485457 www.20 disease change based on the test result. your patient’s bedside. the FPR will be low. the LR − will be much less likelihood ratio (LR) for the test in question.  probability that test  TNR  is negative in nondiseased people  forward fashion from the definitions of Bayes’   theorem and of conditional probability that we provided earlier.75 0. = × . for example. to obtain the post-test A desirable test will have a low FNR and a odds. and thus LR + will be much post-test odds = pretest odds × likelihood ratio greater than 1. we simply multiply the pretest odds by the high TNR. ability at.g. The LR of a test combines the measures of test discrimination discussed earlier to give one num- ber that characterizes the discriminatory power of Example 9 a test.25 probability of result in diseased people LR = The LR for the stress test is: probability of result in nondiseased people TPR 0. Thus. p [D | R ] p [D] p [R | D] Similarly. the TPR will be high. or 3 : 1 1 − p 1 − 0. The odds-ratio form of Bayes’ theorem allows 7 Some authors refer to this expression as the odds-­ rapid calculation.75 0.25 The LR indicates the amount that the odds of FPR 0. p [ −D | R ] p [ −D ] p [ R | −D]  probability that test   is negative in diseased people  FNR This equation is the odds-ratio form of LR − =  = Bayes’ theorem. than 1.75 : 1 comes (e. this result agrees with our earlier answer (see the discussion of Example 8). The pretest odds are: or p 0. defined as: We can calculate the post-test probability for a positive exercise stress test in a 60 p [ R |D ] LR = year-old man whose pretest probability is p [ R | −D] 0.7 It can be derived in a straight.75  probability that test   is positive in diseased people  TPR LR + =  =  probability that test  FPR  is positive in nondiseased people    As expected. so you can determine the prob- likelihood form of Bayes’ theorem.65 LR + = = = 3.

The positive predictive value (PV+) of a test is the likelihood that a patient who It is worth reemphasizing the difference has a positive test result also has disease.97 table from which the PV + was calculated. The PV cannot be generalized to a new population because the prevalence of dis- TN ease may differ between the two populations.99 297 + 2 test odds in one step. you can calculate the post. Bayes’ 98 + 3 theorem provides a method for calculation of the post-test probability of disease for any prior (021) 66485438 66485457 www. in Example 1. The PV gives the probability of true dis- totalnumber of patients with a positive test ease state once the patient’s test result is known. it represents the post-­ Example 6 (see Table 3.84 D.001 in Example Example 10 1 and as 0. The prev- alence of antibody was given as 0.ketabpezeshki.K. The dis- crepancy is due to an extremely important and often overlooked characteristic of PV: the PV of a number of nondiseased patients with negative test PV − = test depends on the prevalence of disease in the Total number of patients with a negative test study population (the prevalence can be calculated as TP + FN divided by the total number of patients From the 2 × 2 contingency table in Table 3.3 Predictive Value of a Test patients with a positive test will have anti- body. therefore. The LR demonstrates that a useful test is one that changes the odds of disease.99.25 in Example 6. Rather. tive or negative test is to calculate the predictive value of the test. The probability that antibody is present in a patient who has a positive index test (EIA) in this study is 0. so we expect 97 of 100 patients with a positive index TP test actually to have antibody.3. Owens and H. that both are calculated from the 2 × 2 table and tingency table: they often are confused. between PV and sensitivity and specificity.4. The likelihood that a patient with a An alternative approach for estimation of the negative index test does not have antibody probability of disease in a person who has a . PV− = = 0. The negative predictive value (PV−) is the What explains the discrepancy in these examples? likelihood that a patient with a negative test does The sensitivity and specificity (and.4 is 0. about 97 of 100 3.97. the not have disease: LRs) in the two examples are identical.97. The PV + calculated from Table 3. Sox The LR is a powerful method for characteriz- ing the operating characteristics of a test: if you 297 know the pretest odds. PV + = TP + FP we found that fewer than one of ten patients with a positive test were expected to have antibody. PV − = TN + FN The difference in PV of the EIA in Example 1 and in Example 6 is due to a difference in the prevalence of disease in the examples. These examples We can calculate the PV of the EIA test should remind us that the PV + is not an intrinsic from the 2 × 2 table that we constructed in property of a test. From the 2 × 2 contingency table in Table 3. given PV + can be calculated directly from a 2 × 2 con.3.4) as follows: test probability of disease only when the preva- lence is identical to that in the 2 × 2 contingency 98 PV+ = = 0. is about 0.C. Yet. The sensitivity and specificity give the probability of a particular test number of diseased patients with positive test result in a patient who has a particular disease PV + = state. Thus. in the 2 × 2 table).

a positive test result can raise ability (see Example 8). calculated with Bayes’ theorem for all values of pretest This discussion emphasizes a key idea of this probability (Source: Adapted from Sox. (1987). curve) was calculated with Bayes’ theorem for all values of ative test. we explore the implications of Bayes’ theorem for test interpretation.4a that. 3. 3. 3. H. spond to a test that has no effect on the probability of dis- ease.4b shows the relationship between post-test probability of disease. 3.C.4a relates pretest and post-test probabilities in a test with a sensitivity and speci- ficity of 0. In: Sox H. (Ed. 3. We produced Fig. Fig.9. and a negative test result has intermediate or if the result contradicts a strongly a large effect. bility is very low. we prefer the use of a Bayes’ theorem to calculate the post-test proba- bility of disease. For that reason. the be 0. A neg.). however. If the pretest probability is high.4a by cal- culating the post-test probability after a positive test result for all possible pretest probabilities of disease. a positive test result has a large 1–17). a positive test result has a small effect.4.4 I mplications of Bayes’ Theorem In this section. yet they often are misunderstood.4  Relationship between pretest probability and Figure  3. (b) The post-test probability of disease corresponding to a negative test result (solid curve) was test probability if the pretest probability is low. indicating a test that is useless. Note from Fig. The 45-degree line in each figure denotes a test in which the pretest and post-test probability are equal (LR = 1).com . and a negative test result has a small effect. If the pretest ­probability is the post-test probability into only the ­intermediate (021) 66485438 66485457 www. We similarly derived Fig. a confir.C. has little effect on the post-­ pretest probability. when the clinician held clinical impression. At high pretest probabilities.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 85 probability. Note that. At high pretest probabilities. the post-test probability after a positive test result is much higher than is the pretest probability. chapter: the interpretation of a test result depends Probability theory in the use of diagnostic tests: Application on the pretest probability of disease. Common diagnostic tests: Use and interpretation (pp. These ideas are extremely important. 3. (pretest probability nearly 0 or nearly 1).4b for a neg- ative test result. Sensitivity and specificity of the test were assumed to ative test result.ketabpezeshki. The dashed lines corre- the pretest and post-test probabilities after a neg. The curve in Fig.4 illustrates one of the most essential concepts in this chapter: The post-test probability of disease increases as the pretest probability of b disease increases. if the pretest proba- matory test has little effect on the posterior prob. with permission) effect. the post-test probability is only slightly higher than the pretest probability. In other words. If the pretest to critical study of the literature. the test result will have a is almost certain of the diagnosis before testing large effect on the post-test probability. American College of Physicians. at low pretest probabilities. probability is low.90 for the two examples. (a) The post-test probability post-test probability after a negative test result is of disease corresponding to a positive test result (solid much lower than is the pretest probability. Figure 3.

In: Sox (Ed.5. This terminology oversim- plifies a diagnostic process that is probabilistic. Diagnostic specificity or a high .4 except that the calculations have been repeated for several values of the sensitivity (TPR true-positive ficity of a test markedly changes the post-test rate) and specificity (TNR true-negative rate) of the test.99. Figure 3. Sox range. Thus. That is. and the calculations were repeated for several diagnosis. if you are trying to rule in a specificity. In both panels. The curves are similar to those shown in tive test results). if the pre- test probability is low. If the clinician believes the pretest probability of coronary artery disease is 0. 3. Owens and H. the post-test probability is in an intermediate range. Probability patient does have the disease. to rule in a disease is to confirm that the results (Source: Adapted from Sox (1987).90. Figure 3. the top family of curves corresponds to positive test results and the bottom family to negative test results. College of Physicians. sitivities. Figure  3.8 you should choose a test with high values of the sensitivity of the test. Common diag- who strongly suspects that his or her patient has a bacterial nostic tests: Use and interpretation (pp. 3. if the pretest probability is very high.4a represents the rela. with permission) Another doctor is almost certain that his or her patient has a simple sore throat but orders a culture to rule out strepto- coccal infection (strep throat).5a shows the post-test probabilities for tests with varying specificities (TNR).g. American infection orders a culture to rule in his or her diagnosis. probability if the test is positive but has relatively (a) The sensitivity of the test was assumed to be 0. Assume that Fig.C.1. it is unlikely that a positive test result will raise the probability of disease suf- ficiently for the clinician to make that diagnosis with confidence. and therefore a posi.5. rule in or rule out a disease. test sensitivity affects primarily the interpretation of a negative test. HIV antibody tests have a specificity greater than 0.5 illustrates another important con- cept: test specificity affects primarily the inter- pretation of a positive test. An exception to this statement occurs when a test has a very high specificity (or a large LR+). a tionship between the pretest and post-test proba- bilities for the exercise stress test. rather. the top family of curves corresponds to positive test results. A doctor cal study of the literature.ketabpezeshki.. and the bottom family of curves corresponds to negative test 8 In medicine. and little effect on the post-test probability if the test the calculations were repeated for several values of test is negative. Similarly.  3. Note that changes in sensitivity produce (021) 66485438 66485457 www. test probability.86 D. if ever.). to rule out a disease is to theory in the use of diagnostic tests: Application to criti- confirm that the patient does not have the disease. Thus. 3. it is unlikely that a nega- tive test result will lower the post-test probability sufficiently to exclude a diagnosis. 1–17). Note that changes in the specificity produce large changes in the top family of curves (positive test results) but have Fig. e. the post-test probabilities for tests with varying sen- tests raise or lower the probability of disease. the post-test probability will be about 0.5  Effects of test sensitivity and specificity on post-­ little effect on the lower family of curves (nega.5b shows the tests rarely. which leaves considerable uncertainty about the diagnosis. Fig. b tive test is convincing.90. In both parts (a) and (b) of Fig. Although there has been a large change in the probability. (b) The specificity of the test was assumed to be 0. an increase in the speci.

if you are trying to probability of a particular result on the second exclude a disease. In a mutually exclusive. Tests for the same (negative test results) but have little effect on the disease are conditionally independent when the top family of curves. you developed in the preceding sections to solve such can use the post-test probability after the first test difficult decision problems. you senting and comparing the expected outcomes use Bayes’ theorem a second time to calculate the of each decision alternative. Common problems are inaccurate estimation of pretest probability. by the assume that all test abnormalities result from one accuracy of the estimated pretest probability. The accuracy of the calculated The fourth common problem arises when you post-test probability is limited. we turn to the problem Bayes’ theorem will be of little value. of treatments) are uncertain. as we have described it. a method for deciding whether new approach is valid. If of a surgical operation. probability often is sufficient. This probability. as are the results Bayes’ theorem to interpret a sequence of tests. as the methods: the decision tree. sequentially in situations in which conditional Bayes’ theorem provides a means to adjust an independence is violated.3. 3. Then. In Sect. however. without paying attention to the possible 3.g. Making formance was measured (see Sect.5 C  autions in the Application of Bayes’ Theorem second test positive | first test positive  p  and disease present Bayes’ theorem provides a powerful method for second test positive | first test negative  = p  calculating post-test probability. calculated with Bayes’ theorem. new information. choose a test with a high sensi. With certain tests. however. given (conditioned on) the disease state. clinicians need a method for disease in the population.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 87 large changes in the bottom family of curves are conditionally independent. as we shall see. tion is satisfied. choosing among treatments when the outcome A third potential problem arises when you use of the treatments is uncertain.5).5. you will obtain inac- estimate of pretest probability to take into account curate post-test probabilities (Gould 2003). probability. For example. 3. when using Bayes’ theorem is to apply published values for the test sensitivity and specificity. (and only one) disease process. Bayesian decision analysis. and clinical prediction rules. If you apply Bayes’ theorem independence and of mutual exclusivity. Thus. the LRs may differ depending on the Medical decision-making problems often cannot pretest odds in part because differences in pretest be solved by reasoning based on pathophysiol- odds may reflect differences in the spectrum of ogy.5 Expected-Value Decision effects of bias in the studies in which the test per. of the possible errors you can = p [second test positive | disease present ].. You should be  and disease present aware. however. make when you use it. and the threshold post-test probability after the second test. only if the two tests information can change a management decision. a method for repre- pretest probability for the second test. Expressed in conditional probability notation for the case in which the disease is present.ketabpezeshki. of decision making when the outcomes of a clini- A second potential mistake that you can make cian’s actions (e. You can use the ideas a patient undergoes two tests in sequence. the post-test odds  = pretest and violation of the assumptions of conditional odds × LR1 × LR2. tivity or a high LR−. If they are not. Here we discuss two result. a range of prior updating must be applied with great care. if the We have shown how to calculate post-test pretest probability assessment is . generally pre- increased by proper use of published prevalence sumes that the diseases under consideration are rates. or LRs. test does not depend on the result of the first test. The Bayesian Accuracy of estimated prior probability is approach. 3. If the conditional independence assump- faulty application of test-performance measures. Nonetheless. (021) 66485438 66485457 www.4. heuristics.

15 3 0.6 Years after and summarized in Table 3. as illustrated of death by the frequency distribution shown in Fig. the results of either are unpredictable (021) 66485438 66485457 www. hunches or on a sixth . Fig.K.40 0.7 illness. Which of the two therapies is preferable? Example 11 demonstrates a significant fact: a choice among therapies is a choice among gam- 3.5. Sox These techniques help you to clarify the decision How can a clinician determine which course of problem and thus to choose the alternative that is action has the greatest chance of success? most likely to help the patient. 3.20 0. we rely on ­ Like those of most biological events.05 which therapy a patient receives.88 D. the out. situations in which chance determines Prospects the outcomes). The length of a patient’s life after Probability either therapy is unpredictable. Two therapies are available.5  Distribution of probabilities for the two There are two available therapies for a fatal therapies in Fig.30 0. choose among gambles? We propose a method Example 11 Table 3. How should we come of an individual’s illness is unpredictable. Figure 3.10 0.45 by the end of the fourth year.. but there is no 4 0.6  Survival after therapy for a fatal disease. but t­herapy B or might survive to the fourth year the patient might die in the first year with with therapy A. How do we usually choose among gambles? More often than not. Each therapy is therapy Therapy A Therapy B associated with uncertainty: regardless of 1 0.ketabpezeshki. he will die 2 0.1 C  omparison of Uncertain bles (i. Owens and H.5.e.6 shows that survival until the fourth year is more likely with therapy B.C. 3.35 way to know which year will be the patient’s last. 3.

Finally. Then. death during an interval is process just described: we multiply the survival assumed to occur at the end of the year. Later. ing is expected utility decision making. Mean survival for therapy chance node.0) + (0.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 89 for choosing called expected-value decision mak. often the patient’s preferences) for the outcomes of the gamble.0) + (0. We shall discuss utility briefly The choice between therapies A and B is repre- (Sect.” In fact.4 × 2. Mean survival for therapy B = (0. By convention. called the expected survival or.7.4). Some physicians Because we cannot be sure of the duration of object to using probability for just this reason: survival for any given patient. nate. Associated with each line is the probability here for discussion only.0)  and we use that number to compare the gambles. unknowable for the individual. node in Fig.3 × 3. chance event.0) + (0. therapy A and therapy B are both Survival after a therapy is under the control of gambles with respect to duration of life after ther. Each line represents one of the possible out- remember that this model is oversimplified.0) + (0. We can cal- culate the mean survival that would be expected 9 Expected-value decision making had been used in many based on the probabilities depicted by the chance fields before it was first applied to medicine. be represented by the expected value at the pies in Example 11.35 × 4. we have chosen to use the term expected value. value associated with each possible outcome by (021) 66485438 66485457 www. We want to assign a measure (or number) to an average survival equal to 2.1 years. used comes. For a single patient.ketabpezeshki. factors. is shown as a circle from which several lines ema- vival) as a criterion for choosing among therapies. such as the quality of life. 3. only one outcome can occur.15 × 2.1 years. but you can pursue this topic and the sented diagrammatically in Fig. a chance node use the average duration of life after therapy (sur. we often life) that would be observed in a large number of must use the frequency of the outcomes of many patients after they were given the therapy. The patients experiencing the same event to inform first step we take in calculating the mean survival our opinion about what might happen to an indi- for a therapy is to divide the population receiving . Utility is the name given to a measure 3.2 R  epresentation of Choices of preference that has a desirable property for deci. patient-specific adjustments and thus estimate the ilar survival rates.0) important and is easy to understand. with Decision Trees sion making: the gamble with the highest utility should be preferred. we should select therapy B. of 3. Therapy B each therapy that summarizes the outcomes such is a gamble characterized by an average survival that we can decide which therapy is preferable. If length of life is our criterion for The ideal criterion for choosing a gamble should choosing. 3.3 years. we multiply the survival probability of each outcome at a chance node. The outcome of a products over all possible survival values.1 × 4. chance.0) + (0. more generally. can We can perform this calculation for the thera. Therapy A is a gamble characterized by apy. 3. be a number that reflects preferences (in medicine.0) + (0. This average length of life is 10 A more general term for expected-value decision mak. we characterize a “You cannot rely on population data. we consider other of the outcome occurring.7. time in each group11 by the fraction of the total A chance node can represent more than just an population in that group. From these frequencies. because therapy by the mean survival (average length of each patient is an individual. The concept of expected value is A = (0.5. Events that details of decision analysis in other textbooks (see are under the control of chance can be represented Suggested Readings at the end of this chapter). Because a full the expected value of the chance node. we sum these event governed by chance.3 years.9 = 3. In Example 11. × 1.05  ing: we characterize each gamble by a number. We calcu- treatment of utility is beyond the scope of this chapter.10 We by a chance node. late the expected value at a chance node by the 11 For this simple example.5.2 × 1. we can make the therapy into groups of patients who have sim.45 × 3. = 2.

and bon dioxide between blood and air. is the technique or expected monetary cost that it will confer on. which measuring outcomes. 3.K. We 3. four steps in decision analysis: His other major health problem is emphy- 1. known as sensitivity analysis. Create a decision tree. because it requires formulating the their ability to exchange oxygen and car- decision problem. in turn causes shortness of breath (dys- 2. this step is the most dif. while he can We clarify the concepts of expected-value deci.90 D. Thus. a 66-year-old Analysis man who has been crippled with arthritis of both knees so severely that. which also has been recog- tion of survival.1 years for therapy B. Sect. expected sense of well-being.C. he must otherwise use a wheelchair. The defense therapy is characterized by the expected dura. or dollars. sema. He is able to breathe comfortably alternative. assigning probabilities. nized by decision analysts. Consider the following clinical problem. the values that are estimates at best. making based on small differences in expected vival. because they recognize the opportunity for value decision making. respectively.3 Many health professionals hesitate when they years for therapy A and 3. vival over all outcomes. We discuss this or incur for. Danby. Use sensitivity analysis to test the conclusions patients were assigned to receive either therapy A of the analysis.7  A chance-node representation of survival after the two therapies in Fig. the patient. the expected survival would be 2. est expected value.3 P  erformance of a Decision The patient is Mr. They rea- the outcome of a therapy. The term expected value error in assigning values to both the probabili- is used to characterize a chance . against this concern. Sox Fig.6. Choose the decision alternative with the high- then sum the product of probability times sur. such as ties and the utilities in a decision tree. the alternative with the highest expected value. 3. There are canes. Owens and H. units of sense of well-being. low this strategy when there are therapy choices The first step in decision analysis is to create a with uncertain outcomes: (1) calculate the expected decision tree that represents the decision prob- value of each decision alternative and then (2) pick lem. or therapy B. we fol. If the outcomes of a son that the technique encourages decision therapy are measured in units of duration of sur. The probabilities times the corresponding years of survival are summed to obtain the total expected survival the probability that that outcome will occur. first learn about the technique of decision analy- We have just described the basis of expected-­ sis. if several hundred 4.5. a disease in which the lungs lose ficult.5.ketabpezeshki. Example 12 3. important fourth step in decision analysis in To use expected-value decision making. (021) 66485438 66485457 www. 3. get about the house with the aid of two sion making by discussing an example. Calculate the expected value of each decision pnea).

and Mr. The patient’s responses are shown in the third column of Table 3. Danby choose to many years with normal mobility do you feel is undergo knee replacement surgery. tance that the patient cannot supply given Next. Should Mr. how mobility. she first lists the doctor that he is reconsidering knee outcomes. Danby: “Many people tive period. and unchanged. however.6. quo). outcomes differ in two dimensions: length of life Mr. during or immediately after knee surgery lyzed.8.6. In your case. Mr. He tells his come. The resulting measure is called a prosthesis. According to these conventions. or equivalent in value to 10 years in your current should he accept the status quo? state of disability?” She asks him this question for each outcome. poor with significant disability. The possible She can convert years in poor health into years outcomes of knee replacement include in good health by asking Mr. Danby outcomes with two dimensions into outcomes then would have to undergo a second risky with a single dimension: duration of survival in operation to remove it. 12 QALYs commonly are used as measures of utility nating from a decision node represents an action (value) in medical decision analysis and in health policy that could be taken. Several To accomplish this task. she now requires a degree of assis. The internist chance that the prosthesis (the artificial can account for this trade-off factor by converting knee) will become infected. As you can see from Table 3. (p[operative death] = 0. Recently. internist must develop a measure that takes into Danby’s ability to survive the operation is in account these two dimensions.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 91 According to the methods of expected-value when he is in a wheelchair. if it occurs at all). patient decides that 10 years of limited mobility the internist sketches the decision tree shown in are equivalent to 6 years of normal mobility. Mr. even with canes. To accomplish this task. chance of recovering full function after surgery largely because his internist told him that (p[full recovery] = 0.60) and the chance of devel- there was a serious risk that he would not oping infection in the prosthetic joint survive the operation because of his lung (p[infection] = 0. the replacement surgery. Fig. The Using the conventions of decision analysis. Possible func. Danby’s internist is familiar with (survival) and quality of life (functional status). She uses her subjective esti- disease. a square box denotes a decision node. she must assign a value to each out- his present state of mobility. he seriously considered knee orthopedic surgeons for their estimates of the replacement surgery but decided against it. the internist asks several years ago. she asks Mr. 3. Danby’s mate of the probability that the patient will die wife had a stroke and was partially para. Danby to indicate death from the first procedure and death the shortest period in good health (full mobility) from a second mandatory procedure if the that he would accept in return for his full expected prosthesis becomes infected (which we will lifetime (10 years) in a state of poor health (status assume occurs in the immediate postopera. but the effort of decision making. and the surgery sometimes does not duration of survival is inadequate because Mr. heavily and feel uncomfortable. Thus. Furthermore.05).com . After removal of the good health. decision analysis. Simply using doubt. there is a small he values 10 years of poor health.ketabpezeshki. She recognizes that this To characterize each outcome accurately.12 able to walk. and each line ema. (021) 66485438 66485457 www. Danby would never again be quality-adjusted life year (QALY). the internist first must assign a walking with canes makes him breathe probability to each branch of each chance node. the problem is filled with uncertainty: Mr.05). restore mobility to the degree required by Danby values 5 years of good health more than such a patient. analysis. say they would be willing to accept a shorter life tional outcomes include recovery of full in excellent health in preference to a longer life mobility or continued.

4 × 0.95 = 7. 1. prosthesis: 0 + 2.85 QALYs. 3. Multiply the probability of opera- (years) Functional status outcome tive death (0. Let us consider. Add the expected values calculated in step 1 (0 QALY) and step 2 (2.ketabpezeshki. of surgery and of no surgery.85 QALYs. Sox Fig.143 QALYs. for each outcome.92 D. the circles represent chance nodes Table 3. in healthy follows: years.6  Outcomes for Example 12 1. Multiply the expected value of never the outcome of surgery to remove an infected developing an infected prosthesis (already ­ prosthesis (Node A in Fig. Similarly. She calcu. The calculation calculated as 8.K.143 + 7.85 QALYs) to obtain whereas 10 years of wheelchair confinement are the expected value of developing an infected equivalent to only 3 years of full function. Add the expected values calculated in step 1 13 In a more sophisticated decision analysis.C.95): 8. The box represents the decision node (whether to have surgery). for example. the expected value at chance node B plete with probability estimates and utility values is calculated: (0. 3.8  Decision tree for knee replacement . Figure  3. 3. the clinician also would adjust the utility values of outcomes that (0. quo or unsuccessful surgery) Multiply the probability of surviving the oper- 10 Wheelchair-bound 3 ation (0. replacement surgery (Node C).9). surgery) 2.9 shows the final decision tree—com. Multiply the expected value of an infected pros- lates the expected value at each chance node.85 × 0. expected value at the chance node representing 2.98 QALYs) to get require surgery to account for the pain and inconvenience the expected value of surviving knee replace- associated with surgery and rehabilitation. Calculate the expected value of surviving sur- 10 Poor mobility (status 6 gery to remove an infected knee prosthesis.05 × 0 = 0 QALY.4 × 6) =8. Owens and H. (021) 66485438 66485457 www.95 × 3 = 2. Calculate the expected value of operative Years of full death after surgery to remove an infected function Survival equivalent to prosthesis.05 =0.85 = 2.123 QALYs.13 To obtain the expected value of surviving knee The second task that the internist must under.85 QALYs) by moving from right (the tips of the tree) to left (the the probability that the prosthesis will become root of the tree).98 = 8.6 × 10) + (0.143 QALY) and step 2 (7.95) by the number of healthy years (the outcome if a second equivalent to 10 years of being wheelchair-­ surgery is necessary) bound (3 years): 0.4 QALYs) by the probability requires three steps: that the prosthesis will not become infected (0.05): 2. thesis (already calculated as 2. she proceeds as take is to calculate the expected value.05) by the QALY of the out- 10 Full mobility (successful 10 come—death (0 years): 0. the infected (0.98 QALYs. ment surgery: 0.4 QALYs. 0 Death 0 3.

Thus. and well. the best result on the average in similar patients. Not have similar probabilities. Mr. the best surgery. by accepting surgery. the average of the analysis represent the outcomes that would length of life.ketabpezeshki. measured in years of normal . The average the expected value of surgery has been calcu. however. the 7. will be 6.7. Thus. length of life.7. and some do not choice for the individual is the alternative that gives gain any improvement in mobility after surgery. the results In the analysis for no surgery. is 6. Mr. adjusted to the individual. The patient’s valuations of outcomes The clinician performs this process. have similar utilities and for whom uncertain events lent to 10 years of continued poor mobility. mal mobility. (021) 66485438 66485457 www. (measured in years of perfect mobility) are assigned to Probabilities have been assigned to each branch of each the tips of each branch of the tree chance node. Nonetheless. We can understand ysis is tailored to a specific patient in that both the what this value means for Mr.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 93 Fig. If Decision analysis can help the clinician to identify the clinician had 100 similar patients who under. Danby only by utility functions and the probability estimates are examining the alternative: no surgery. called all patients will experience this outcome. that. Because 6. 7. What does this value mean? It does not mean internist recommends performing the surgery. 10 years. The key insight of expected-value decision mak- teed 7. Danby is guaran. on average the surgery will provide an out- tancy. measured in years of normal mobil. is come with higher value to the patient. some develop infection.7 years of mobile life. however. working back to the root of the tree. Danby considered equiva. occur on average in a population of patients who ity. the average number aged over many similar patients. 3. expressed in years of nor- lated. which Mr.7. The outcome of the analysis is as follows. Danby’s average life expec.9  Decision tree for knee-replacement surgery.0 is less than For surgery. an individual patient has no guarantees. for node D as who have poor mobility will live longer than. One look at the ing should be clear from this example: given the decision tree will show that some patients die in unpredictable outcome in an individual.0. until some will live less than. the therapy that will give the best results when aver- went the surgery. some averaging out at chance nodes. The decision anal- of mobile years would be 7.

it also assesses a person’s willingness to life with the outcome. The utility of a health time-trade-off technique. chance of immediate death (which has an ble with a chance of ideal health or immediate expected value of 20 years). (1988). respectively. Because the time trade-off does not include gam- To illustrate use of the standard gamble. we use ideal health itself. might be indifferent when the probability of ideal Several other approaches are available to health is 0. achieving the outcome (e. health (usually ideal health or best attainable Utilities are typically expressed on a 0 to 1 scale.5.8) + (0 × 0. and 0.5. a study of patients who had infection.8. 12 months of life with asymptomatic HIV infec- 0.95.3. and severe angina as 0. immediate death systematically until the sub. it does not assess a person’s risk we seek to assess a person’s utility for the health attitude. atic HIV infection as 8  ÷ 12 = 0. Use of the standard QALYs.g. the utility of asymptom- In Sect. Perhaps the strongest assumption under- state of asymptomatic HIV infection. Although the (021) 66485438 66485457 .K. the utility of the scale. Nonetheless. suppose bles. the time-trade-off technique is used frequently to ject is indifferent between asymptomatic HIV value health outcomes because it is relatively infection and the gamble.94 D. In practice. because length of life is not the only out. How ment is the time-trade-off technique (Sox et al. Patients’ prefer. however. ences for a health outcome may depend on the Because the standard gamble involves chance length of life with the outcome. Thus in this example. and on the risk involved in take risks—called the person’s risk attitude. For example. as shown by Von Neumann and gains (or losses) in both length and quality of life. of death. asymptomatic HIV are equal. For example. time-trade-off technique provides a convenient gest theoretical basis of the various approaches to method for valuing outcomes that accounts for utility assessment. Morgenstern and described by Sox et al. we can represent patients’ utility of asymptomatic HIV infection using the preferences with utilities. then we calculate the utility of asymptom- There are several methods for assessing utilities. a person simply rates the quality of life gamble and that of asymptomatic HIV infection with a health outcome (e. can we incorporate these elements into a decision 1988. value health outcomes. gamble enables an analyst to assess the utility of come about which patients care. Torrance and Feeny 1989). For example. we introduced the concept of atic HIV infection is 0. few people are risk-neutral. moderate. health) that he or she would find equivalent to a where 0 represents death and 1 represents ideal longer period of time with asymptomatic HIV health. 1995). The The standard-gamble technique has the stron. To assess the analysis? To do so.2)] = 0. we ask our subject to compare utility is that people are risk neutral. outcomes that differ in length or quality of life. We then ask our subject to choose with a 50 % chance of living 40 years and a 50 % between asymptomatic HIV infection and a gam.8 and the probability of death is 0.ketabpezeshki. We calculate the utility of the g­ amble infection) on a scale from 0 to 100.g. To use the visual analog At this point of indifference.. To use the lying the use of the time trade-off as a measure of standard gamble. if our subject says that 8 chest pain (angina) with exercise rated the utility months of life with ideal health was equivalent to of mild.82 (Nease et al. tion. We vary the probability of ideal health and course. Owens and H. Often. on the quality of events. 3.C. a risk-neutral decision maker (assigned a utility of 1) and immediate death would be indifferent between the choice of living (assigned a utility of 0) for the comparison of 20 years (for certain) and that of taking a gamble health states.4 R  epresentation of Patients’ as the weighted average of the utilities of each Preferences with Utilities outcome of the gamble [(1 × 0. a subject easy to understand. we ask a person to state is a quantitative measure of the desirability determine the length of time in a better state of of a health state from the patient’s perspective..2. For example.92. a cure for cancer A second common approach to utility assess- might require a risky surgical operation). Sox 3.8. A risk-­ the desirability of asymptomatic HIV infection to neutral decision maker is indifferent between those of two other health states whose utility we the expected value of a gamble and the gamble know or can assign.

Figures 3. tion will lead to perfect mobility. respectively. time trade-offs.ketabpezeshki. Sensitivity analysis is a test of the validity of the Each point (value) on these lines represents one conclusions of an analysis over a wide range of calculation of expected survival using the tree in assumptions about the probabilities and the val. Expected survival is lower is available. The point at that differ in length and quality of life and in risk. the Health Utilities Index. 1995). but there often is a wide range of with surgery.8. The expected clusions regarding the preferred choice change survival. is higher for surgery as when the probability and outcome estimates long as the probability of perfect mobility exceeds are assigned values that lie within a reasonable 20 %. Ratings with the visual analog scale. effect of varying the probability that the opera- ity analysis to answer this question: Do my con. Computer-based tools with an interactive The solid line represents the preferred option at a given format have been developed for assessing utili. the internist can proceed with c­ onfidence (021) 66485438 66485457 www.5. As the In summary. the recommendation is trustworthy.11 shows the with nearly equal confidence. they often include text and multimedia pre- sentations that enhance patients’ understanding of the assessment tasks and of the health out. Figure 3. 1996. If the estimated the chance of full recovery at 60 %. however. (2009) and (2012). and outcome measures. probability ties. (In The knee-replacement decision in Example 12 Example 12. 4).10 and 3. correlate modestly well with utilities assessed by the standard gamble and time trade-­ off. a much lower figure than is expected from range? previous experience with the . and the visual analog scale to assess utilities in patients with angina. Figure 3. which the two lines cross represents the probability of operative death at which no surgery becomes preferable. Each of these instruments assesses how people value health outcomes and therefore may be appropriate for use in decision analyses Fig. see Joyce et al. in patient living with HIV. or utilities. mortality on length of healthy life (Example 12). 3. and the EuroQoL (see Gold et al. the relative val- how patients value complicated health outcomes ues of surgery versus no surgery change. 1991.10 shows that expected survival ues. the consulting orthopedic surgeons illustrates the power of sensitivity analysis. We use sensitiv. Lenert et al. 3. it has no theoretical justification as a valid measure of utility.5 P  erformance of Sensitivity and without surgery under varying assumptions Analysis of the probability of operative death and the prob- ability of attaining perfect mobility. For a demonstration of the use of standard gambles. in healthy years. to no surgery) remain the same despite a wide comes (Sumner et al. Fig. see Nease et al. when the operative mor- ­reasonable probabilities that a clinician could use tality rate exceeds 25 %. Nease and Owens range of assumed values for the probabilities 1994. The probability of an outcome is higher with surgery over a wide range of opera- at a chance node may be the best estimate that tive mortality rates. Other approaches to valuing health outcomes include the Quality of Well-Being Scale. we can use utilities to represent probability of operative death increases.11 show the expected survival in healthy years with surgery 3. however.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 95 visual analog scale is easy to explain and use. ch.) conclusions of the analysis (surgery is ­preferable Thus. (1995).10  Sensitivity analysis of the effect of operative or cost-effectiveness analyses. iness.

The software then uses As the probability of a successful surgical result increases. For each evalua- The point at which the two lines cross represents the prob. But what if we want to model events that might occur in the dis- tant future? For example.6 R  epresentation of Long-Term the probability of operative death at which the Outcomes with Markov two therapy options have the same expected sur. poor mobil- should select surgery. which was ­surgery when the probability of operative death is compromised by arthritis. From the probabilistic sensitivity analysis. The analyst may have uncertainty about many param- eters in a model. a decision tree or decision model may have a 100 or more . in the future. tion might develop AIDS 10–15 years after however.10 cross is 3. Danby cannot be sure is preferred. We assumed that each of 25 %. this set of parameter values and calculates the the relative values of surgery versus no surgery change. the analyst can determine the proportion of times an alternative to recommend surgery.C. the analyst must specify a probability distribution for each model parameter.000 times. Danby took action on his decision. thus. Owens and H. occur 10–15 years. But in a complex problem. this value is correct when we use QALYs as the infection. Mr.) would occur shortly after Mr. Probabilistic sensitivity analy- sis is an approach for understanding how the uncertainty in all (or a large number of) model parameters affects the conclusion of a decision analysis. probability distribution. If expected survival is to be the basis for choosing therapy. death. a therapy to prevent or delay the basis for choosing treatment. Sox The approach to sensitivity analyses we have described enables the analyst to understand how uncertainty in one. tion of the model. Models vival. see the arti- surgery than he is without it. To perform a probabilistic sensitivity analysis. etc. end of the chapter. accounting for all uncertainty in of a good outcome. the internist and the patient In Example 12. 3. a patient with HIV infec- 14 An operative mortality rate of 25 % may seem high. When it is higher. For more thinking that he is more likely to do well with information on this advanced topic. should select no surgery. 3. they ity.11  Sensitivity analysis of the effect of a success- ful operative result on length of healthy life (Example 12). sitivity analysis is to indicate the range of proba- bilities over which the conclusions apply. Danby’s decision should be indifferent between surgery and no to have surgery to improve his mobility. or three parameters affects the conclusions of an analysis. two. The point at which the two lines in Fig.5. A decision maker perform- ing a more sophisticated analysis could use a utility func.96 D.ketabpezeshki. outcomes for each alternative. The ana- lytic software then chooses a value for each model parameter randomly from the parameter’s Fig. cle by Briggs and colleagues referenced at the Another way to state the conclusions of a sen. the software will determine ability of a successful result at which surgery becomes preferable. The process is at a given probability usually repeated 1.000–10. they the possible outcomes (full mobility. but he has valid reasons for model parameters simultaneously. we evaluated Mr.14 When the probability is lower. development of AIDS could affect events that tion that reflects the patient’s aversion to risking death.K. The solid line represents the preferred option which alternative is preferred. A similar (021) 66485438 66485457 www. or more.

How determine from a Markov model the expected can we include such events in a decision analysis? length of time that a person spends in each health The answer is to use Markov models (Beck and state.. alternative represented by a Markov model. Sonnenberg and Beck 1993.7  Transition probabilities for the Markov model in Fig.7 for the cycle length of 1 year. we could model a cancer-­ cycle. 3. 3. Sanders et al. 3.7 that a person who is in the arrows represent allowed transitions between health states well state will remain well with probability 0. tree representation is convenient for decisions for the probabilities that a person remains well. or quality-adjusted life expectancy. (See the articles by Beck we can calculate the probability that a person will and Pauker (1983) and Sonnenberg and Beck be in each of the health states at any time in the (1993) for further explanation of the use of future. we first specify the In decision analyses that represent long-term set of health states that a person could experience outcomes. We can also include events that could occur in the future.0 Death to cancer 0.06.8. 2005).12). Therefore. develops cancer. will develop cancer with probability 0. and Death (see Fig. Siebert tancy. As an illustration.13 Health state transition Annual probability Well to well 0. and Death in Fig.12). Cancer. which model the decision (Owens et al.) (021) 66485438 66485457 www. but it is not always sufficient for problems that over time is shown in Table 3. Cancer. 1997. tion in the transition probability from Well to fied number of cycles. We model in conjunction with a decision tree to then specify the transition probabilities. we can determine life expec- Pauker 1983. and will die from non-cancer causes with probability problem arises in analyses of decisions regarding 0.4 Cancer to death 0.0 Markov model that has three health states: Well. by using a Markov model.0 Death to death 1.06 Well to death 0.g. Salpeter are the probabilities that a person will transit et al. The calculations for a Markov many chronic diseases: we must model events that model are performed by computer software. or dies from non-cancer causes zon.6 Death to well 0.12.0 Cancer to cancer 0. Thus. This period—often 1 in the probability of going from one state to month or 1 year—is the length of the Markov another.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 97 Table 3. 3. consider a simple Markov models. the analysts will often use a Markov (e.04 Cancer to well 0. The states of health Table 3. Well. We have speci- fied each of the transition probabilities in Fig. for any et al. occur over the lifetime of the patient.7.04 during 1 year. we that a person can experience are indicated by the circles.9. which all outcomes occur during a short time hori.ketabpezeshki. note from Table 3. 3. 1995. For example.9 Well to cancer 0. The decision Based on the transition probabilities in Table 3. The analyst from one of these health states to another during models the effect of an intervention as a change a specified time . 2012). The Markov model then simulates the prevention intervention (such as screening for transitions among health states for a person (or breast cancer with mammography) as a reduc- for a hypothetical cohort of people) for a speci. Cancer in Fig. To build a Markov model.12  A simple Markov model.

(In this example. ease is the probability of disease at which you 3. this kind of problem. 3. tests nor treat the patient). (Fig.0757 0. but you have already learned all the principles that you need to solve The clinician who is evaluating a patient’s symp. (test) before choosing whether to treat or do The treatment threshold probability of dis- nothing.e.0780 0. Owens and H. such as drugs disease of concern. Learning the patient’s true state. clinicians You can use this tree to learn the treatment often are willing to treat a patient even when they threshold probability of disease by leaving the are not absolutely certain about a patient’s true probability of disease as an unknown.6561 0. 3.0600 0. Determine the pretest probability of disease. There are risks in this course: the clinician expected value of surgery equal to the expected may withhold therapy from a person who has the value for medical (i.1120 0. should be indifferent between treating and not When the clinician knows the patient’s true state. At probabilities of disease that are greater than the treatment threshold probability.7290 0.1912 0. 3. Treat without obtaining more information. you should treat in Example 12). Below the testing is unnecessary.5314 0.4645 Fig. misleading FP or FN results.9000 0.6  he Decision Whether T Deciding among treating.K. the preferred action is to treat . Obtain additional diagnostic information decision to treat.0572 Death 0. or Do Nothing nothing sounds difficult. or he may administer therapy to or physical therapy) treatment. Above assess the trade-offs among therapeutic options (as the treatment threshold. and the doctor needs only to treatment threshold. gery corresponds to the “treat” branch of the tree (021) 66485438 66485457 www.14.0633 0.5905 0. Test. Decide whether a test result could affect your 2.8100 0. may require costly. At probabilities of disease that are less than the treatment threshold probability. setting the state. Therefore.ketabpezeshki.3399 0. Determine the treatment threshold probability the following actions: of disease. 1. Whether to treat when the diagnosis however. and doing to Treat. the preferred action is to withhold therapy.4053 0. and solving for someone who does not have the disease yet may the probability of disease..0696 0.8 Probability Health of future health states state Probability of health state at end of year for the Markov model Year 1 Year 2 Year 3 Year 4 Year 5 Year 6 Year 7 in Fig. sur- suffer undesirable side effects of therapy. such as the one shown in Fig. Sox Table 3.0798 0. Do nothing further (neither perform additional 2.13).4783 Cancer 0.0400 0.13 Depiction of the treatment threshold probability. 3. 3.C. treating (Pauker and Kassirer 1980). nonsurgical.12 Well 0. and is not certain is a problem that you can solve with often risky diagnostic procedures that may give a decision tree. you should not treat. testing. time-­consuming.98 D.2682 0. There are three steps: toms and suspects a disease must choose among 1.

sion making is this: do not order a test unless it responds to the “do not treat” branch. If the pretest probability is above should be indifferent to the choice. In contrast. We define H as the differ- ence in utility of nondiseased patients who are not treated and nondiseased patients who are treated (U[−D.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 99 treat]. The equation for the treatment threshold probability fits with our intu- ition: if the benefit of treatment is small and the harm of treatment is large. therefore. This analysis where p* = the treatment threshold probability. Once you know the pretest probability. we calculate the who are not treated (U[D. 1988). If the pretest probability is patients who are treated and diseased patients above the treatment threshold. (021) 66485438 66485457 www. this principle and surgery at this probability. 3. you should not compute the probability at which the clinician and patient treat the patient. We define B as would move the probability of disease toward the the difference between the utility (U) of diseased treatment threshold. the treatment threshold probability will be low. as shown in Fig. The utility of dis- eased patients who are treated should be greater than that of diseased patients who are not treated. mation about the patient. the treatment threshold probability is to use the if the pretest probability is below the threshold equation: probability.14). you know what to do in the absence of further infor- Fig. you should treat the patient. 3. we can is below the treatment threshold. the test result would alter your deci- H+B sion of whether to treat the patient.14). To decide whether a test could alter manage- diseased patient. Thus. We calcu- with treatment of a diseased patient (Pauker and late the post-test probability after a test result that Kassirer 1980. result could cause the probability of disease to pletes step 1. means that you should order a test only if the test ment threshold probability. as shown in Fig. 3. In you are indifferent between medical treatment our framework for decision making.ketabpezeshki. 3. Using the tree com. completes step 3.14. either case. Recall that p the threshold. Conversely. therefore. treat]. treat] − U[D. The utility of nondiseased patients who are not treated should be greater than that of nondiseased patients who are treated. and B = the benefit associated ment. In H p* = . negative test result must lead to a post-test prob- An alternative approach to determination of ability that is below the threshold. it is the treat. if the pretest the treatment threshold intuitively rather than probability is above the treatment threshold. people often determine cross the treatment threshold. we simply use Bayes’ theorem. H = the harm associated with treatment of a non. a positive result must lead to a post- test probability that is above the threshold. do not treat] − U[−D. B is positive. a analytically. do not probability of disease if the test result is negative. By setting the utili.) Because could change your management of the patient. if the benefit of treatment is large and the harm of treat- ment is small. If the pretest probability ties of the treat and do not treat choices to be equal. In practice.14  Decision tree with which to calculate the treat- ment threshold probability of disease. Thus [−D] = 1 − p [D] you have completed step 2. H is positive. and nonsurgical intervention cor. One of the guiding principles of medical deci- in . the treatment thresh- old probability will be high. Sox et al.

This example is especially useful for two rea- A relatively low treatment threshold is justifiable sons: first.053.. to do. result would change your management of the old means that you are willing to treat nine patient. You estimate the pretest probability of pul- You suspect that a patient of yours has a monary embolus to be 0. (021) 66485438 66485457 www. probability of pulmonary embolus is higher than bility was below 0.10. You review the literature and learn that the LR for a positive CTA scan is approximately 21 (Stein et al.e. it demonstrates one method for mak- because treatment of a pulmonary embolism with ing decisions and second. Decide whether a test result could affect your test probability were above the treatment thresh- decision to treat for an embolus. dye flows into the vessels of the lung. if you believed that there was greater than a A post-test odds of 1.10? If you could obtain no fur- Post-test odds = pretest odds × LR ther information. Sox If the pretest probability is below the treatment (mortality of less than 1 %) in treating someone threshold. a test in which a computed result could raise the probability of pulmonary tomography (CT) of the lung is done after a embolus to above 0. ability equal to 0. One approach is ability is lower than the treatment threshold. Estimate the pretest probability of pulmonary of disease toward the treatment threshold and embolus.10. You patients without pulmonary embolus to be sure of have completed step 3. Owens and H. 2006). the treatment threshold. the benefit of treatment is high and the harm of treatment is low. you would treat for pulmonary embolus if the pretest probability was above 0. If the scan is negative.53.ketabpezeshki.K. patient. A positive result will move the probability 2.05. To decide whether this strategy is correct. Determine the treatment threshold probability threshold and will be of no help in deciding what of pulmonary embolus. the treatment threshold proba- bility will be low. A decision to treat when the treatment threshold. assume you decide that the treatment Bayes’ theorem to calculate the post-test proba- threshold should be 0. old. you to do a computed tomography angiography should do nothing unless a positive CTA scan (CTA) scan. The step 2. You have completed radiopaque dye is injected into a vein.11. it shows how the blood-thinning medication substantially reduces concepts that were introduced in this chapter all the high mortality of pulmonary embolism. The To decide whether a test result could affect CT scan can then assess whether the blood your decision to treat. as discussed . as high probability.10. (i. which is equal to a pulmonary embolus (blood clot lodged in pretest odds of 0. ability of disease of 0. You therefore use the odds-ratio form of First. could alter your management decision if the post-­ 3. a positive CTA scan the pretest probability is at the treatment thresh. fit together in a clinical example of medical deci- whereas there is only a relatively small danger sion making. you must decide whether a vessels are blocked. You are a pulmonary medicine specialist. and you should order the lung scan. You have Example 13 completed step 1. What bility of disease if the lung scan result is reported does it mean to have a treatment threshold prob. positive CTA scan result would raise the proba- you do no further tests and do not treat the bility of pulmonary embolism to more than 0. treating one patient who has pulmonary embolus. Because the pretest prob- the vessels of the lungs). we calculate the probability of disease who does not have pulmonary embolus.C. you A negative CTA scan result will move the take the following steps: probability of disease away from the treatment 1. Because the post-test and would withhold therapy if the pretest proba.053 × 21 = 1.100 D. Because if the test result is positive.10 = 0.1 is equivalent to a prob- 1 in 10 chance that the patient had an embolus).10 in this patient.

With the appro- tionship may exist between the chance nodes priate software.5. thus. 1997). 1997). 3. in which the events branches of the tree. Influence diagrams represent decision determine the expected value of each alternative nodes as squares and chance nodes as circles. tioning is complex or in which communication of however. With a decision tree. Thus. these nodes advantages and limitations relative to decision have a probabilistic relationship.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 101 3. Owens determine the alternative with the highest expected et al. influence diagrams from Obtain PCR? to Treat? are an important alternative representation for The probabilities and utilities that we need to such problems (Nease and Owens 1997. value are contained in tables associated with As shown in Fig. formed with computer . In contrast. Arcs between deci- and Belief Networks sion nodes indicate the timing of decisions: the arc points from an initial decision to subsequent In Sects. 3.5 and 3. The arc points from the conditioning probabilistic relationships among variables event to the conditioned event (PCR test result (Owens et al.15. deciding whether to treat. we can use influence diagrams to (Owens et al. always indicates that the nodes are such conditioning is important (such as may occur independent or conditionally independent. if the occurrence of one of the events thus the conditioning is apparent immediately by does not affect the probability of the other event inspection.6. tioning is revealed by the probabilities in the Unlike a decision tree. we can ments. A probabilistic relation. Influence dia- An arc from a chance node to a decision node grams also are particularly useful for discussion indicates that the chance event has been observed with content experts who can help to structure a at the time the decision is made. In by averaging out at chance nodes and folding back contrast to decision trees. events between branches of the tree. 1997). 3. as indicated by trees. but they also have additional graphical ele. To determine whether events usually are represented from left to right in the are conditionally independent in a decision tree order in which the events are observed. In an influence diagram. given a listic conditioning is indicated by the arcs. we used decision trees to decisions. For example.15. problems that have decision (021) 66485438 66485457 www. Although decision must decide whether to obtain a PCR test before trees are the most common graphical representa. 3. Thus. use in a decision tree. the diagram also has arcs between nodes and a calculation of expected value is more complex diamond-­shaped value node. perform the same analyses that we would perform ship exists when the occurrence of one chance with a decision tree.3). For influence diagrams. and generally must be per- chance nodes indicates that a probabilistic rela. the influence the tree (Sect. 3. The tageous for problems in which probabilistic condi- absence of an arc between two chance nodes.15 indicates analysis. them to perform probabilistic inference. probabilistic condi- conditioned on the occurrence of the third event. the arc problem but who are not familiar with decision from PCR result to Treat? in Fig.15. 3. as indicated by the arc tion for decision problems.7 Alternative Graphical that the decision maker knows the PCR test result Representations for Decision (positive. In a decision tree.16). the decision maker represent decision problems. 1997). An arc between two (Owens et al. the probability of a positive or negative PCR test Why use an influence diagram instead of a result (PCR result) depends on whether a person decision tree? Influence diagrams have both has HIV infection (HIV status). Diagrams that have only event affects the probability of the occurrence of chance nodes are called belief networks. influence diagrams chance nodes and the value node (Fig. in Fig.ketabpezeshki. 3. influence requires that the analyst compare probabilities of diagrams use arcs to indicate the timing of events. however. 3. probabi- events are conditionally independent. we use another chance event. or not obtained) when he or Models: Influence Diagrams she decides whether to treat.15). Two in large models). These have certain features that are similar to decision tables contain the same information that we would trees. Such representation is advan- is conditioned on HIV status in Fig. in Fig. negative. Influence diagrams represent graphically the the arc. and third event.

whereas the influ- easier for people to understand when represented ence diagram does not. HIV infection. The choice of whether to with a decision tree.K. QALE quality-adjusted life expectancy.102 D. HIV+ HIV infected. because the tree shows the use a decision tree or an influence diagram depends (021) 66485438 66485457 www.ketabpezeshki. Test results are shown in quotation marks (“HIV+”). (1997).com . and to treat. PCR polymerase chain reaction. The influence diagram highlights probabilistic relationships. HIV human immunodeficiency virus. The structural asymmetry of the alternatives is explicit in the decision tree. Sox Fig. 3. Owens and H. HIV− not infected with HIV.15  A decision tree (top) and an influence diagram (bottom) that represent the decisions to test for. Reproduced with permission) alternatives that are structurally different may be structural differences explicitly.C. whereas the true disease state is shown without quotation marks (HIV+) (Source: Owens et al.

ments (for example. Markov models in the compartment. results are shown in quotation marks (“HIV+”). similar to For more information on these types of models. PCR polymerase chain reaction. such as might occur with cancer.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 103 Fig. Dynamic the analyst. and the transmission models are particularly well-suited purpose of the analysis. The depend on the number of infected or susceptible choice of modeling approach depends on the individuals.8 Other Modeling Approaches tions. 3. Test The information in these tables is the same as that associ. dead). with the adjusted life expectancy. and the ated with the branches and endpoints of the decision tree true disease state is shown without quotation marks in Fig. (021) 66485438 66485457 www. Discrete event simulation models problem and the objectives of the analysis.ketabpezeshki. Reproduced with HIV infected. the paper by Caro and colleagues useful when the clinical history of a problem is noted in the suggested readings at the end of the complex. infected. An entity making. These models are composed of entities (a is beyond our scope. the experience of disease. 3. NA not applicable. TX+ treated. for assessing the outcomes of infectious diseases. also are often used to model interactions between Although how to choose and design such models people. heart chapter is an overview of this series of . level health state transition models.15. and other chronic diseases. that provide a means to model we suggest a recent series of papers on best mod- very complex events flexibly over time. probability and value tables associated with the nodes. HIV− not infected with HIV. TX− not treated. an important feature for and influence diagrams. the availability of software. QALE quality. influence diagrams provide a powerful graphical These models divide a population into compart- alternative to decision trees. recov- ered.16  The influence diagram from Fig. 3. HIV human immunodeficiency virus. For selected problems. An analyst also can infectious diseases in which the transmission may choose several other approaches to modeling. uninfected. HIV+ (HIV+) (Source: Owens et al.15. Microsimulation models are individual. and that analysts use commonly for medical decision that experience events (a heart attack). The rate of transition between compart- ments depends in part on the number of individuals We have described decision trees. permission) on the problem being analyzed. (1997). Markov models. They are eling practices. and transitions between compartments are governed by differential or difference equa- 3.. can interact with other entities and use resources. we note other type of models patient) that have attributes (clinical history).

Fortunately. however. simply drawing a tree decision model available for analysis to anyone that denotes the possible outcomes explicitly will who has access to the Web (Sanders et al. that uncertainty in the clinical AIDS virus. clinical populations (Owens and Nease 1997). In Example 1. comes.6 to 0. we can screen for HIV.C. reservations about decision analysis because the sion analysis might be integrated smoothly into analysis may depend on probabilities that must be medical practice. In so doing. appeared to be an accurate way to is difficult to obtain from the medical literature. cover that it is necessary to estimate only a range tions for drug use. one-third of the positive tests particular variable should concern us. on casual the information needed to make the estimate often inspection. and coupled with the need to control costs. for example. uncertainty is sensitivity analysis: we can exam- cally relevant population. estimated. Decision models have many advantages as Although we should try to interpret every kind aids to guideline development (Eddy 1992): they of test result accurately. we can decide whether uncertainty about a ees used drugs. even a chapter.8 makes a difference test results is widespread. Other decisions involve lit. and utilities of potential out- decisions require decision analysis. we the estimate may be in error. 1999). such as the pretest probability. MEDLINE and using expert clinicians’ subjective probability other bibliographic retrieval systems (see Chap. much more accurate tests are used to to the final recommended decision. we often dis- ment screens civil employees in “sensitive” posi. Our quantitative analysis. to incorporate tle uncertainty. however. associ- more selective role in medicine. in actual ine any variable to see whether its value is critical practice. clarify the question sufficiently to allow you to We have not emphasized computers in this make a decision. guideline developers and users by making the For many problems.9  he Role of Probability T might be. and to tailor guidelines for specific have outcomes that cannot be known with cer. determine. would be FPs. should be of great interest to the public. and if 10 % of the employ. lines. els can provide distributed decision support for aging these situations. 22). If the of probabilities for a particular variable rather drug test used by an employer had a sensitivity than a precise value. particularly because discussed a hypothetical test that. data is a problem for any decision-­making method revealed that the hypothetical test results were and that the effect of this uncertainty is explicit misleading more often than they were helpful with decision analysis. screen blood donors for previous exposure to the We argue. Thus. The federal govern. in the final decision. Decision models can help guideline sions depend on physiologic principles or on developers to structure guideline-­ development deductive reasoning.K.95. ysis. whether a change in pre- The need for knowledgeable interpretation of test probability from 0. addition. A ity and measures of test performance will prevent thoughtful decision maker will be concerned that any number of misadventures.104 D. Nonetheless. Web-based interfaces for decision mod- Decision analysis provides a technique for man. you can perform an analysis quickly and Decision Analysis and learn which probabilities and utilities are the in Medicine important determinants of the decision. Some deci. The method for evaluating because of the low prevalence of HIV in the clini. Thus. 21) estimates and asking what the patient’s utilities make it easier to obtain published estimates of (021) 66485438 66485457 www. with a sensitivity anal- and specificity of 0. By of decision analysis (see Chap.ketabpezeshki. An understanding of these issues The growing complexity of medical decisions. Health care professionals sometimes express You may be wondering how probability and deci. decision analysis has a make explicit the alternative interventions. In tainty at the time that the decision is made. although they can simplify many aspects “quick and dirty” analysis may be helpful. Owens and H. Sox 3. problems (Owens and Nease 1993). must be based on imperfect data. An understanding of probabil. as do many companies. many decisions patients’ preferences (Nease and Owens . Not all clinical ated uncertainties. and they will Owens 1998). When time is limited. has led to health professionals should be prepared to answer major programs to develop clinical practice guide- the questions of their patients.

respectively. Probability theory in the use of diagnostic tion of population datasets. The probability of death M. D. as well as problems with uncertain outcomes. 60–66. D. H. Sox. M. D. analysis offers such aid. & Owens. patient. A. Judgment under important medical problems for which decision uncertainty: Heuristics and biases. a patient about to undergo CABG sur- Caro. (021) 66485438 66485457 www. Medical Decision Making. and dyspnea). (2001). 1996). Siegel. to calculate expected values. This book provides authoritative guidelines for the conduct of symptoms is 0. for example. Model parameter estimation and uncertainty analysis: A report of the ISPOR-SMDM modeling good research practices task force-6. & Paltiel. Karnon.. Jr.. K. Tversky. M. P. Use of influence used on data collected by hospital information sys. (b) Two known complications of heart Einstein. Researchers continue Representation and analysis of medical decision prob- to explore methods for computer-based automated lems with influence diagrams. (1986). Briggs. mutually exclu- A report of the ISPOR-SMDM modeling good sive outcomes of surgery are research practices task force-1. F. Decision analysis software. What is the of cost-­effectiveness analyses. This textbook addresses in detail most of the topics introduced in this chapter. This article provides a comprehensive introduction to the use of influence diagrams. Higgins.05. Addison-Wesley. H.. K.. models and use of computer-based systems to Raiffa. & Kahneman. Reading: growing maturity of this field.. programs for performing statistical analyses can be Nease. J.80.02.. E. diagrams to structure medical decisions. This now classic book provides an panies that offer formal analytical tools to assist advanced. (1974). A.. & Nease... relief of symptoms (angina papers that describe best modeling practices. 15. Value in Health. perform sensitivity analyses. Chichester: Wiley- welcome tools that help them make decisions Blackwell. With the tures on choices under uncertainty. Calculate the following probabilities for model parameters and for performing sensitivity ­analyses. it contains a summary of uncertainty for the clinician and risk for the the c­ oncepts of probability and test interpretation. tinue to have symptoms? Hunink... This Medical decision making often involves article is written for clinicians. This now classic article provides a clear and interesting discussion of the experimental evidence for the use and misuse of heuristics in situations of uncertainty. & Owens. decision trees. This article pro- sonal computers. and decision trees. A. and the probability of relief cine. Medical Decision Making. M. development of practice guidelines from decision 17(3). There are discussing many other topics. This introductory textbook covers the sub- when they are confronted with complex clinical ject matter of this chapter in greater detail. Decision analysis: Introductory lec- implement guidelines (Musen et al. 32(5). available for per..15 tests. Cambridge: Cambridge surgery are stroke and heart attack. J. gery (see Example 2): Modeling good research practices – overview: (a) The only possible. Suggested Readings Briggs. K. This article describes best practices for ­estimating 1. D. the Archimedes tools described at ­ . 1124. The patient asks what chance he or she has of having both 15 See. L.ketabpezeshki. Russell.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 105 d­ isease prevalence and test performance.. K. Fenwick. & Weinstein. 263–275. Medical tems. Sculpher. & Weinstein.. and to Owens. E. with clinical outcome assessment and interpreta- Sox. (2012). J. ­symptoms. and continuation of Gold. Decision Making. Glasziou. nonmedical introduction to decision analy- sis. (1970). can help clinicians to structure vides a comprehensive introduction to the use of influ- ence diagrams. & Kuntz. including probabilistic sensitivity analysis. Weinstein. This paper is an introduction to a series of death. 104(1). D. is 0. (2012).. A. Siebert.. Weeks. M. with probabilities of 0. J. R. R. Jr. Cost effectiveness in health and medi. B. J. U. (1996). Computer University Press. Pliskin. (1997). R. A. C.. Most health care professionals would Medical decision making. C. Science.. Siegel. (1997). M.02 and 0. H. Chapter 4 discusses probability that the patient will con- approaches for valuing health outcomes. 17(13). C. utility theory. Schacter. C. 241–262.. 796–803. J. 185. Decision mak- ing in health and medicine. 722– Questions for Discussion 732. F.. (2013). New York: Oxford University Press.. M.. Annals of Internal Medicine.. there are now com.

would you (b) Use the TPR and TNR calculated in change the TPR or TNR of the part (a) to fill in the 2 × 2 table in test? Table 3. If the patient chooses sur- from a diagnostic test for HIV in an gery. (c) The patient wants to know the (b) The man tells you that two people probability that he or she will have with whom he shared needles sub- a stroke given that he or she has a sequently died of AIDS. Which heart attack as a complication of heuristic will be useful in making the surgery. The results of a hypothetical study to the patient having HIV after a measure test performance of a diagnos. Assume that 1 in 500 a subjective adjustment to the pre- patients has both complications.9. tancy = 2 years). Owens and H. your community is 1 in 500 and that he or she has a 5 % chance of death the prevalence in people who have (life expectancy = 0). You have a patient with cancer who prevalence.K.ketabpezeshki. he or she has a 2 % chance of asymptomatic man whose test was dying from the operation (life expec- positive when he volunteered to donate tancy = 0).9  A 2 × 2 contingency table for the hypo. Calculate the disease 4. a 50 % chance of being blood. (021) 66485438 66485457 www. Table 3. You are asked to interpret the results motherapy. You know that the cured (life expectancy  = 1 year). positive test result. Calculate your . specificity. and calculate your answer. Create a decision tree. positive and negative test. culate the post-test probability of 2.10. that you worked out in 2(a) to cal- pendent. and PV–. After taking his history. test probability of disease given a disease prevalence. x quantities that the question ask students to calculate plications are conditionally indepen- dent. If overall prevalence of HIV infection in the patient chooses chemotherapy. has a choice between surgery or che- 3.10. learn that he has a history of intrave.C. Assume that the com. table in Table 3.900 x Total 50 55 105 PCR PCR polymerase chain reaction Total x x x PCR polymerase chain reaction complications.05.106 D. you cured (life expectancy  = 15 years). PV+. the community at large. a 65 % chance injected drugs is 20 times as high as in of cure (life expectancy = 15 years). and PV–. (d) If you wanted to increase the post- (a) Calculate the sensitivity. and a 48 % chance of not being nous-drug use.10  A 2 × 2 contingency table to complete thetical study in problem 2 for problem 2b PCR test Gold standard Gold standard Gold Gold result test positive test negative Total PCR test standard test standard test Positive 48 8 56 result positive negative Total PCR Positive x x x Negative 2 47 49 PCR PCR Negative 100 99. and a 30 % chance that the cancer will (a) Estimate the pretest probability be slowed but not cured (life expec- that this man is infected with HIV. PV+. and that the events are inde. Assume tic test for HIV are shown in the 2 × 2 that the pretest probability is 0. Sox Table 3. test probability in part (a)? that the probability of heart attack (c) Use the sensitivity and specificity is 0.

what is the most important question to con. 10.4. p [ D] × p [ R | D] See the January 1998 issue of Medical p [D | R ] = p [ D] × p [ R | D] + p [ − D] × p [ R | − D] Decision Making for articles that dis- cuss this question.D ] = p [ D ] × p [ R | D ] .D ] + p [ R − D ] post-test probability to compensate for each. and the visual analog scale.D ] p [D | R ] = (3.75. given a would require antibiotic therapy (as test result. Why do you think clinicians sometimes p [ R . we obtain: can influence measurement of test per- formance? Explain what each one is. D. 3.D ] + p [ R . (021) 66485438 66485457 www. When you search the medical litera. we Do you think the reasons that you iden- obtain Bayes’ theorem: tify are valid? Are they valid in only cer- tain situations? Explain your answers.1) ity of bacterial infection as 0. A test p [R ] is available (TPR = 0. (3.3  Biomedical Decision Making: Probabilistic Clinical Reasoning 107 Appendix: Derivation Calculate the expected value of each of Bayes’ Theorem option in terms of life expectancy. p [ R. the time trade-off.4) order tests even if the results will not affect their management of the patient? Substituting Eqs.1. p[D|R]. The definition of conditional no treatment is available). Again. The prior (pretest) proba- opposed to a viral infection. −D] = p [ −D ] × p [ R | −D ]. −D ].8. TNR = 0. −D] 8. and based on the examination you estimate the probabil. p [ R.2. the task of performing a complex deci- sion analysis? p [ R. These expressions can be rearranged: sider? How should you adjust proba- bilities in light of the answer to this p [ R. 3. a significant risk to the patient? 6. Your treat. sum of its probability in diseased patients and its form the test? Explain your reasoning. How could a computer system ease probability.D ] p [ R . from the definition of conditional 7.3) question? . You are concerned that a patient with a Bayes’ theorem is derived as follows. Should you per. probability in nondiseased patients: How would your analysis change if the test were extremely costly or involved p [ R ] = p [ R. for which bility of D is p[D].4 into Eq. (3.85) that indicates the presence or absence The probability of a test result (p[R]) is the of bacterial infection. What are the three kinds of bias that Substituting into Equation 3. R. Explain the differences in three approaches to assessing patients’ ­preferences for health states: the stan- dard gamble.3 and 3. probability is: ment threshold is 0.ketabpezeshki. p [ R |D ] = and p [ R | − D ] = ture to find probabilities for patients p [ D] p [ −D] similar to one you are treating.2) and state how you would adjust the p [ R. We denote sore throat has a bacterial infection that the conditional probability of disease.D ] p [D | R ] = (3. 5.

R. Kaufman. Scottsdale.).J. DSc (*) Department of Center for Cognitive Studies in cal information systems embody ideals in design Medicine and Public Health. interac- methods that are useful for healthcare infor. tem is slow to understand information technology tion? What role do these considerations play and effectively incorporate it into the work envi- in system design? ronment (Shortliffe and Blois 2001). © Springer-Verlag London 2014 (021) 66485438 66485457 www. Shortliffe.L. USA infiltrate clinical practice and settings. enhanc- • What are the basic HCI and cognitive science ing and extending user experiences.1007/978-1-4471-4474-8_4. Patel and David R. Arizona State University. The rapid pace of technological develop- science differs from behavioral science? ments such as the Internet. in implementation. vimla. PhD organization. you should know the 4. development technologies and more generally. As computer-based systems New York. wireless technologies. • What are the attributes of system usability? Yet many observers note that the healthcare sys- • What are the gulfs of execution and evalua. 109 DOI J. These transfor- safety? mations necessitate adaptation at many different levels of aggregation from the individual to the larger institution. Patel.ketabpezeshki.1 Introduction answers to these questions: • How can cognitive science theory meaning.H. The New York Academy that often do not readily yield practical solutions of Medicine. in computing and assessment of healthcare information over the course of the past two decades have systems? begun to permeate diverse facets of clinical prac- • What are some of the ways in which cognitive tice.patel@uth. NY 10029. and cognitive . Enormous advances in health information fully inform and shape design. AZ 85259 effects resulting in systemic inefficiencies and e-mail: kaufman@dbmi. Biomedical Informatics. the conse- e-mail: vpatel@nyam. tions and communications (Rogers 2004). Kaufman After reading this chapter. sometimes causing disruptions of workflow and user dissatisfaction. eracy among healthcare professionals afford the ence between experts and novices? potential for great improvement in healthcare. in the last decade affords characterize the structure of knowledge? significant opportunities for supporting. Cognitive Science and Biomedical Informatics 4 Vimla L. This impact can have deleterious Biomedical Informatics. PhD. • What are some of the ways in which we can and hand-held devices. biomedi- V. Cimino (eds. 1216 Fifth Avenue. Innovative • Why is it important to consider cognition and technologies often produce profound quences often can be felt through all levels of the D. These mation system evaluation and design? advances coupled with a growing computer lit- • What are some of the dimensions of differ.columbia. human factors in dealing with issues of patient suboptimal practice. 13212 East Shea Blvd. Similar to other complex domains. which can lead to frustrated E..

It also has How can cognitive science theory meaning- a close affiliation with educational research. performance. clinicians may revolt and the hospital dotes and case studies is inadequate for produc- may be forced to pull the plug on an expensive ing robust generalizations or sound design and new technology. How can we manage change? implementation principles. and problem-solving (Gillan and Schvaneveldt 1999). patients and health consumers. attention.110 V. In this respect. 4. unnecessary delays in knowledge about system design and implementa- healthcare delivery. encoding and retrieving information from memory. and systems that extend Blois 2001). neuroscience. and the ways in which systems impact ing curve yielding incremental improvements in individual and team performance. an understanding of necessitates an individual and collective learn.L. linguistics. reasoning.1. ity. and methods from the cognitive sciences can illumi- philosophy. applied cognitive research is closely Biomedical Informatics aligned with the disciplines of human-computer interaction (HCI) and human factors.1. Cognitive artifacts are human. mastery of the system and intuitive user interfaces.ketabpezeshki. There is a growing role for the social sciences. The primary disciplines dational role of cognitive science in biomedical include cognitive psychology. the process of medical judgment The past couple of decades have produced and decision-making. Kaufman healthcare practitioners. performance and satisfaction. Biomedical informatics to be more intuitive and also implemented to be is more than the thin intersection of biomedi- coherent with everyday practice? cine and computing (Patel and Kaufman 1998). In tical knowledge embodies the need for sensible the best-case scenario. people’s abilities in perceiving objects. principles of system usability and learnabil- intensive cognitive activity to the external world. we focus on the foun- role in intelligent agency.1 Introducing Cognitive Science in biomedical . However.2 Cognitive Science and In this regard. development everyday life. workflow. Applied grams and devising measures to reduce errors or cognitive research is focally concerned with the increase efficiency. In this chapter. devices. In the worst-case experiential knowledge in the form of anec- scenario. They can also play a larger role in characteriz- mance in technology-mediated settings. In fully inform and shape design. cognitive sci- development and evaluation of useful and usable ence represents one of the component basic sci- cognitive artifacts. nate different facets of design and implementation cognitive science can provide a framework for the of information and knowledge-based systems. we interact with cognitive artifacts to and assessment of health-care information sys- receive and/or manipulate information so as to tems? Cognitive science provides insight into alter our thinking processes and offload effort. anthropology. early language These tasks may include developing training pro- acquisition) as well as applied research. Weinger and Slagle 2001). There is a need for a How can we introduce systems that are designed theoretical foundation. and (021) 66485438 66485457 www. This prac- (Lin et al. Patel and D. including the cognitive and behavioral sciences. From the perspective of informatics. artificial intelli. Cognitive ing and enhancing human performance on a wide science incorporates basic science research focus.g.R. and even adverse events tion that can guide future initiatives. 1998. 4. ences of biomedical informatics (Shortliffe and made materials. the training of healthcare a cumulative body of experiential and practical professionals. healthy consumers of biomedical information. informatics research and practice. memory.. particularly as they pertain to human-computer interaction and other Cognitive science is a multidisciplinary domain of areas such as information retrieval and decision inquiry devoted to the study of cognition and its support. the mediating role of technology in clinical thereby reducing mental workload. analysis and modeling of complex human perfor. Theories and gence. range of tasks involving clinicians. patients and ing on fundamental aspects of cognition (e.

theories of human memory actions.ketabpezeshki. In this chapter. various studies of medical cog- dences between basic cognitive science research. even when scientific knowl.. (5) provide rich descriptive that can then be contrasted with representation accounts of clinicians employing technologies in (021) 66485438 66485457 www.g. nition pertaining to doctor-patient interaction).4 Cognitive Science and Biomedical Informatics 111 Table . This in turn has influ- Precisely how will cognitive science theory enced applied cognitive research on information and methods make such a significant contribution retrieval (Chap. standing biomedical texts. This understanding pro- edge is highly relevant in principle. Heuristics/ Medical problem solving Medical artificial intelligence/decision reasoning strategies and decision making support systems/medical errors Perception/attention Radiologic and Medical imaging systems dermatologic diagnosis Text comprehension Learning from medical texts Information retrieval/digital libraries/ health literacy Conversational analysis Medical discourse Medical natural language processing Distributed cognition Collaborative practice Computer-based provider order and research in health care entry systems Coordination of theory Diagnostic and therapeutic Evidence-based clinical guidelines and evidence reasoning Diagrammatic reasoning Perceptual processing Biomedical information visualization of patient data displays the design of a safer workplace. Furthermore. medical cognition and applied cognitive research in medical informatics Cognitive Science Medical Cognition Biomedical Informatics Knowledge organization Organization of clinical Development and use of medical and human memory and basic science knowledge knowledge bases Problem solving. including: (1) seed basic research findings that can mance (e. ment is that it can inform our understanding of research in text comprehension has provided human performance in technology-rich health.1. significant challenge. (3) medical cognition and cognitive research in present an analytic framework for identifying biomedical informatics along several dimen. problems and modeling certain kinds of user inter- sions. therapeutic reasoning. of such knowledge in clinical systems. (4) characterize the relationship between and knowledge organization lend themselves to health information technology. and (b) research in the areas and memory. 21) from biomedical knowl- towards these important objectives? The transla. information (e. aspects of the visual system). there are correspon. Similarly. (2) pro- of medical errors and patient safety as they vide an explanatory vocabulary for characterizing interact with health information technology). we demonstrate that cognitive (a) basic cognitive science research and theories research. tion of research findings from one discipline into Similarly. making that vides a basis for developing biomedical artificial knowledge effective in a design context can be a intelligence and decision support systems.1 Correspondences between cognitive science.. findings pertaining to the structure illuminate dimensions of design (e. edge sources and research on health literacy. attention of human memory). a theoretical framework for research in under- care environments (Carayon 2007).g.. we discuss In this chapter. how individuals process and communicate health As illustrated in Table 4.g. theories and methods can contribute to that provide a foundation for understanding the applications in informatics in a number of ways underlying mechanisms guiding human perfor. human factors characterizations of expert clinical knowledge and patient safety. theories of problem solving and rea- practical concerns that can be applied to another soning can be used to understand the processes is rarely a straight-forward process (Rogers and knowledge associated with diagnostic and 2004). The central argu. For example.

but recognition of a fundamental set of erally attempted to characterize the changing common concerns shared by the disciplines of psy. the experimental analysis of observable behavior. and applied psychology as well as the social sci- usability testing and human factors) found an ences for the better part of the twentieth century increase of almost 70 % in the last 5 years over (Bechtel et al. Empiricism is the view that experience is the in informatics. Behaviorism represented an the previous 5 years. complex environment in which they exist. epistemology. science of behavior and more specifically. P 33) (H. It also tional definitions so they could be manipulated. interven. A. as the case may be. and (6) furnish a generative are concerned with systems that are adaptive-that are what they are from being ground between the approach for novel designs and productive applied nether millstone of their physiology or hardware. a behavior that was information processing systems. and the social sciences generally. and all of them followed by a satisfying state of affairs is more (021) 66485438 66485457 www. linguistics. Skinner proposed that behavioral events may be understood and ana- lyzed entirely in relation to past and present envi- 4. there has been a significant growth in cognitive research In this section. Similarly. entiate it with competing theoretical frameworks ing informatics journals.g. serves to introduce basic concepts in the study of measured and quantified for empirical investiga- cognition. and could partially emerged as a response to the limitations not be subjected to empirical validation. Kaufman the context of work. Behaviorism prehensive framework of scientific inquiry around constitutes a framework for analyzing and modi. A keyword 1985). We conducted an emergence of cognitive science in view to differ- informal comparison of studies across three . of behaviorism.F. Radical behav- iorism as espoused by B. learn- gestive of a strong growth of cognitive research ing. All behavioral sciences that emphasizes the points of constructs had to be specified in terms of opera- difference between the two approaches.ketabpezeshki. It is an approach that has had an Behaviorists eschewed the study of thinking as an enormous influence on the social sciences for unacceptable psychological method because it most of the twentieth Century. Patel and D. For example.g. fying behavior. research programs in informatics (e. 1975). The next section of the chapter hypothetical constructs (e. Although this doesn’t con. A. computer science. not really a new uli and the responses emitted. These studies gen- discipline.L. of course. Simon 1980) Since the last edition of this text. only source of knowledge (Hilgard and Bower The social sciences are constituted by multi. Journal of Medical Informatics of two time peri. the first being 2001– underlying a particular science of behavior (Zuriff 2005 and the second 2006–2010. Journal of the American Medical introduce core concepts that constitute an explan- Informatics Association and the International atory framework for cognitive science. empirically based stitute a rigorous systematic analysis. mental processes as contains a brief history of the cognitive and mechanisms in a theory) were discouraged. under different reinforcement and punishment All of these disciplines are concerned with contingencies. Cognitive science was inherently subjective. Behaviorism is the conceptual framework ods over the last decade.R. tion (Weinger and Slagle 2001). This framework dominated experimental search of ten common terms (e. Herbert tions in health information seeking). Simon (1980.. we sketch a brief history of the in biomedical informatics. relationship between stimulus and response chology.2 Cognitive Science: The ronment and evolutionary history without any Emergence of an reference to internal states (Baum 2011)..112 V. 1998). Journal of Biomedical in the social sciences.g. error prone. attempt to develop an objective. The section also serves to Informatics. Behaviorism endeavored to build a com- ple frameworks and approaches. it is sug. economics. and the upper millstone of a tion strategies for supporting low literacy popula. Explanatory Framework Behavioral theories of learning emphasized the correspondence between environmental stim- Cognitive science is.. cognition.

size. and mathematical word importantly the advent of the digital computer. the first (021) 66485438 66485457 www. knowledge is lication in the nascent field of cognitive science is nothing more than sum of an individual’s learn. and a set of operators.” It also (A. Noam Chomsky’s writings on syntactic and the modeling of subjects' cognitive and overt grammars in linguistics (see Chap. These tasks They cite Bruner. cybernetics. and their behavior in the form of a computer simulation. study of high-level cognition. behaviorist tenets tion for the formal investigation of symbolic- continue to play a central role in public health information processing (more specifically. The “computer metaphor” provided a The Tower of Hanoi (TOH) is a relatively framework for the study of human cognition as straight-forward task that consists of three pegs the manipulation of “symbolic structures. methods that have become ubiquitous in the ological foundation in a wide range of social sci. A problem space has an initial objectivity and a sufficiency test of a theory and state. behaviorism cognition.1.g. elementary deductive logic. in particular. state to a successor state. own logic theorist program in computer science The Tower of Hanoi. framework. A to peg C one at a time with the constraint that a cessing theory (Atkinson and Shiffrin 1968). This was the culmina- play no part in the learning process. infor. It laid the founda- ence disciplines. classical behavioral theories have been was a mature thesis that described a theoretical largely discredited as a comprehensive unify. the disavowal of the focused primarily on investigations of experi- unobserved such as mental states) of behavior. 1983). chology. a goal state. tion of more than 15 years of work on problem For reasons that go beyond the scope of this solving and research in artificial intelligence. These tasks aroused substantial interest in “information pro. The goal is to move the three disks from peg which was a prerequisite for an information pro. present in the problem statement.4 Cognitive Science and Biomedical Informatics 113 likely to increase the frequency of the act. the Tower of Hanoi.ketabpezeshki. The development of models of places an emphasis on antecedent variables and human information processing also provided a environmental contingencies that serve to sustain foundation for the discipline of human-computer unhealthy behaviors such as smoking (Sussman interaction and the first formal methods of analy- 2001). 4. lyzing problem solving behavior. the most significant landmark pub- According to behavior theories. and introduced protocol-analytic continues to provide a theoretical and method. were well structured. Goodnow and Austin’s “Study allowed for a complete description of the task of Thinking. Arguably. and perhaps most illustrated in Fig. In addition. there was an increasing sis (Card et . The early investigations of problem solving ological constraints (e. The larger disk can never rest on a smaller one. dissatisfaction with the limitations and method. a step-by-step description of the publication “The magic number seven” in psy. implementation of models of human performance Problem solving can be construed as search in a as computer programs provided a measure of problem space. B. In particular. mation theory. Newell and Simon’s “Human Problem Solving” ing history and transformations of mental states (Newell and Simon 1972). required very little background knowledge and cessing (Gardner 1985). 8).” George Miller’s influential journal environment. and C) and three or more disks that vary in provided the foundation for a model of memory. research. However. in the sense that all the vari- Newell and Simon (1972) date the beginning ables necessary for solving the problem were of the “cognitive revolution” to the year 1956. also serves to increase the objectivity of the study Operators are any moves that transform a given of mental processes (Estes 1975). It chapter. sequential behavior of the subjects’ performance. Around 1950. mentally contrived or toy-world tasks such as ism.. served as an as the pivotal works. For example. extended a language for the study of ing theory of behavior. problems (Greeno and Simon 1988). health behavior research problem solving). For example. Cognitive scientists placed important test bed for the development of an “thought” and “mental processes” at the center of explanatory vocabulary and framework for ana- their explanatory framework. developments in logic.

g. there are a total of 27 protocols (Greeno and Simon 1988).1 Tower of Hanoi task illustrating a start state and a goal state move could be to move the small disk to peg B or techniques for representing verbal think-aloud peg C. a book. Patel and D. retrospec- strategy that will minimize the number of steps. The minimum number of moves studies. Comprehension refers to cog- 1972). memory and the verbal reports are inevitably dis- lenges they confront at each stage of a problem torted. data used in studies of problem solving. 4. tive think-aloud protocols are viewed as The metaphor of movement through a problem somewhat suspect because the subject has had space provides a means for understanding how the opportunity to reconstruct the information in an individual can sequentially address the chal. mov.g.. It involves the processes The term protocol refers to that which is produced by a that people use when trying to make sense of a subject during testing (e. In these ber of disks. be described at multiple levels of realization from The most common method of data analysis is surface codes (e. words and syntax) to deeper known as protocol analysis1 (Newell and Simon level of semantics.R. or other 1 informational resources. bal utterance. It also involves the final product of (021) 66485438 66485457 www. more global level in terms of strategies. tasks performed by either clinicians or patients. Think aloud protocols recorded in concert and the actions that ensue. means ends analysis is a commonly investigated the processes and properties of lan- used strategy for reducing the difference between guage and memory in adults and children for many the start state and goal state.ketabpezeshki. a verbal record). Think possible states representing the complete prob.114 V. and sentences (as in a sentence completion task) Although TOH bears little resemblance to the (Anderson 1983). van Dijk and Kintsch (1983) developed an influen- the example illustrates the process of analyzing tial method of analyzing the process of text com- task demands and task performance in human prehension based on the realization that text can subjects. We can characterize with observable behavioral data such as a sub- the problem-solving behavior of the subject at a ject's actions provide a rich source of evidence to local level in terms of state transitions or at a characterize cognitive processes. Protocol analysis refers to a class of nitive processes associated with understanding or deriving meaning from text. subjects are instructed to verbalize their necessary to solve a TOH is 2n−1. For instance.. Ericsson and Simon (1993) specify the at a time. In a three-disk TOH. conversation. TOH has 3n states where n is the . For Cognitive psychologists and linguists have example. aloud protocols are the most common source of lem space. decades. Beginning in the early 1970s.L. conditions under which verbal reports are accept- The search process involves finding a solution able as legitimate data. or a ver- or set of procedures governing behavior or a situation. It differs from the more common use of protocol as defining a code piece of text. Kaufman Fig. Early research focused on basic labora- ing all but the largest disk from peg A to peg B is tory studies of list learning or processing of words an interim goal associated with such a strategy. For example. such as a sentence. Problem solvers thoughts as they perform a particular experimen- will typically maintain only a small set of states tal task.

and artifacts. A (Elstein et al. 4. In the mid to late Fortunately. the clinic. cultures. For example. and the abil. ences about the patient’s underlying condition edge-intensive domains provided a theoretical and may direct the physician’s information- and methodological foundation to conduct both gathering strategies and contribute to an evolv- basic and applied research in real-world settings ing problem representation. Processing opment of the basic methods and principles of problem solving. but is also dependent most compelling proposals has to do with a shift on perceptual processes that focus attention. Problem-solving Cognitive science serves as a basic science research was studying performance in domains and provides a framework for the analysis and such as physics (1980). but were adequately consider cognitive and physiological not fully adequate in accounting for cognition constraints and imposes an unnecessary burden in knowledge-rich domains of greater complex. essentially what people have processing approach has come under criticism understood. there was a shift in research to complex vide us with greater insight into designing sys- “real-life” knowledge-based domains of enquiry tems for the human condition. It is well known that product design often fails to atory vocabulary (e. the mental representa. This paralleled a similar change in terized as a series of operations or computations artificial intelligence research from “toy pro. Furthermore.ketabpezeshki. In recent years. such as the workplace (Vicente 1999) and the Two interdependent dimensions by which classroom (Bruer 1993). many of the problems distributed cognition in greater detail in a subse- associated with decision making are the result of quent section. However. The shift to real-world problems world. One of the ing and decision making. They also provide a rich explan.g. on task performance (Preece et al. computational theory of mind provides the fun- Similarly the study of text comprehension shifted damental underpinning for most contemporary from research on simple stories to technical and theories of cognitive science. Mental represen- grams” to addressing “real-world” problems and tations are internal cognitive states that have the development of expert systems (Clancey and a certain correspondence with the external Shortliffe 1984). processes of the solitary individual. These are likely to elicit further infer- ratory experiments. interaction. In uted across groups. we can characterize cognitive systems are: (1) tion provided a fertile test bed for assessing and architectural theories that endeavor to provide extending the cognitive science framework. This fact. problem space). The basic premise scientific texts in a range of domains including is that much of human cognition can be charac- medicine.. We explore the concepts underlying hension. either lack of knowledge or failure to understand the information appropriately. the conventional information- tion of the text. for its narrow focus on the rational/cognitive Comprehension often precedes problem solv. Most of the after noticing an abnormal gait as he entered early investigations on expertise involved labo. 1978) and architecture (Akin 1982).3 Human Information constrained artificial environment for the devel. they may reflect a clini- in cognitive science was spearheaded by research cian’s hypothesis about a patient’s condition exploring the nature of expertise. advances in theory and methods pro- 1970s. a unified theory for all aspects of cognition and (021) 66485438 66485457 www. some of the more important differences in claim has significant implications for the study medical problem solving and decision making of collaborative endeavors and human-computer arise from differences in knowledge and compre. The early investigations provided a . the from viewing cognition as a property of the soli- availability of relevant knowledge. the shift to knowl.4 Cognitive Science and Biomedical Informatics 115 such processes. ity and involving uncertainty. These areas of applica. tary individual to viewing cognition as distrib- ity to deploy knowledge in a given context. 2007). on mental representations. which is. medical diagnoses modeling of complex human performance. (Greeno and Simon 1988).

or problem solving vide a single set of mechanisms for all cognitive strategies. The goal of such a theory is to pro- speed.. These can be either structural regulari. vides a means to put together a voluminous and tive performance (e. vision and motor skills change denced in human behavior. extended discussion of architectural theories and (021) 66485438 66485457 www. between perceptual.3.ketabpezeshki.L. 2009). seemingly disparate body of human experimental and their development during the lifespan. such as processing of cognition. limitations on working memory). there have systems and memory capacity limitations— been several attempts to develop a unified theory or processing regularities. The architectural approach capital. Cognitive architecture regards to the lifespan issue. It represents a scientific hypothesis about those digm has contributed to our understanding of the aspects of human cognition that are relatively nature of medical expertise and skilled clinical constant over time and independent of task performance. Cognitive architectures also play a role in providing blueprints for building future intelligent systems that embody a broad range 4. Although there is much plasticity evi- tion. nity. This are bound by biological and physical constraints. memory. can instrumentally contribute to such an endeavor ties—such as the existence of and the relations (Zhang et .116 V. Kaufman (2) distinction the different kinds of knowl. programs. problem solving and com- enable (e. and knowledge. functionally in terms of the capabilities they to decision making. An made significant inroads in the design commu. A graphical user interface or more capacities and limitations of the human cogni- generally. 2004). many of whom suffer (e. These from chronic health conditions such as arthritis constraints reflect the information-processing and diabetes. selective attention. the expert-novice para. cognitive processes as a function of aging (Fisk et al.g. focused attention on selective visual prehension (Newell 1990). and memory Over the course of the last 25 years. relatively fixed permanent structure that is (more Differences in knowledge organization are or less) characteristic of all humans and doesn’t a central focus of research into the nature of substantially vary over an individual’s lifetime. are invariably challenges in applying basic izes on the fact that we can characterize certain research and theory to applications. 2008). Cognitive systems are characterized behaviors from motor skills. there is a growing represents unifying theories of cognition that are body of literature on cognitive aging and how embodied in large-scale computer simulation aspects of the cognitive system such as atten. and long-term memories that store content about and psychomotor skills over the course of the last an individual’s beliefs. expertise. (Carroll 2003). There capabilities. In medicine.R. Patel and D. 2009). A more regularities of the human information processing human-centered design and cognitive research system. basic science research is of growing importance Cognitive architectures specify functional rather to informatics as we seek to develop e-health than biological constraints on human behavior applications for seniors. language. limitations on memory). and endowed tions about the human cognitive system. Such a theory pro- features). cognition. goals. there remains a significant gap in applying edge necessary to attain competency in a given basic cognitive research (Gillan and Schvaneveldt domain.1 Cognitive Architectures and of capabilities similar to those of humans (Duch Human Memory Systems et al.g.g. Architectural systems embody a may not be suitable for older adults. the way they constrain human cogni. Designers routinely violate basic assump- of their knowledge. the 50 years has provided a foundation for design representation of elements that are contained in principles in human factors and human-computer these memories as well as their organization into interaction. Individuals differ substantially in terms 1999). a website designed for younger adults tive system. In data into a coherent form. Cognitive architectures include short-term Fundamental research in perception. attentional. experiences.. Although cognitive guidelines have larger-scale structures (Langley et al. memory.

for example. For example.g. High velocity/high work. humans actively construct and interpret infor- Working memory is an emergent property of mation from their environment. etc. Studies have also shown the indi- Knowledge viduals at different levels of expertise will differ- entially represent a text (Patel and Kaufman Architectural theories specify the structure and 1998). to an excess of information that competes for few Propositions are a form of natural language cognitive resources. Problems impose a varying time from our initial experience of the phenom- cognitive load on working memory. On the other content. we employ the architectural frame of independent constructs that account for the reference to introduce some basic distinctions in variability of mental representations needed to memory systems. However. The power of cognition is reflected in the the environment (e. written text. objects. reference to linguistic content. whereas working memory the cognitive system needs to be attuned to differ- (WM) refers to the resources needed to maintain ent representational types to capture the essence information active during cognitive activity (e. we recover the propositional structure when we load clinical environments such as intensive care read or listen to verbal information.. reconstruct. and transform events. whereas WM is limited to other than that in which they have occurred with- five to ten “chunks” of information. Representations enable us to remem- become familiar from repeated exposure and is . non-experts will often remember more the kinds of knowledge that reside in LTM and information. For idea (i.4 Cognitive Science and Biomedical Informatics 117 systems is beyond the scope of this represent percep- knowledge activated from long-term memory. subsequently stored in memory as a single unit images. LTM is infinite. are expressed as language and translated into ment. experts are more likely to mechanisms of memory systems. semantics) or concept without explicit example. Similarly. and conversations absent in space and (Larkin et al. Representations reflect states of knowledge. music. of these inputs. 1980).e.g. Given that envi- interaction with the environment. experiences and thoughts in some medium theory.. These ideas cognitive load. maintaining a seven-digit phone num. that people recover the gist of a text or spoken communication (i. representation that captures the essence of an ing memory (Chandler and Sweller 1991).2 The Organization of Kintsch 1983). creating a burden on work. but much of the recalled informa- that support decisions and actions. psychology has furnished a range of domain- However. text differently than we do mathematical equa- tained in working memory includes stimuli from tions. ber in WM is not very difficult. Long-term ronmental stimuli can take a multitude of forms memory (LTM) can be thought of as a repository (e. cally be interpreted as a greeting containing iden- versation is nearly impossible for most people. whereas theo. a crowded computer display speech or text when we talk or write. divided into at least two structures: long-term A central tenet of cognitive science is that memory and short-term/working memory. images. selectively encode relevant propositional infor- ries of knowledge organization focus on the mation that will inform a decision. A chunk is out extraneous or irrelevant information (Norman any stimulus or patterns of stimuli that has 1993). words on a display) and ability to form abstractions .. tical propositional content even though the literal Multi-tasking is one factor that contributes to semantics of the phrases may differ. “hey”. There are several ways to characterize hand.e.3. is another contributor. The information main. van Dijk and 4. propositional structure) not the specific words (Anderson 1985. of all knowledge.). The structure of the task environ.. For example. we process written text comprehension). For example.g. Numerous units also impose cognitive loads on clinicians psychological experiments have demonstrated carrying out task. speech. Human memory is typically engage the external world. Cognitive tion may not be relevant to the decision (Patel (021) 66485438 66485457 www. to “hello”. and “what’s happening” can typi- maintain a phone number while engaging in con. This refers ena.. In tions.

g. is divided into the past) and the textbase representation. (Patel et al. clinical practice (e. This situation model is typ- tions constitute an important construct in theories ically derived from the general knowledge and of comprehension and are discussed later in this specific knowledge acquired through medical chapter. the sentation plus the domain-specific and everyday schema serves as a “filter” for distinguishing rel- knowledge that the reader uses to derive a broader evant and irrelevant information. The predicate calculus rep. text. teaching. events.g. and Groen 1991a. 1986). b). geometric Kintsch (1998) theorized that comprehension shapes.1. therapeutic plan or prognosis (Patel dial infarction may contain the findings of “chest (021) 66485438 66485457 www. using a predicate calculus formalism or as a knowledge of associations between clinical find- semantic network. b). birds fly) that repre- problem as written in a patient chart. Schemata have constants (all birds have is called the textbase (the propositional content of wings) and variables (chairs can have between the text).g. the focus is on a female who Since the knowledge in LTM differs among phy- has the attributes of being 43 years of age and sicians. based on tures for representing categories of concepts their summary or . in medicine the textbase one and four legs).ketabpezeshki. fruits. Schemata can meaning from the text.. tion model is constituted by the textbase repre. A subject’s tions or treatment procedures that have worked in response. a schema for myocar- diagnosis. The formalism have been widely used in the study of medical is informed by an elaborate propositional lan.L. other forms of knowledge representation.. sent the prototypical circumstance. Theories and methods of text comprehension cede the event of 1. When a person interprets information. the situation be considered as generic knowledge structures model would enable a physician to draw that contain slots for particular kinds of inferences from a patient’s history leading to a propositions. The method provides us with Schemata represent higher-level knowledge a detailed way to characterize the information structures.2 (diarrhea). and thyroid conditions). 4. For instance. Kaufman Fig.2 Propositional analysis of a think-aloud protocol of a primary care physician and Groen 1991a. In medicine.2. the resulting situation model generated white. knowledge of medica- resentation is illustrated below. The text information mation. indicates that the events of 1. The variables may have asso- could consist of the representation of a patient ciated default values (e. theories and findings Propositional knowledge can be expressed from biomedical research). conveys and the schemata in long-term memory. the sit- The formalism includes a head element of a seg.g.R. cognition and have been instrumental in charac- guage (Frederiksen 1975) and was first applied to terizing the process of guideline development the medical domain by Patel and her colleagues and interpretation (Arocha et al. Patel and D. proposition 1.118 V.3 (GI upset) pre. Like sentences or segments and sequentially analyzed.g. For instance. perception of the patient). There are sche- involves an interaction between what the text mata for concepts underlying situations. To process Comprehension occurs when the reader uses information with the use of a schema is to deter- prior knowledge to process the incoming infor. chairs. sequences of actions and so forth. 4.... 2005). Propositional representa. as given on Fig. readings (e. The TEM:ORD or temporal order relation by any two physicians is likely to differ as well. mine which model best fits the incoming infor- mation presented in the text. ings and specific diseases. For example in information (e. The situa. They can be construed as data struc- subjects understood from reading a text. uation model is used to “fit in” the incoming ment and a series of arguments. stored in memory (e.

The chological evidence to suggest that mental research also documented various conceptual images constitute a distinct form of mental repre. like other forms of mental knowledge is a kind of knowing related to how representation. The claim is that clinicians and medical nal models of systems. to perform various activities. of the system as a function of expertise. Mental images are a form of cians’ mental models of the cardiovascular sys- internal representation that captures perceptual tem (specifically. Conceptual knowledge is acquired simulation to generate possible future states of a through mindful engagement with materials in system from observed or hypothetical state. language or Conceptual knowledge and procedural knowl- from one’s imagination (Payne 2003). To engage in such a predict the effects of network interruptions on conversation.” “sweating. tion of a physical system. els can be derived from perception. Figure 4. images of the objects and their spatial arrange.” which is part of the dictive and explanatory capabilities of the func- schema for thyroid disease. The study information recovered from the environment. Patel and Magder (1996) characterized clini- ment in the room. Kaufman. sitate the acquisition of procedural knowledge. niture in your living room. as is the preserve literal information about the external case in reasoning about the behavior of electrical world. Mental models are students have variably robust representations of designed to answer questions such as “how does the structure and function of the system. models are a particularly useful construct in However.ketabpezeshki.. There are numerous plete. cardiac output).” “shortness of breath. when one initiates a Google Search. flow in the pulmonary and cardiovascular sys- struct for describing how individuals form inter. trates the four chambers of the heart and blood Mental models are an analogue-based con. one needs to be able to construct getting cash from an ATM machine). effects of perturbations in the system on blood resentation explicitly shares the structure of the flow and on various clinical measures such as left world it represents (e.3 illus- diagnosis such as dermatology and radiology. a transition from a declarative or interpretive stage (021) 66485438 66485457 . or summary repre. tems.. mental models are always incom. Imagine that you are having a conversa. a set of connected visual ventricular ejection fraction. impacted subjects’ predictions and explanations ticularly important role in domains of visual of physiological manifestations. This is in contrast to an vide another useful way of distinguishing the abstraction-based form such as propositions or functions of different forms of representation.g. the development of skills may involve understanding human-computer interaction. Procedural sentation. This it work?” or “what will happen if I take the fol. schemas in which the mental structure consists of Conceptual knowledge refers to one’s under- either the gist. Running edge are acquired through different learning of a model corresponds to a process of mental mechanisms.” but not An individual’s mental models provide pre- the finding of “goiter. Procedural knowl- one may reasonably anticipate that system will edge is developed as a function of deliberate return a list of relevant (and less than relevant) practice that results in a learning process known websites that correspond to the query. imperfect and subject to the processing technical skills in medical contexts that neces- limitations of the cognitive system. Mental as knowledge compilation (Anderson 1983). model enables prediction and explanation of the lowing action?” “Analogy” suggests that the rep.g. images of a partial road map from your home to Conceptual and procedural knowledge pro- your work destination). However. circuits (White and Frederiksen 1990). standing of domain-specific concepts. an abstraction. The tion at the office about how to rearrange the fur.4 Cognitive Science and Biomedical Informatics 119 pain. flaws in subjects’ models and how these flaws sentation (Bartolomeo 2008) Images play a par. characterized the development of understanding There is compelling psychological and neuropsy. For a range of contexts (from reading texts to con- example. Mental mod. construct has been used to characterize models tions reflect abstractions and don’t necessarily that have a spatial and temporal context. versing with colleagues). model can be used to simulate a process (e. More often the The schematic and propositional representa.

which is a condition-action We have employed an epistemological frame- rule that states “if the conditions are satisfied. Patel using empirical evidence from psychological et al. mies.3 Schematic model of circulatory and cardiovascu- lar physiology. (1997) address this issue in the context of record patient data (Kushniruk et al. work developed by Evans and Gadd (1989).com . in learning to use an electronic health considerable importance for research in bio- record (EHR) system designed to be used as part of medical informatics. In view to understand the nature Fig. For even clinical) knowledge. it is necessary to charac- of the pulmonary and systemic circulation system and the terize the domain-specific nature of knowledge process of blood flow. a production rule.g.R. Clearly. risk factors for heart disease) without any in-depth understand- ing. Much research has been a consultation. The diagram illustrates various structures of medical cognition. 1996. The acquisition of factual knowledge alone is not likely to lead to any increase in understanding or behavioral change (Bransford et al. The acquisition of conceptual knowledge involves the integration of new information with prior knowledge and necessitates a deeper level of understanding. with the aim of developing biomedical ontolo- a more experienced user of this system can more gies for use in knowledge-based systems.120 V. 2000b). Facts are routinely disseminated through a range of sources such as pamphlets and websites. Production rules are a common method for representing knowl- edge in medical expert systems such as MYCIN (Davis et al. Factual knowledge involves merely knowing a fact or set of facts (e.ketabpezeshki. Patel and D. They (021) 66485438 66485457 www.L. In addition to differentiating between proce- dural and conceptual knowledge. but is often validity of the AI systems. (1999). This is contrast to a new medical student who may have largely factual knowledge. this explaining and predicting physiologic behavior can be a rather daunting task. nomenclatures and vocabulary systems Procedural knowledge is often modeled in such as UMLS or SNOMED (see Chap. whereas. Biomedical taxono- used without conscious awareness. 7) are cognitive science and in artificial intelligence as engaged in a similar pursuit. 1977). Patel effortlessly interview a patient and simultaneously et al. a less experienced user will need to conducted in biomedical artificial intelligence attend carefully to every action and input. 4. For example. Thus far. The illustration is used to exemplify organization in medicine. one can differ- entiate factual knowledge from conceptual knowledge. Procedural knowledge supports experiments on medical expertise to test the more efficient and automated action.. risk factors may be associated in the physician’s mind with biochem- ical mechanisms and typical patient manifesta- tions. there is no single way to represent all biomedical (or toward increasingly proceduralized stages. we have only considered domain- general ways of characterizing the organization of knowledge. but it is an issue of example. Kaufman then execute the specified action” (either an inference or overt behavior). Given the vastness the concept of mental model and how it could be applied to and complexity of the domain of medicine.

used by researchers in medical artificial intelli.4 Medical Cognition vide us with insight into the organization of clini- cal practitioners’ knowledge (see Fig. Finally. such as the ethnic background of the levels at which biomedical knowledge may be a patient. The study of expertise is one of the principal par- The framework consists of a hierarchical adigms in problem-solving research. Establishing A goal of this approach has been to character- a finding reflects a decision made by a physician ize expert performance in terms of the knowledge that an array of data contains a significant cue or and cognitive processes used in comprehension. (Lesgold et al. and decision making. using consist of clusters of findings that indicate an carefully developed laboratory tasks (Chi and underlying medical problem or class of prob. while novice chess players levels of abstraction. Diagnosis is the level of could only reproduce approximately 20 % of the classification that subsumes and explains all lev. The grandmaster chess players were able to manageable sub-problems and to suggest possi.4). Facets problem solving. Facets resemble the kinds of constructs novice differences. nity to explore the aspects of performance that facets. 1988). followed by findings. . In one of his experiments. 4. permits investigators to develop domain-specific However. the systems level consists the knowledge used for medical understanding of information that serves to contextualize a par- and problem solving. Findings are composed of observations that assessment and training purposes. cues that need to be taken into account. reconstruct the mid-game positions with better ble solutions. Glaser 1981). Glaser 2000). correct positions. have potential clinical significance. experts to novices provides us with the opportu- tions at the lowest level. This framework represents a formal- ization of biomedical knowledge as realized in textbooks and journals. (1965) pioneering research in chess represents tions such as left-ventricular failure or thyroid one of the earliest characterizations of expert- condition. 4. they do not constitute clinically useful models of competence that can be used for facts. solving skill (Lesgold 1984. deGroot’s lems. and also for differentiating ticular problem. They reflect general pathological descrip. and can be used to pro. Facets also vary in terms of their than 90 % accuracy. Comparing structure of concepts formed by clinical observa. Clinical observations are undergo change and result in increased problem- units of information that are recognized as poten.4 Epistemological frameworks representing the structure of medical knowl- edge for problem solving proposed a framework that serves to characterize els beneath it. They are interim hypotheses that serve to duce the position of the chess pieces from mem- divide the information in the problem into sets of ory.4 Cognitive Science and Biomedical Informatics 121 Fig. subjects were allowed to view a chess board for gence to describe the partitioning of a problem 5–10 seconds and were then required to repro- space. When the chess pieces were (021) 66485438 66485457 www.ketabpezeshki. It also tially relevant in the problem-solving context. and diagnoses. 4.

Although competent performers. experts physicians who jects varying in levels of expertise in terms of solve a case outside their field of specialization) memory.. music study of medical expertise. In the 1980). The concept of an lem solution to the givens of the problem state. which novices cannot from detailed experience within a medical sub- perceive. Chi et al.. This phenomenon is accounted for it. may be able to encode relevant for. and long-term memories for materials (e.. The expert worked forward from the givens passes competency in a domain (Sternberg and to solve the necessary equations and determine Horvath 1999). (021) 66485438 66485457 www. refers to an individual who sur- ment. 1995). or errors. a distinc- an informative general overview of the area.g.L. The results indicated certified by a professional licensing body. 1988 Patel and Groen (1991a.e. the particular quantities they are asked to solve for instance. Hoffman 1992) provide between general and specific expertise. while novices tend to spend more time working by a process known as “chunking. Differences in the directionality of reasoning information and generate effective plans of action by levels of expertise has been demonstrated in in a specific domain.ketabpezeshki. however. (3) they have superior short-term expertise. uation (Hoffman et al. determine the strategies a subject uses to solve a Usually. tuned to the particular problems at hand. An deployment of different skills required for individual may possess both. someone is designated as an expert problem.g. but is a result of an enhanced superficial level of representation. as exem- novice subject with an expert subject in solving plified by Elo ratings in chess. superior memory ences impact the problem representation and for domain materials).R. it has been useful to (Sloboda 1991).122 V. and medicine (Patel et al.. of perceiving large patterns of meaningful infor. The novice solved most of the problems simply based on years of experience or peer eval- by working backward from the unknown prob.. but not outside of it. they often lack the speed diverse domains from computer programming and the flexibility that we see in an expert. stantially in terms of exhibiting these kinds of per- It is well known that knowledge-based differ. expert. and in particular and experts (i.g. General expertise corresponds to expertise research are the following: (1) experts are capable that cuts across medical subdisciplines (e. or engineering. possesses an extensive. This tion supported by research indicating differences research has focused on differences between sub. between subexperts (i. Edited volumes (Ericsson 2006. (2) they are fast at processing and at domain. superior recognition ability is not a function of more principled levels whereas novices show a superior memory. Kaufman placed on the board in a random configuration. such as graduate degrees. clini- not encountered in the course of a normal chess cal findings in medicine) within their domain of match. distinguish different types of expertise. law. Among reasoning strategies and organization of knowl- the expert’s characteristics uncovered by this edge. (Perkins et al.e. formance characteristics (e. Specific expertise results mation in their domain. general medicine).g. Simon and Simon (1978) compared a based on a certain level of performance. such as cardiology or endocrinology. 1990) to medical diagnoses (Patel A domain expert (e. a medical practitioner) and Groen 1986). b) distinguished Ericsson and Smith 1991. sports (Allard and Starkes 1991). domain specialist) in terms of the role of domain specific knowledge. Patel and D.” It is the most on the solution itself and little time in problem general representational construct that makes the assessment.. by virtue of being textbook physics problems. or only generic problem solving. as in that the expert solved the problems in one quarter . 1994). on the basis of of the time required by the novice with fewer academic criteria. (4) they typically fell to that of novices. expert chess masters’ recognition ability expertise. (5) they spend ability to recognize typical situations (Chase and more time assessing the problem prior to solving Simon 1973). reasoning strategies. This result suggests that represent problems in their domain at deeper. accessible knowledge The expertise paradigm spans the range of base that is organized for use in practice and is content domains including physics (Larkin et al. (6) individual experts may differ sub- fewest assumptions about cognitive processing.

in recall studies. than intermediates and novices. The tasks used include comprehension and explanation of clinical prob- lems. It is important to note. The intermediate effect is also partly due to the tuning of skills. a clinical levels of expertise may perform more poorly than case). b). the development is monotonic. generation of Fig. re-learning. In other words. students and residents. ity) gradually increases. including senior medical made. As compared to a novice student (Fig. Furthermore. and requests for laboratory data. interaction.. illustration of this learning and developmental the reasoning pattern of an intermediate student phenomenon known as the intermediate effect. increasing accumulation of knowledge and fine. Patel et al. research has intermediates. as a relevancy of the stimuli to a problem is taken into person becomes more familiar with a domain. suggesting that Furthermore. with an intermediate somewhat unusual trajectory. The performance indicators used have included recall and infer- ence of medical-text information. For instance. a reorganization of knowledge and tively error-free performance. an appreciable monotonic phenomenon or her level of performance (e. clinical findings from a patient in doctor-patient The straight line gives a commonly assumed representa- tion of performance development by level of expertise. his account. skills takes place. learning process involves phases of error-filled During the periods when the intermediate performance followed by periods of stable. when knowledge and fine-tuning of skills. The Y-axis may represent any of a .g. punctuated by peri. 1988. 4. experts recall more relevant propositions those at lower level of expertise on some tasks. accuracy. ety of tasks however. The research has also identified novice to expert.4 Cognitive Science and Biomedical Informatics 123 The intermediate effect has been found in a variety of tasks and with a great number of performance indicators. Rather. in diagnostic accuracy. effect occurs. novices. For instance. Cross.5 presents an is not. qual. 1994). when propositions are divided in terms novices have shown that people at intermediate of their relevance to the problem (e. developmental levels at which the intermediate ber of performance variables such as the number of errors phenomenon occurs. The curved line represents the actual development from among others.g. 4. that in some tasks. intermediate levels typically generate a great ods of apparent decrease in mastery and declines deal of irrelevant information and seem incapa- in performance. generation of diagnostic hypotheses.5 Schematic representation of intermediate effect. intermediates.6). number of conceptual elaborations. number of concepts recalled. That is. However. Figure 4. which may be necessary for ble of discriminating what is relevant from what learning to take place. it requires the arduous unintended changes that take place as the person process of continually learning. or number of hypotheses generated in a vari. shows the generation of long chains of discussion (021) 66485438 66485457 www. recall and explanation of laboratory data.ketabpezeshki. People at exercising new knowledge. there is a longstanding body of intermediates have difficulty separating what is research on learning that has suggested that the relevant from what is not. appears.. It is often assumed exhibiting higher degree of accuracy than the that the path from novice to expert goes through novice and the expert demonstrating a still higher a steady process of gradual accumulation of degree than the intermediate. rela. The development of expertise can follow a there is a gradual increase. showing the typical non-monotonic effect. and problem solving (Patel and Groen 1991a. and experts are assessed in terms shown that this assumption is often incorrect of the total number of propositions recalled (Lesgold et al. recall and inference of diagnostic hypotheses. characterized by shifts in per- human learning does not consist of the gradually spectives or a realignment or creation of goals. and However. sectional studies of experts. and reorganizes for intended changes. doctor-patient communication.

4.6 Problem interpretations by a novice medical stu. knowledge structure of a senior level student There are situations. intermediate effect disappears. in which the leads him more directly to a solution (Fig. shown from right to left (solid light line).R. informa. This suggests that under time-restricted (021) 66485438 66485457 www. Thick solid dark tion in the box represents diagnostic hypothesis. maybe as a pears when short text-reading times are used.ketabpezeshki. 4. intermediates. Patel and D. 4.L.7).8).. Schmidt reported Thus. 4. and experts given only tors involved in the intermediate effect may help a short time to read a clinical case (about thirty in improving performance during learning seconds) recalled the case with increasing accu- (e. Forward driven or data driven inference dent. The given information from patient problem is rep. the intermediate effect can be explained as that the intermediate recall phenomenon disap- a function of the learning process. however. Identifying the . Kaufman Fig. by designing decision-support systems or racy.g. A well-structured focusing on relevant information). resented on the right side of the figure and the new Backward or hypothesis driven inference arrows are generated information is given on the left side. line represents rule out strategy Intermediate hypothesis are represented as solid dark Fig.124 V.7 Problem interpreta- tions by an intermediate medical student evaluating multiple hypotheses and reasoning in intelligent tutoring systems that help the user in haphazard direction (Fig. arrows are shown from left to right (solid dark line). Novices. circles (filled). necessary phase of learning.

In other words. This model of is a decrement in performance until a new level problem solving has had a substantial influence of mastery is . both on studies of medical cognition and medical education. consider more likely to engage in unnecessary search. This latter stage is most amenable to the case. intermediates cannot engage in extra.ketabpezeshki. the greater ability of expert physicians to selec- nized to be efficiently used. which proposed that physicians rea- Ledley and Lusted (1959) into the nature of clini. Although intermediates may ing students with medical experts in simulated have most of the pieces of knowledge in place. 4.e.8 Problem interpreta- tions by a senior medical student conditions. it occurs repeatedly at stra. The results emphasized this knowledge is not sufficiently well orga. First. They were the first to use experimen- tal methods and theories of cognitive science to 4. cal thinking is associated with the seminal work tegic points in a student or physician’s training of Elstein. generated a small set of hypotheses very early in tion stage. intermediates that est empirical studies of medical expertise can are not under time pressure process too much be traced to the works of Rimoldi (1961) and irrelevant information whereas experts do not. physicians eration stage followed by a hypothesis evalua.e. rea- clinical reasoning involving a hypothesis gen. Their research findings led to the development The systematic investigation of medical expertise of an elaborated model of hypothetico-deductive began more than 50 years ago with research by reasoning. These physicians by drawing on then contemporary periods are followed by intervals in which there methods and theories of cognition.. The intermediate effect is not a one-time The origin of contemporary research on medi- phenomenon.. soning from hypothesis to data). novices lack the knowledge to tal studies of diagnostic reasoning by contrast- do much searching. and Sprafka (1978) who and follow periods in which large bodies of new studied the problem solving processes of knowledge or complex skills are acquired. as soon as the first pieces of data became (021) 66485438 66485457 www. Until this knowledge tively attend to relevant information and narrow becomes further organized. fewer hypotheses). Rather. Shulman.4. formal decision analytic techniques.1 Expertise in Medicine investigate clinical competency.4 Cognitive Science and Biomedical Informatics 125 Fig. The earli- neous search. soned by first generating and then testing a set of cal inquiry. the intermediate is the set of diagnostic possibilities (i. They proposed a two-stage model of hypotheses to account for clinical data (i. Kleinmuntz (1968) who conducted experimen- On the other hand. problem-solving tasks.

which depends on the physician pos- Feltovich. Moller. sessing a highly organized knowledge base about drawing on models of knowledge representation the patient’s disease (including sets of signs and from medical artificial intelligence. Kaufman available. Shulman.126 V. Although data-driven reasoning is highly nations of a complex clinical problem (Patel efficient. reasoning. Medicine is clearly a knowledge-rich ices and experts. characterized symptoms). The disease models were differently suggests that they might reach differ- described as sparse and lacking cross-referencing ent conclusions (e. The results of this soning. can result in a complete chain of inferences from (021) 66485438 66485457 www. each cue Patel and Groen (1991a. Pure data-driven rea- reasoning strategy—using patient data to lead soning is only successful in constrained toward a complete diagnosis (i. deductive model of reasoning as espoused by tion. and Swanson (1984). driven reasoning strategy. intermediates use more hypothesis-driven rea- novice’s knowledge was described as “classically. decisions or understand- between shared features of disease categories in ings) when solving medical problems.ketabpezeshki. evant data. They a range of contexts of varying complexity. physicians made use of a They tended to use a backward or hypothesis- hypothetico-deductive method of diagnostic rea. since there are no subjects who accurately diagnosed the prob. reasoning from situations. experts’ memory store of dis. focusing only on the rel. Experts tend to use data-driven domain and a different approach was needed. Elstein. nature and directionality of clinical reasoning in contributory to each hypothesis generated. clinical problem solving. reasoning during clinical performance. 1980). built around the prototypical instances patterns. this type Patel and colleagues studied the knowledge. they make. nosed or partially diagnosed the patient . The fact that experts and novices reason of a disease category. and Sprafka (1978). Similar memory. employed a forward-oriented (data-driven) ences that a person makes. soning resulting often in very complex reasoning centered”. Patel and D. ing. where one’s knowledge of a problem data to hypothesis). hypothesis generation. domains (Larkin et al. built-in checks on the legitimacy of the infer- lem. of reasoning sometimes breaks down and the based solution strategies of expert cardiologists expert has to resort to hypothesis-driven reason- as evidenced by their pathophysiological expla. Second. cue interpretation. novices and the domain of pediatric cardiology. In contrast. According to the authors. patterns of reasoning have been found in other ease models was found to be extensively cross-ref. ences about diagnostic cues and evaluation of Although experts typically use data-driven competing hypotheses. It has been established The previous research was largely modeled after that the patterns of data-driven and hypothesis- early problem-solving studies in knowledge-lean driven reasoning are used differentially by nov- tasks.g. The results indicated that adequate domain knowledge. b) investigated the was interpreted as positive.R. which did and hypothesis evaluation. For example. The were unable to find differences their diagnostic objectives of this research program were both to reasoning strategies between superior physicians advance our understanding of medical expertise (as judged by their peers) and other physicians and to devise more effective ways of teaching (Elstein et al. negative or non. hypotheses. Johnson. 1978). it is often error prone in the absence of and Groen 1986).. experts typically skip steps in their These differences accounted for subjects’ infer. Due to their exten- erenced with a rich network of connections among sive knowledge base and the high level inferences diseases that can present with similar symptoms.L.. physicians were selective in This is in contrast to subjects who misdiag- the data they collected. Attention to initial not differentiate expert from non-expert reason- cues led to the rapid generation of a few select ing strategies. The hypothetico-deductive process was study presented a challenge to the hypothetico- viewed as consisting of four stages: cue acquisi. Because of their lack of substantive fine-grained differences in knowledge organization knowledge or their inability to distinguish rele- between subjects of different levels of expertise in vant from irrelevant knowledge. reasoning. Third.

However. for instance. as illustrated in Fig. It is. For instance. 4. 4. in some circumstances. such as in explanations. the situation produces absence of relevant prior knowledge and when a heavy load on cognitive resources and may there is uncertainty about problem solution. therefore. These lem. and trying to fit the loose Visual diagnosis has also been an active area of ends within it. as the person searches for an explana. and learning the content of the material. the physician to a conditional relation. unfamiliarity with the prob. or consulting a domain expert. in a hypothesis-driven reasoning inquiry in medical cognition. sis. other characteristics associated conditions include the presence of “loose ends” with expert performance were observed. It has Studies have shown that the pattern of data. COND refers cal findings on the left side of figure). their ability to diagnose skin lesions presented on tion for them. The presence of loose ends may foster tigated clinicians at varying levels of expertise in learning. Hypothesis-driven tive load.9 Diagrammatic representation of data-driven (right of figure). and uncertainty (Patel et al. a medical student or a slide. In diminish students’ ability to focus on the task. when students are given reasoning is usually exemplary of a weak method problems to solve while training in the use of of problem solving in the sense that is used in the problem solving strategies.4 Cognitive Science and Biomedical Informatics 127 Fig. The results revealed a monotonic increase (021) 66485438 66485457 www.ketabpezeshki. been found that when subjects used a strategy driven reasoning breaks down in conditions of based on the use of data-driven reasoning. the use of to be used when domain knowledge is inadequate data-driven reasoning may lead to a heavy cogni- or the problem is complex.g. attention. in a backward-directed fashion. they case complexity. strong methods engage The reason is that students have to share cogni- knowledge whereas weak methods refer to gen. Loose ends trigger reasoning strategy. ratory failure. where some particular piece of a reduced number of moves to the solution. goals and hypotheses. reading a specialized because one has to keep track of such things as medical book. Weak does not necessarily imply learning to solve the problem-solving method ineffectual in this context. ilar cases seen in the past. For instance. their problem solving perfor- explanatory processes that work by hypothesizing mance suffered (Patel et al. which is inconsistent with the main diagno- From the presence of vitiligo. tive resources (e. a prior history of progres. problem-solving terms. CAU indicates a causal relation.. In addition. were more able to acquire a schema for the prob- lem. a physician may encounter a sign or a symptom tion. the anomalous finding of respi- (top down) and hypothesis-driven (bottom-up) reasoning. In contrast. in a patient problem and look for information that hypothesis-driven reasoning is slower and may may account for the finding by searching for sim- make heavy demands on working . information remains unaccounted for and isolated However. reasons forward to conclude the diagnosis of Myxedema and RSLT identifies a resultive relation the initial problem statement to the problem solu. 1990). 1990). the patient. memory) between eral strategies. Studies have inves- fashion. a disease. and examination of the thyroid (clini. when subjects used a hypothesis-driven from the overall explanation. most likely However. is accounted for as a result of a hypometabolic state of sive thyroid disease.

are rife with 1981). Stewart. Koppel accuracy of diagnosis. (1988) describes a litany of poorly designed which resulted in a gross anatomical localization artifacts ranging from programmable VCRs and served to constrain the possible interpreta. The results . innovative and promising medical information tions despite discrepant findings in the patient technologies that have yielded decidedly sub- history.128 V. that around 50 % of software code was devoted tation. Nielsen (1993) reported detection. There have been numerous and edited volumes (e.g. problematic interfaces can gests systematic differences between subjects at have serious consequences for patient safety varying levels of expertise corresponding to (Lin et al. feature Twenty years ago. ers indicated that. Novices experienced greater difficulty are inherently non-intuitive and very difficult focusing in on the important structures and were to use. both in medical problem The history of computing and more generally. therapeutic reasoning and 1998) and the importance of enhancing the user mental models of physiological systems. Evans and Patel 1989). that reflected the underlying pathophysiological structure. optimal results and deep user dissatisfaction Crowley.ketabpezeshki. on average. 1994) years (Jaspers 2009). tifaceted discipline devoted to the study and ples of a problem is considered to be one of the practice of design and usability (Carroll 2003). At minimum. employed a similar protocol-analytic approach to difficult interfaces result in steep learning the Lesgold study to examine differences in curves and structural inefficiencies in task per- expertise in breast pathology. 2004. 6 % of their ment of effective search strategies. Zhang et al. to answering machines and water faucets that tions. Medical experience has been widely acknowledged by cognition remains an active area of research and both consumers and producers of information continues to inform debates regarding medical technology (see Chap. Naus. In a classi. 11).L. texts devoted to promoting effective user inter- Other active areas of research include medical face design (Preece et al. At worst. surface features (e. solving and in other domains (Chi and Glaser the history of artifacts design. formance. greatly increased over the course of the last 10 marized in a series of articles (Patel et al. curricula and approaches to learning (Patel et al. “scaly lesions”). Kaufman in accuracy as a function of expertise. Shneiderman text comprehension. In able to rapidly invoke the appropriate schema the Psychology of Everyday Things. project budgets were spent on usability evalua- rate recognition of anatomic location. fication task. in the interpretation of chest x-ray pic. and all aspects of task per. On the other hand. 1998. fast and accu. Norman and initially detect a general pattern of disease. Lesgold et al. novices categorized lesions by their 2005. unusable by anyone except for the team of tures. formance including microscopic search. it is likely that more feature identification strategies that results from a than 50 % of code is now devoted to the GUI. well-organized knowledge base. 2007. 11). Part of the impetus is (021) 66485438 66485457 www. Similarly. et al. Given the ubiquitous presence of graphi- of visual data interpretation skills and explicit cal user interfaces (GUI). and Friedman (2003) when implemented in practice. The authors propose a model of visual to the user interface and a survey of develop- diagnostic competence that involves develop. hallmarks of expertise. usability evaluations have The study of medical cognition has been sum. (1988) investigated the stories of dazzlingly powerful devices with abilities of radiologists at different levels of remarkable capabilities that are thoroughly expertise.5 Human Computer according to superordinate categories. interme- diates grouped the slides according to diagnosis and expert dermatologists organized the slides 4.R. Schmidt and Rikers 2007). feature identification and data interpre. such as Interaction viral infections. Patel and D. acquisition tion..g. The results revealed that the experts were designers and their immediate families. 2005) (see Chap. Human computer interaction (HCI) is a mul- The ability to abstract the underlying princi. there have been numerous more likely to maintain inappropriate interpreta.

ketabpezeshki. an experienced user can attain a high level of pro. consider new interface requirements. design yield significant contributions. development and applied social by the system should be easy to retain once science research (Carroll 2003). and research web readily usefully deployed to improve human environments for clinicians (Elkin et al. 2005). The interface is based are now “touchstones in the culture of comput. focus groups. (2) efficiency: devoted to the study and practice of usability. tion retrieval and useful summary reports). Physicians in benefits. usability techniques are increasingly in a subsequent section. usability evaluation of medical infor. In pulmonary graph displays (Wachter et al. to reduce medical errors. communication and is not optimally designed 2001). physician to support patient care (e. efficient informa- order entry (Ash et al.g. Chan usability and interface design will be a perpetu- and Kaufman 2011). there also active academic HCI communities that have remain formidable challenges to designing and contributed significant advances to the science of developing usable systems. it is safety. There are Even with growth of usability research.g. In performance. 2008).. ally moving target. took too much time and was difficult to very clear that investments in usability still yield use (Benko 2003). Human computer lowing five attributes: (1) learnability: system interaction (HCI) is a multifaceted discipline should be relatively easy to learn. The methods include obser. it receives between $10 and $100 a few months after implementation. and (5) satisfaction: the user experience integration and evaluation of applications of should be subjectively satisfying. prominence.g.. However. 2000.. seniors and low case for the instrumental value of such research literacy patient populations). and administrative staff). substantial rates of return (Nielsen et al. the capabilities of this system are not information retrieval systems. 2003a. In rather comprehensive and supports a wide range our view. Koppel et al. nature of user interactions.g. these studies make a compelling across the digital divide (e. Although tively seamless integration with current workflow evaluation methodologies and guidelines for and practices. there is a What do we mean by usability? Nielsen need for a scientific framework to understand the (1993) suggests that usability includes the fol. as health infor- vations. of functions and user populations (e. compromised patient the flourishing of the World Wide Web). The interface emphasizes Usability methods have been used to evaluate completeness of data entry for administrative a wide range of medical information technolo. To provide another example. Although much has changed in the complained that the system. technology to support human activities. nurses. user. on a form (or template) metaphor and is neither ing” (Carroll 2003). user acceptance and rela. 2002). there is a need to to improve efficiency.4 Cognitive Science and Biomedical Informatics 129 that usability has been demonstrated to be highly by the events at Cedar Sinai Medical Center in cost . The concept of usability as well as clinicians find it exceptionally difficult and the methods and tools to measure and promote it time-consuming to use. Karat (1994) reported that for which a decision was made to suspend use of a every dollar a company invests in the usability computer-based physician order entry system just of a product. (4) errors: system should be designed to HCI has spawned a professional orientation minimize errors and support error detection and that focuses on practical matters concerning the recovery. mation technology reaches out to populations Collectively. which was designed world of computing since Karat’s estimate (e. surveys and experiments. In addition. used to assess patient-centered environments Innovations in technology guarantee that (Cimino et al. purposes rather than the facilitation of clinical gies including infusion pumps (Dansky et al. learned. 2003. Kaufman et al. ventilator management systems. This is exemplified computing.. We further discuss issues of EHRs addition. physi- mation technologies has grown substantially in cians.nor task-centered. 2003). HCI has emerged as a central area of both com- ductivity. HCI researchers have been devoted (021) 66485438 66485457 www. general. (3) memorability: features supported puter science. we have been working with a mental health It remains far more costly to fix a problem after computer-based patient record system that is product release than in an early design phase.

a behavioral approach to understanding and furthering software design. The user may or may decision making. In this context. Carroll (1997) traces the history of HCI back to the 1970s with the advent of software psy- chology. which in as an exciting development. (1983) envisioned HCI as a test bed for apply. ing an action. behaviors.R. not perceive a change in system state (e. and Selection Rules cognitive system. Operators. and analyti.10.. perception. multimodal interfaces. methods and tools to assess retrieve information. Most medical information technologies such as illustrated in Fig. Our own research is grounded action sequence. Models of cognitive engineering are typi- growth of this initiative. and user populations. In this regard. Although we view this involves the formation of an intention. it is concerned with developing analytic frameworks for characterizing how tech- nologies can be used more productively across a range of tasks. Card et al. tributed to a certain scientific fragmentation The intention leads to the specification of an (Carroll 2003). 4. design concepts. 4. which may include logging onto in a cognitive engineering framework. Human factors. ergonomics. The specification results in execut- In supporting performance. this case might be to retrieve the record online. The approach is centrally con. . This pattern is embodied in analysis of routine skills and expert performance. let’s presup- approaches with an abundance of new theoreti. medical record. The (021) 66485438 66485457 www. The system responds in some fashion memory. rapid prototyping).L. collaborative work- spaces. GOMS is a powerful cally predicated on a cyclical pattern of interac- predictive tool.g. engaging a search facility to ment of principles. it has also con. 2002). ubiquitous computing. The the processing constraints imposed by the human Goals.. problem solving. The second stage cal foci (Rogers 2004). Norman’s (1986) seven stage model of action. record and an electronic record. and entering the patient’s and guide the design of computerized systems to medical record number or some other identifying support human performance (Roth et al. Patel and D. Fig. sys- cerned with the analysis of cognitive tasks and tem provides no indicators of a wait state). Kaufman to the development of innovative design concepts such as virtual reality. social.g. (GOMS) approach to modeling was a direct out. and industrial engineering research were pursuing some of the same goals along parallel tracks. the focus is on cog.130 V. for example. HCI research has also been focally concerned with the cognitive. retrieving a patient’s cognitive skills.ketabpezeshki. and cultural dimensions of the computing experience. several actions). settings. which the system (which in itself may necessitate is an interdisciplinary approach to the develop. The action cycle begins provider-order entry systems. engage complex with a goal. and immersive environments. HCI research has been instrumental in transforming the software engineering process towards a more user-centered iterative system development (e. which may necessitate several nitive functions such as attention. comprehension. In the early 1980s.10 Norman’s seven stage model of action ing cognitive science research and also furthering theoretical development in cognitive science. information. and (or doesn’t respond at all). but it is limited in scope to the tion with a system. The goal is abstract and indepen- HCI research has embraced a diversity of dent of any system. pose that the clinician has access to both a paper cal frameworks.

changes to the system design and educating users therapeutic reasoning. clinical tem. tion technologies. The gulf of evaluation reflects the degree performing it. formance is predicated on an analysis of both the after the primary selection had been made. like many medical informa- an alternative course of action is necessary. . it is some. Similarly. The designer’s model is the conceptual cating orders in view to admit a patient remains model of intent of the system. described below. transparent. There are two of the actions as reflected in the state of the sys- primary means in which the action cycle can tem (e. This will then determine of greater complexity that incorporates a wide whether the user has been successful or whether range of functions. entry task can be completed using written orders Gulfs are partially attributable to differences or one of many diverse computer-based order- in the designer’s model and the users’ mental entry systems.. desired outcomes. the state of a sys- be achieved. The users’ men. Norman’s theory of action has given rise to (or ing of subgoals. especially as they learn the sys. This information-processing demands of a task and is a source of confusion. For example. and patient monitoring to foster competencies that can be used to make and management. the particular implementa- on an estimation of the user population and task tion will greatly impact the performance of the requirements (Norman 1986). it may introduce gaining an understanding of how actions (e. Goals that necessitate multiple state or nitive demands and have a common underlying screen transitions are more likely to present diffi. many systems require a 1992). This analysis is often referred to to which the user can interpret the state of the as cognitive task analysis. For example. There is cycle may appear to be completely seamless. structure that involves similar kinds of reasoning culties for users. For example.. For example.4 Cognitive Science and Biomedical Informatics 131 perceived system response must then be inter. On the other hand. partially based the same. methods that inform this approach can be applied tions have been met. the process may break actions (e. an admission order- better use of system resources. However. a need to provide good mappings between the However to a novice user. The underlying task of communi- models. The principles and system and determine how well their expecta. clicking on a button) and the results down at any of the seven stages. and patterns of inference. For example. Generic tasks necessitate similar cog- place. screen transitions).ketabpezeshki. the action tem should be plainly visible to the user.g. task. The study of human per- goal completion action. A complex task will involve substantial nest. involving a series of actions in some case reinforced) the need for sound that are necessary before the primary goal can design principles.g. To an experienced user. The gulf of execution reflects the designed system will provide full and continuous difference between the goals and intentions of feedback so that the user can understand whether the user and the kinds of actions enabled by the one’s expectations have been met.g. Graphical user interfaces that These are a class of usability evaluation meth- involve direct manipulation of screen objects ods performed by expert analysts or reviewers represent an attempt to reduce the distance and unlike usability testing. a system may eliminate the tal models of system behavior are developed need for redundant entries and greatly facilitate through interacting with similar systems and the process. especially for novice the kinds of domain-specific knowledge required users. unnecessary complexity leading to suboptimal clicking on a link) will produce predictable and performance. don’t typically (021) 66485438 66485457 www. system. a well- break down. from the analysis of times difficult to interpret a state transition and written guidelines to the investigation of EHR determine whether one has arrived at the right systems. A user may not know the appropriate The model has also informed a range of cog- action sequence or the interface may not provide nitive task-analytic usability evaluation methods the prerequisite features to make such sequences such as the cognitive walkthrough (Polson et al. Bridging gulfs involves both bringing about tasks in medicine include diagnostic reasoning.. to a wide range of tasks. The preted and evaluated to determine whether the distance is more difficult to close in a system goal has been achieved. between a designer’s and users' model. such as pressing “Enter”.

R. The specific aims of the 6. Action: Press Button Next to Checking frustrate real users. The method is performed by System response: Enter Dollar Amounts in an analyst or group of analysts ‘walking through’ Multiples of 20 (Screen 6) the sequence of actions necessary to achieve a 10. Goal: Obtain $80 Cash from Checking Account ciples such as visibility of system status.g. goals and subgoals. Action: Enter $80 on Numeric Key Pad goal. The CW process emphasizes the sequential pro- ple. and six screen transitions. the ways in which a user’s In general. Action: Press Button Next to No Response procedure are to determine whether the typical System response: Select Transaction-8 Choices user’s background knowledge and the cues gen. Subgoal: Decide whether a printed record ing sequences of actions and goals needed to is necessary accomplish a given task. a home telecare system.. the displays. Action: Select Correct mouse clicks and cognitive actions (e.g. user 1.. infer. (021) 66485438 66485457 www.e. multiple screen transitions are more system response (or absence of one). 1996). Both behavioral or physical actions such as 11. Kaufman involve the use of subjects (Nielsen 1994)).com . Subgoal: Choose Between Quick Cash to produce the correct goal-action sequence and Cash Withdrawal required to perform a task. Action: Enter Card (Screen 1) control and freedom. not unlike problem solving. it is desirable to minimize the num- objectives can be translated into specific actions). For exam. involved in involve locating an icon or shortcut on one’s completing a computer-based task. The walkthrough of the ATM reveals that ence needed to carry out a physical action) are process of obtaining money from the ATM coded. Action: Enter “Pin” on Numeric keypad task analytic method that has been applied to the 4. (Screen 4) erated by the interface are likely to be sufficient 7. This is illustrated below in a tion methods. into a series of subgoals and actions. In As in Norman’s model. Action: Press Button Next to Cash intended to identify potential usability problems Withdrawal that may impede the successful completion of a System response: Select Account (Screen 5) task or introduce complexity in a way that may 9. which is likely to confuse the user. The CW is a cognitive 3. The system response (e. action structure (i. ber of actions necessary to complete a task. ton next to screen) tinct medical information technologies System response: “Do you Want a Printed (Kushniruk et al. to characterize the cognitive processes of users Binary Option: Yes or No (Screen 3) performing a task.. We have employed a duly noted.132 V.ketabpezeshki. which can be decomposed infusion pumps used in intensive care settings. each action results in a addition. update of values) is also char- cognitive walkthrough (CW) and the heuristic acterized and an attempt is made to discern evaluation are the most commonly used inspec. The purpose of a CW is Transaction Record”. potential problems. and analysis include goals. the basis of a small set of well-tested design prin. The method involves identify.L. Patel and D. System response: Enter PIN > (Screen 2) flexibility and efficiency of use (Nielsen 1993). The principal assumption underlying this necessitated a minimum of eight actions. Subgoal: Interpret prompt and provide We illustrate HE in the context of a human factors input study later in the chapter. Action: Hit Enter (press lower white but- study of usability and learnability of several dis. five method is that a given task has a specifiable goal. consistency and standards. opening an Excel spreadsheet (goal) may cess. The codes for including EHRs. 5. Heuristic evaluation (HE) is a partial walkthrough of an individual obtaining method by which an application is evaluated on money from an automated teller system. similar approach to analyze the complexity of a The CW method assumes a cyclical pattern of range of medical information technologies interaction as described previously. 2. The focus is desktop (subgoal) and double clicking on the more on the process rather than the content of application (action). The method is 8. The change in screen.

It is clear that understanding. comfortable. engineering not of medicine “(Senders. 2010) with a particular icy issues inform and shape healthcare pro- focus on a) system usability and learnability. resources. In addition. and jobs. The methods are complementary and serve as health consumers and patients (Flin and Patey a means to triangulate significant findings. cerned with the full complement of technologies ods such as the cognitive walkthrough and and systems used by a diverse range of individu- usability testing (Beuscart-Zephir et al. 2007). and environ- problems. (2003) conducted a cognitive study of health practices from several perspec- evaluation of the IDEATel home telemedicine tives or levels of analysis.4 Cognitive Science and Biomedical Informatics 133 Usability testing represents the gold standard 4. limitations. Human factors and human computer (021) 66485438 66485457 www. Human factors of a small number of subjects. It involves in-depth testing technologies (Henriksen 2010). The focus in this (HTU). nature of the interaction (Jaspers 2009). ways in which organizational. It is known to capture a higher percentage which may follow. and involve five to ten subjects who are asked to think effective human use (Chapanis 1996). (2) electronic transmission having to do with patient safety. A full exposition of human factors in b) the core competencies. hospital administrators. safety. A typical usability testing study will ments for productive. context of healthcare. which provided the following functions: chapter is on cognitive work in human factors (1) synchronous video-conferencing with a and healthcare. abilities. 2003. b). Starren et al. and the adverse events tasks. Weinstock et al. The complex setting requires a detailed understand- usability study revealed dimensions of the inter. people work with them or are impacted by these Kaufman et al. significant obstacles cor. skills and knowledge medicine is beyond the scope of this chapter. The For a detailed treatment of these issues. als including clinicians. and health literacy were cepts and discuss illustrative research in patient documented.ketabpezeshki. five or six sub. and five or six subjects. The assumption is research discovers and applies information that a test can be perfectly valid with as few as about human behavior. are problems of psychology and of the more serious usability problems and pro. that patient safety is a systemic challenge at a ings. 1993)” vides a greater depth of understanding into the (cited in (Woods et al. The focus is on the system (Shea et al. and pol- 2002. 2005a. 18 for more details on IDEATel). Human factors work approaches the Kaufman et al. In the aloud as they perform the task. 2002. necessary to productively use the system. tasks. 2009). 2009). jects may find upwards of 80 % of the usability machines. ing of both the setting and the factors that influ- face that impeded optimal access to system ence performance (Woods et al. other characteristics to the design of tools. systems. mental theoretical foundation. It refers to a and Patient Safety class of methods for collecting empirical data of representative users performing representative Human error in medicine. the study employed both a cognitive walkthrough reader is referred to the Handbook of Human and in-depth usability testing. The focal point of Factors and Ergonomics in Health Care and the intervention was the home telemedicine unit Patient Safety (Carayon 2007). safe. establish important con- models of the system.6 Human Factors Research in usability evaluation methods. predicting access to Web-based educational materials (see and transforming human performance in any Chap. and cesses. In addition. 2007). particularly in relation to issues nurse case manager. (4) review of one’s clinical data and (5) vidual. We recognize of fingerstick glucose and blood pressure . human factors is con- It’s not uncommon to employ multiple meth. Usability testing commonly employs video capture of users Human factors research is a discipline devoted performing the tasks as well as video-analytic to the study of technology systems and how methods of analysis (Kaufman et al. Our objective in this section is to introduce a responding to perceptual-motor skills. (3) email to a physician and nurse case multiple levels of aggregation beyond the indi- manager. cultural. 2009).

there is no workload.L. The erable concern for the past quarter century. they were able to determine that an adverse highway. Human factors analysis care errors. of evaluation and both strongly emphasize a user-centered approach to design and a systems- centered approach to the study of technology use. but are not restricted to operations (Woods et al. but objective is to ensure their effectiveness. The categorization of information a consequence. equipment. Any system will be easier and less bur. industrial pro- whereas human factors focus on a broad range cess control. us to selectively prioritize and attend to certain The Harvard Medical Practice Study was pub- stimuli and attenuate other ones. in a patient who is receiving the treatment. HCI is more focused on com. Kaufman interaction are different disciplines with dif.6. However. The research has computing technologies (Carayon 2007).R. Both human fac. Based on an to multitask by dividing our attention between extensive review of patient charts in New York two activities. surprising fact that 98. and the elimination or mitigation of applies knowledge about the strengths and limi. Human able to human error. They also have lished several years prior to the IOM report and the property of being sharable. patient injury caused by healthcare errors (Patel tations of humans to the design of interactive and Zhang 2007). starting point. Mental models and issues of Institute of Medicine . if we are driving on a State. Human factors research leverages theories and ferent histories and different professional and methods from cognitive engineering to character- academic societies. underlying patterns in seemingly disparate tors and HCI employ many of the same methods settings (Woods et al. which enables us was a landmark study at the time. Patient elucidated empirical regularities and provides safety is one of the central issues in human fac. It has been an issue of consid- systems. the greater community was galvanized by the and ease of use.ketabpezeshki. and their environment.1 Patient Safety Researchers and professionals in both domains draw on certain theories including cognitive When human error is viewed as a cause rather than engineering. 1991). and may argue over the specific numbers. 2007). They tainous roads. military command control and space of systems that include. safety. Although one tive capacity. explanatory concepts and models of human per- tors research and we address this issue in greater formance. This report communicated the analysis. ing cause of death in this country. we can easily have a conversation with event occurred in almost 4 % of the cases (Leape a passenger at the same time. challenging situations in aviation. Our perceptual system inundates us disputing that too many patients are harmed or with more stimuli than our cognitive systems can die every year as a result of human actions or possibly process. By serving as an end point rather than a or HCI is sometimes capricious. you will have to allocate more of further determined that almost 70 % of these your attentional resources to driving and less to adverse events were caused by errors and 25 % of the conversation. attention. For example. it retards further understanding. (021) 66485438 66485457 www.000 preventable deaths densome to use to the extent that that it is co. all errors were due to negligence. ize human performance in complex settings and puting and cutting-edge design and technology. Attentional mechanisms enable absence of action. Patel and D. (Henriksen 2008) The field of human factors is guided by prin- ciples of engineering and applied cognitive psy. it serves as a cloak for our igno- technology-based work as either human factors rance. as the et al.” decision making are central to human factors released in 1999. including memory. This enables us to discern common detail in a subsequent section.134 V. 4. 2007). every single year in the United States are attribut- extensive with users’ mental models. An adverse event refers to any unfa- skies get dark or the weather changes or suddenly vorable change in health or side effect that occurs you find yourself driving through winding moun. “To Err is Human. which makes it the 8th lead- factors focus on different dimensions of cogni. Patient safety refers to the prevention of health- chology (Chapanis 1996).

They argue for the error. uncertainties. researchers in other safety-critical tant. Human factors researchers recognized when everything is operating smoothly. gaps in supervision. they may be rule-based. Reason (1990) intro. ence (e.4 Cognitive Science and Biomedical Informatics 135 We can only analyze errors after they hap. Johnson. Error is the failure of the vast majority of errors will be trapped and a planned sequence of mental or physical activi. but it was executed inappropriately. inad. zero defect. resulting in an immediately felt. or pharmacists who are actively respond. error (Morel et al. rect knowledge was available. duced an important distinction between latent In addition. in which case the cor- ditions that were operative prior to the mishap. demands and other latent con. However. conducted empirical investigations into error tant technologies. Reason 1990). the situational context. the patient has been harmed. shifting attention A mistake involves an inappropriate course of demands and competing goals that characterized action reflecting an erroneous judgment or infer- a situation prior to the occurrence of an error. communication breakdown detection and recovery by experts (attending phy- between key actors. tied to ability to detect and recover from errors Zhang. citing it as an impractical goal. stress. The effects of active failure are increase the potential for an error. if there is a system of checks approach that recognizes that error could be and balances that is part of routine practice or if attributed to a multitude of factors as well as the there is systematic supervisory process in place. and Shortliffe (2004) and not so much to the ability not to make errors. Latent conditions are enduring systemic domains have long since abandoned the quest for problems that may not be evident for some time. interaction of these factors. Hindsight bias masks the dilem. and combine with other system problems to weaken choosing to focus instead on the development of the system’s defenses and make errors . suggesting They further characterize medical errors as a some sort of human deficiency or irresponsible progression of events. event. 2005. If these measures or ties to achieve its intended outcome when these practices are not in place. and a search for the guilty culprits. 2011). owing to factors such as incorrect knowledge or mas. Then that this approach error is inherently incomplete an unsafe practice unfolds resulting in a kind of and potentially misleading. based on the distinctions proposed by Reason pened and they often seem to be glaring blunders (1990).. Mistakes may either be knowledge-based hindsight bias. an error can propagate failures cannot be attributed to chance (Arocha and cross the boundary to become an adverse et al. Active failure represents workload or intense time pressure. We can further classify errors in terms of after the fact. conditions are less visible.ketabpezeshki. strategies to enhance the ability to recover from There is a lengthy list of potential latent condi. The latent perception of the performance of most clinicians. and absence of a safety culture critical care domain. committed by providers such as nurses. Patel and her colleagues tions including poor interface design of impor. This leads to assignment of blame slips and mistakes (Reason 1990). A slip occurs or search for a single cause of the error. The notion that human error should not be tol- cians.g. but equally impor. but not necessarily leading to an adverse need for a more comprehensive systems-centered event. if an individual is subject to a heavy and active failures. For example. when the actor selected the appropriate course in hindsight. active errors are adverse event. it is exceedingly difficult to recreate of action. a wrong diagnosis or misreading of an This sort of retrospective analysis is subject to x-ray). but there was Too often the term ‘human error’ connotes blame a problem in applying the rules or guidelines. defused in this middle zone. At this point. However. using both laboratory-based in the workplace—a culture that emphasizes safe and naturalistic approaches (Patel et al. Patel. practices and the reporting of any conditions that These studies show that expertise is more closely are potentially dangerous. In healthcare. have developed a taxonomy of errors partially The study results show that both the experts and (021) 66485438 66485457 www. then that will the face of error. There is a period of time behavior. 2008). sicians) and non-experts (resident trainees) in the equate training. erated is prevalent in both the public and personal ing to patient needs at the “sharp end”. physi.

ketabpezeshki. the efforts of researchers and practi.R. those related to recover from knowledge based errors is better entering and retrieving information and those than that of trainees. decision-making. Holden and Karsh (2007) argue that purpose of prevention. 2003).com . that suggests that medical devices can also cause eted theoretical framework incorporating theories substantial harm (Jha et al. related to communication and coordination. and to self-administer analgesic drugs as required. Kaufman non-experts are prone to commit and recover enumerate the primary kinds of errors caused from errors. apparent that technology-induced errors are a method of pain relief that uses disposable or deeply consequential and have had deleterious electronic infusion devices and allows patients consequences for patient safety. There has been a wealth of studies regarding Medical devices include any healthcare patient safety and medical errors in a range of product. and social-cognition. The end result is a clinicians often do not receive adequate training model that can be applied to health information (Woods et al.g. Although medi- which is a perspective that interweaves technol. tunities to facilitate clinical care and medical decision making.L. have been partially automated and offer a com- tioners will yield greater success in the under. States (Bright and Brown 2007). Ash. resulting in rapid detec. Patel and D. monitoring. tive overload. Lin and col- healthcare in a multitude of ways including the leagues (1998) conducted a series of studies on reduction of errors. and implementation of health et al. reported that more than one million adverse med- vation. 2007). Although this affords oppor- information technology. cian and this limits the maximum level of drug munity. many interfaces are not suitable for tion and correction. It has been from different spheres of research including moti. Studies on expertise and settings that are highly interruptive (e. they are complex in nature and for the design of systems. ical device events occur annually in the United They also draw on a sociotechnical approach. design. diagnosis. excluding drugs. that are used for the contexts. The patient controlled analgesia (PCA) devices. 2007). 4. They also decision-making are important if we are to build characterize a problem in which an informa- robust decision-support systems to manage the tion entry screen that is highly structured and boundaries of risk of error in decision making requires completeness of entry can cause cogni- (Patel and Cohen 2008). They propose a multifac. However. example. it may add layers of complexity and uncertainty. tion in complex real-time critical care situations The authors characterize several problems that appears to induce certain urgency for quick action are not typically found in usability studies. Kuperman (2003b) were among the first to give The device is programmed by a nurse or techni- voice to this problem in the informatics com. but experts’ ability to detect and by health information systems. There is considerable evidence of the lessons learned. cal devices are an integral part of medical care in ogy. of principles for design and implementation. For in a high alert condition. In addition.. people. They also endeavored to describe and administration to keep the dose within safe lev- (021) 66485438 66485457 www. a clut- understanding of the limits and failures of human tered display with too many options). Stavri. and authors propose that through iterative testing of bar coded medication administration systems the model. plex programmable interface (Beuscart-Zephir standing. much of the work is atheoretical in nature and treatment or alleviation of an illness (Ward and that this diminishes the potential generalizability Clarkson 2007). many medical technology usage behavior and that guides a set devices such as such as smart infusion pumps.136 V. it is increasingly a patient controlled analgesic or PCA device. and the social context of interaction hospital settings.2 Unintended Consequences There is evidence to suggest that a poorly designed user interface can present substantial It is widely believed that health information challenges even for the well-trained and highly technologies have the potential to transform skilled user (Zhang et al. 2010). Error detection and correc. 2005a.6.

a simple than the other ones. a single action). They found the programming sequence was unnecessarily that particular usability problems were associ- complex. Kushniruk et al. and cognitive evalu- paring simulations of the old and new interface. logs.. They They found that programming the new interface found that the error was due to a confluence of was 15 % faster.e. its design served as a .g.5. the problem of inap- back and to structure the user experience (e. This is a compelling demonstration administered. and in fact. Lin and colleagues investigated the effects Zhang and colleagues employed a modified heu- of two interfaces to a commonly used PCA ristic evaluation method (see section 4.ketabpezeshki.. On the basis of the cognitive patient safety. They con.. tual operators utilized by clinicians when calcu- losophy that emphasizes simplicity and function. consequential and have the potential to impact essarily complex. the system responses (e. Several of the violations changes in the display as a result of an action) were classified as problems of substantial sever- and inferences that are needed to interpret the ity. documented violations. It is an effective gauge of the pumps was likely to induce more medical errors complexity of a system. It’s important to note that the revised screen actual patient. The poor interface design did not that medical equipment can be made safer and provide assistance with the decision-making pro- more efficient by adopting sound human factors cess. they found the existing PCA inter.4 Cognitive Science and Biomedical Informatics 137 els. On the basis of an analysis by 4 original interface so that it was more in line with evaluators. medication. Their results suggested that one of the two state of the system. ation of the order entry system involved. The authors used a range of inves- was a computer simulation and was not actually tigative methods including inspection of system implemented in the physical device. associated with mistakes (not detected by users) On the basis of the CTA analysis. (2005) exam- task analysis. not duration). of potassium chloride being administered to an back. cognitive task analysis is a method and visibility (the ease in which a user can dis- that breaks a task into sets of subtasks or steps cern the system state) were the most widely (e. For example. Consistency previously. As described interface design were documented. (2005) according to sound human factors principles. ated with the occurrence of error in entering able on the screen to provide meaningful feed. (021) 66485438 66485457 www. ined the relationship between particular kinds face to be problematic in several different ways. the electronic health record. task that necessitates 25 steps or more to com. examination of ducted a cognitive study with 12 nurses com.. In addition. 2003). design principles. above) device including the original interface. There was a lack of information avail. a nurse ing the screen was found to be correlated with would not know that he or she was on the third of errors in entering wrong dosages of medica- five screens or when they were half way through tions. lating medication dosage (i. It is clear that usability problems are plete using a given system is likely to be unnec. the labeling of functions and ambiguous interface led to 10 errors as compared to 20 for dating of the dates in which a medication was the old one. For example.g. Lin and while others were associated with slips pertain- colleagues (1998) also redesigned the interface ing to unintentional errors. based on volume ality over intricacy of design and presentation. where the interface was a poor fit for the concep- This methodology embodies a particular phi. For example. Horsky et al. analyzed a problematic medication order placed The new system was designed to simplify the using a CPOE system that resulted in an overdose entire process and provide more consistent feed. propriate default values automatically populat- negotiating the next steps). they redesigned the et al. The new display. of usability problems and errors in a handheld For example. semi-structured interviews. a total of 192 violations with the user sound human factors principles.g. The average workload rating factors including problems associated with the for the old interface was twice as high. Based on to test the safety of two infusion pumps (Zhang a cognitive task analysis. certain types of errors were the task. the structure of many subtasks in prescription writing application.

Patel and D. When patients undergo surgery. These changes cy’s warehousing and not on clinical guidelines.ketabpezeshki. audible sounds. The duplicative medication orders and thereby increas. The study describes three of communication and cultural transmission. letters and numbers).com . including a need or a routine dose for a particular kind of patient. there was a discrepancy between their study of the causes and mitigation of unintended expectations and the dose listing. it is classical model of information-processing cog- relatively easy to select the wrong patient because nition viewed external representations as mere (021) 66485438 66485457 www. again. medication or patients not receiving the intended tions and a survey to document the range of medication. Physicians must reenter CPOE and reacti- many of them occurred with some frequency. fragmentation of data and failure to integrate the The growing body of research on unintended hospital’s information systems and (2) human. Frequently. although normal dosages are 20 or consequences. On a similar note. means usability problems.. practices that have little experience with health tion errors. resulting in enhanced perfor- 30 mg. a in the external world which has the potential to reminder that prior orders exist and may need to be internalized. In this case. ish their impact (Bloomrosen et al.R. for more cognitively-oriented research to guide However. screen. When selecting a patient.L. patient safety as well as greater user doses.3 External Representations CPOE systems.138 V. the errors. The It is a well-known phenomenon that users matter is especially pressing given the increased come to rely on technology and often treat it as implementation of health information tech- an authoritative source that can be implicitly nologies nationwide including ambulatory care trusted. term external representation refers to any object ing the possibility of medical errors. Once The errors were classified into two broad cate. the pharmacy might stock only 10-mg mance. The study. Kaufman Koppel and colleagues (2005) published an names and drugs are close together. internal representations reflect mental nicians fail to cancel the existing orders leading to states that correspond to the external world. icons. The authors outline displays to determine the minimum effective dose a series of recommendations. and. patients’ names do not appear on all computer-provider order-entry systems (CPOE) screens. According to the authors.g. 2011). used a series of methods result in either unintended patients receiving including interviews with clinicians.6. Clinicians mistakenly assumed that this was the minimal dose. The kinds of problems. The system expects a clinician to and Information Visualization (1) order new medications and (2) cancel existing orders that are no longer operative. Perhaps. physicians can order facilitated medical errors. a reminder to do so may serve to reduce gories: (1) information errors generated by the frequency of such mistakes. which was medications at computer terminals not yet published in JAMA (Journal of the American “logged out” by the previous physician. Medication discontinuation failures are a commonly documented problem with 4. CPOE systems suffer from a range of textures are vital sources of knowledge. so 10-mg units are displayed on the CPOE acceptance. This can result in information/fragmenta. vate each previously ordered medication. the system CPOE system cancels their previous medica- facilitated 22 types of medication error and tions. the font is influential study examining the ways in which small. To reiterate. graphs. The dosages consequences resulting from health information listed on the display were based on the pharma. observa. images. technology implementations. This can Medical Association). texts with As is the case with other clinical information symbols (e. cli. consequences spurred the American Medical machine interface flaws reflecting machine rules Informatics Association to devote a policy meet- that do not correspond to work organization or ing to consider ways to understand and dimin- usual behaviors. External representations such as be canceled may serve to mitigate this problem. clinicians relied on CPOE information technologies. could facilitate improved management of those For example. shapes and systems.

e.. which develops an internal copy of a slice of a common abstract structure can have a sig- of the external world and stores it as a mental rep.. differ- tion is then retrieved when needed. Scaife and Rogers 1996). The external representa. Consider a out for precisely determining the time (Norman simple illustration involving multi-digit multipli. many of whom their place value). complex medication regimens. Now consider the use of tions (e. Medical prescriptions are an interesting case The individual brings to the task knowledge in point. First.e.g. For example. an analog clock pro- cation with a pencil and paper.g. For example.. they will exert a reasonably heavy load on work. There are various tions include the positions of the symbols. (Zhang and Norman 1994). A simple example (1993) argues that external representations play is that Arabic numerals are more efficient for a critical role in enhancing cognition and intel. rows and columns). These external It is widely understood that not all representa- representations served as a stimulus to be inter. imagine cal. memory retrieval and computationally pencil and paper as illustrated below. Chronic illness affects over 100 million of the meaning of the symbols (i. intensive reasoning) and that displays can reduce the amount of time spent searching for critical 3 7 information. Although these spatial relations (i. by holding partial results are inherently hard for patients to follow. ent forms of graphical displays can be more or This view has changed considerably. 1993). computations is a central argument in support of classifying dermatological lesions). leading to distributed cognition. prescriptions written by clinicians visual representation. tions are equal for a given task and individual. Unless ily determine time intervals (e. Norman less efficient for certain tasks. 1999). 3 × 37 = 111). tables and lists can dramatically change 3 3 3 decision-making strategies (Kleinmuntz and 3 4 4 1 Schkade 1993.g. Larkin and Simon (1987) argued that effective ing memory in relatively short order. The The representational effect is a well-documented hard work is then done by the machinery of the phenomenon in which different representations mind. Calculations can rap- the visual system would process the informa. The appropriate internal representa. One may displays facilitate problem-solving by allowing have to engage in a serial process of calculation users to substitute perceptual operations (i. On the other hand..e. These durable representations (at als (XXXVII × XCIII) even though the repre- least those that are visible) persist in the external sentations or symbols are identical in meaning. arithmetic operators. and suffer from multiple of these individuals suffer addition and multiplication tables (that enable a from multiple afflictions and must adhere to look-up from memory).4 Cognitive Science and Biomedical Informatics 139 inputs to the mind (Zhang 1997). The are helpful. elapsed or you are unusually skilled in such computations. remaining time) without recourse to . which is the subject of the knowledge being retrieved from memory and next section.ketabpezeshki.g. idly become computationally prohibitive with- tion in a display that would serve as input to the out recourse to cognitive aids. nalized (e. reasoning. extends a person’s working following prescriptions were given to a patient (021) 66485438 66485457 www. memorized) by the system. vides an interface that enables one to more read- culating 37 × 93 without any external aids.g. Research has demonstrated that dif- x 9 3 ferent forms of graphical representations such as 1 1 1 graphs.. rec- and maintain partial products in working mem. the pill organizers and mnemonic devices designed partial products of interim calculations and their to promote patient compliance. and computation. The outside the mind. digits and individuals in the United States. The offloading of cognitive system for further processing (e..g. arithmetic (e. world and are continuously available to augment Similarly. 37 × 93) than Roman numer- ligent behavior. nificant effect on reasoning and decision making resentation. a digital clock provides an easy read- memory. memory (Card et al.... resulting in a decision or action. ognition processes) for effortful symbolic opera- ory (e.

. Norman 1993).e. reading a table requires a certain level of numeracy that is beyond the The physician’s list is concise and presented abilities of certain patients with very basic edu- in a format whereby a pharmacist can readily fill cation. I took my last dose of Lanoxin 6 h ago) blood pressure values. Carafate –1 tablet before meals and at bedtime However. blood pressure monitoring device and math- umns) and by medication (rows). Day relationship in a tabular format (i. Patel and D.140 V. when do I take Zantac) and computation (e. The familiar monitoring device or medication. However.R.e. the organization by terized the difficulties some older adult patients medication does not facilitate a patient’s deci. provided an easy readout and patients could tation can transform a cognitively taxing task readily make appropriate inferences (e. 1988 reported in into a simpler one that facilitates search (e. memory retrieval example.. Carafate X X X X The results suggest that even the more literate Zantac X X patients were challenged when drawing infer- Coumadin X ences over bounded periods of time. Kaufman following a mild stroke (Day. experienced in dealing with numeric data.. The patient ematical representation in tabular format (sys- can simply scan the list by either time of day tolic/diastolic). the values expressed on the interface of their the items can be organized by time of day (col.g.g.L. This simple change in represen. several patients appeared and inference (what medications to bring when to lack an abstract understanding of covaria- leaving one’s home for some duration of hours) tion and how it can be expressed as a functional are necessary to make such a decision.. They tended Adapted from Norman (1988) to focus on discrete values ( . Kaufman and colleagues (2003) charac- the prescription.11. For example. Tables Inderal –1 tablet 3 times a day can support quick and easy lookup and embody Lanoxin –1 tablet every AM a compact and efficient representational device. ficulty establishing the correspondence between In this matrix representation in Table 4. a single reading) Fig.2).ketabpezeshki.11 Mapping values between blood pressure monitor and IDEATel Table (021) 66485438 66485457 www.2. Some computation.g.2 Tabular representation of medications priate measures. cells and proposed an alternative tabular representation rows) as illustrated in Fig. when reviewing their blood glucose or (e.. how many pills are taken at dinner time). a particular external representation Zantac –1 tablet every 12 h (twice a day) is likely to be effective for some populations Quinaglute –1 tablet 4 times a day of users and not others (Ancker and Kaufman Coumadin –1 tablet a day 2007). For of day. However.g. sys- tolic value is higher than usual) and take appro- Table 4. Others had dif- (Table 4. espe- sion of what medications to take at a given time cially when represented in tabular form. certain patients had Lanoxin X difficulty recognizing anomalous or abnormal Inderal X X X results even when these values were rendered as Quinaglute X X X X salient by a color–coding scheme. 4. when interpreting Breakfast Lunch Dinner Bedtime the same values in a table. 4..

the law of cal instructions can be remarkably complex. methorphan hydrobromide U. problem-solving. Even when calculations were cor- the medium of representation rather than the rect. more readily perceived. They of course have a long history as tools of or interpreting medication instructions). symmetry indicates that symmetric objects are Each teaspoonful (5 mL) contains 15 mg of dextro. The psychological study understanding of which often demands that the of information displays similarly has a long his- user translate minimal.g. a graph- for proper functioning. Card and colleagues (1999) define information DOSAGE CHILDREN: 1 mg/kg of body weight daily in 3 or 4 divided doses. proximity states that visual entities that are close Consider the following instructions for an over. Branch. together are perceptually grouped. “a picture is worth 10. where science and cultural inventions that augment correct processing of information is necessary thinking. Diagrammatic representations configuring and using a digital video recorder are not new devices for communicating ideas. proce. the correct municating patient data. Medical errors involving the use of ther.g. visual representations of abstract data If you wish to administer medication to a 22 lb to amplify cognition”. Medical data can include single 22lbs / 2. (DVR).S. laboratory data) or non-numeric lay subjects’ responses to this problem. as well as a minimal familiarity with be an ancient Chinese proverb (Larkin and Simon the application domain. cooking something for the first time. the calculations are as follows: application (Kosara and Miksch 2002. interactive.. Starren and Johnson 2000).. in a palatable Advances in graphical user interfaces afford yellow. tion for describing how we see patterns in visual The calculation of dosages for pharmaceuti.000 words” is believed to soning. tsp / 3doses / day = 2 / 9 tsp / dose Representations can also be characterized as either Patel. 1 or 2 teaspoonfuls three or four times daily. and Arocha (2002a) studied 48 numeric (e.P. there is a significant mismatch between the Instructions can be embodied in a range designer’s conceptual model of the pharmaceuti- of external representations from text to list cal text and procedures to be followed and the of procedures to diagrams exemplifying the user’s mental model of the situation. DOSAGE ADULTS: a wide range of novel external representations. The comprehension of ical representation of regions of geographical written information in such cases frequently space. but maintained a daily diary background. Information visualization child three times a day and wish to determine the of medical data is a vigorous area of research and dosage. quantitative formulas tory dating back to Gestalt psychologists begin- into qualitative. aggregating and com- the-counter pharmaceutical labels. visualization as “the use of computer-supported. information storage. There were no significant considerable difficulty reading the table on the differences based on cultural or educational computer display.. One patient experienced amount to administer. the earliest maps. and domains (e. Everyday nonexperts are called upon to Diagrams are tools that we use daily in com- follow instructions in a variety of application . ning around the turn of the twentieth century. images (Ware 2003). Visual (021) 66485438 66485457 www.ketabpezeshki. lemon flavored syrup. The information such as symptoms and diseases. they were unable to estimate the actual form of representation. completing income-tax forms. the problems correctly calculate the appropriate dosage of with the representation seemed to be related to cough syrup. External representations have always been apparent. planning.5 %) were unable to expected range.4 Cognitive Science and Biomedical Informatics 141 in noting whether it was within their normal or majority of participants (66.2lbs / kg × 1mg / kg / day / 15 mg / data elements or more complex data structures. dures. and frequently complex. The phrase involves both quantitative and qualitative rea. For example. The law of the-counter cough syrup. date back thousands of years. steps. In at least one case. than in the case of over. One of the central problems is that with very similar representational properties. They produced a set of laws of pattern percep- apeutic drugs are amongst the most frequent. For example. This is nowhere more 1987). and critical. a vital means for storing.

They characterized (changing as additional temporal data become five major classes of representation types includ- available). and blood pressure a representation affect latency? (Gardner and Shabot 2001) (see Chap. including both numeric Each of these data types has distinct measure- and nonnumeric (Tang and McDonald 2001).. reducing the search for information (grouping larly in relation to genetic sequencing and align. Electronic Health Records nitive studies characterizing how different kinds of medical data displays impact performance. representation).. radiology). drawing attention to events that we know very little about what constitutes a require immediate attention).. (4) produced at a very fast pace. and processing resources available to the users Information visualization is an area of great (offloading cognitive work to a display). which mental representations mediate all activity tions. monitoring (e. 1998). Medical imaging systems authors propose some criteria for evaluating the are used for a range of purposes including visual efficacy of a representation including: (1) latency diagnosis (e. (2) by importance in bioinformatics research. computational advances in infor- discussed previously. and more six major ways: (1) by increasing the memory reliable performance. There is a significant opportunity for cognitive methods and theories to play an instrumental role in this area. graph. EHRs need to include a wide range of ing list. Patel and D. there have been several efforts to model of information-processing cognition in develop a typology of medical data representa. In this chapter. tasks and users. the user can select different possible views to sary to use such representations effectively? highlight variables of interest). Starren and Johnson (2000) proposed a tax.L. respiratory rate. (2) accuracy. (1999) propose could lead to significantly faster. The different kinds of users. what inferences can be more readily updated observations) for the presentation of the gleaned from a tabular representation versus a presentation of physiological parameters such as line chart? How does configuration of objects in heart rate. The tools and applications are being tations to enhance the detection of patterns. (the amount of time it takes a user to answer a ning.g. Further research is needed to cal structures in either two or three dimensions. research and administrative tasks by different kinds of data. data strategically). The purposes of tion). Although there is by using perceptual attention mechanisms for tremendous promise in such modeling systems. assessment and plan. EHR ment properties (e. data representation types.ketabpezeshki. and education and training question based on information in the representa- (Greenes and Brinkley 2001). 4. icon. Card et al.g.R. and generated text.6. ordinal scales are useful for data representations are employed in a wide range categorical data) and they are variably suited for of clinical. impacted performance (Lin et al. What sorts ing information in a manipulable medium (e. Kaufman representations may be either static or dynamic onomy of data representations. They . As At present. easier.g.g. we have considered a classical However. dynamic representations (e.4 Distributed Cognition and In general. 19). continuously For example. explore the cognitive consequences of different Patient monitoring systems employ static and forms of external medical data representations.142 V. Lin et al. of competencies or prerequisite skills are neces.. and (5) by encod- usable interface for particular tasks. The (021) 66485438 66485457 www. we are gaining a better understanding also demonstrated that redesigning interface in a of the ways in which external representations can manner consistent with human factors principles amplify cognition. table. particu. communication. (3) by using visual presen- ment. found that that the mation visualization have outstripped our original display of a PCA device introduced sub..g. and (3) compactness (the rela- these representations are to display and manipulate tive amount of display space required for the digital images to reveal different facets of anatomi. there have been relatively few cog. and constitute the central units of analysis. understanding of how these resources can be stantial cognitive complexity into the task and most effectively deployed for particular tasks.

“Insert” menu and then execute the action. In the dis. nurses vidual to the organization. Cole and Engestrom (1997) ory each time an action is to be performed. crete. For suggest that the natural unit of analysis for the example. The traditional a spectrum of viewpoints on what constitutes cognitive analysis might account for this skill by the appropriate unit of analysis for the study suggesting that the user has formed an image or of cognition. This suggests that the display and gynecology department is seen as an inter- can have a central role in controlling interaction related assembly of things (including humans) in graphical user .. uated at several levels of analysis from the indi- orative nature of cognition (e. the conventional information.ketabpezeshki. He names of menu headers. but internal (i. yet they could routinely argues that “work practices are conceptualized make fast and accurate menu selections. visual displays. there are individuals. and retrieves this information from mem. In Berg’s view. Berg (1999) goes on processing approach has come under criticism to emphasize that the “elements that constitute for its narrow focus on the rational/cognitive these networks should then not be seen as dis- processes of the solitary individual. study of human behavior is an activity system. Draper. if the goal is to “insert a clip art icon”. knowledge) and external representa. of cognition from being the sole property of the Although there are compelling reasons for adapt- individual to being “stretched” across groups.. cognition has two central points of inquiry. The distributed individual and collective. Mayes. outpatient clinic or obstetrics menu selections. In the previ. 2002b). This is consistent with a distributed cognition tributed approach to HCI research. whether (021) 66485438 66485457 www. An relied on cues in the display to trigger the right emergency ward. ing a strong socially-distributed approach. also considers the importance of an individual’s tions (e.g. Berg is McGregor. manuals). individual’s mental representations and external Suchman 1987). as a single indivisible unit of analysis. we consider the relevance of external characteristics (p 89). delivery of patient care. The emerg. well-circumscribed entities with pre-fixed ous section. Patel et al. Distributed internal representations (Perry 2003). the study of information sys- ing perspective of distributed cognition offers tems must reject an approach that segregates a more far-reaching alternative. cognition is framework that embraces the centrality of exter- viewed as a process of coordinating distributed nal representations as mediators of cognition. The as networks of people. one The mediating role of technology can be eval- that emphasizes the inherently social and collab. imagine an expert one that characterizes the mediating effects of user of a word processor who can effortlessly technology or other artifacts on cognition.e. as view of cognition represents a shift in the study well as the social and technical dimensions of IT. “culturally-organized environments”.. and technical support staff in neonatal care unit lates internal representations of the external jointly contributing to a decision process). the user would simply recall that this is subsumed comprising relations among individuals and their under pictures that are the ninth item on the proximal. documents and so forth” (Berg 1999). representations to cognitive activity. This viewpoint is increasingly representations are both instrumental tools in gaining acceptance in cognitive science and cognition (Park et al. Technologies. tools. whose functioning is primarily geared to the As discussed.g.4 Cognitive Science and Biomedical Informatics 143 analysis emphasizes how an individual formu. mation-processing. human and machine. and world. an material artifacts and cultures (Hutchins 1995. and Koatley (1988) demonstrated that a leading proponent of the sociotechnical point even highly skilled users could not recall the view within the world of medical informatics. However. groups of thereby achieving the goal. Let us first consider a more radi- schema of the layout structure of each of eight cal departure from the classical model of infor- menus. human-computer interaction research. 2001. negotiate tasks through a combination of key The distributed cognition perspective reflects commands and menu selections. organizational rou- results indicate that many or even most users tines. and technologies can be construed some problems with this model. To illustrate the point. doctors. A system consisting of individuals.

The our studies (see Kushniruk et al.L. Patel and D.g. DCI enhance or expedite performance. transform the ways individuals and enduring effects of an EHR (see Chap. directly enter information into the EHR. during the interview).144 V.12 Display of a structured electronic medical record with graphical capabilities (021) 66485438 66485457 www.12. (1991) use the system while collecting data from patients introduce an important distinction in considering (e. laboratory tests. of present . although a (Dossier of Clinical Information).. and the prescription of medication. The former is concerned allows the physician to record information about with the changes in performance displayed the patient’s differential diagnosis. in several of given technology may do all of these things. 12).R. 1996) and in actual clinical settings using cogni- ties (knowledge and skills) as a consequence tive methods (Patel et al. history individual? We believe it is important to under. what becomes of the the patient’s chief complaint. skills. groups think. They do not merely augment. The graphical interface provides tive load associated with a given task and per. a highly structured set of resources for represent- mit physicians to focus on higher-order thinking ing a clinical problem as illustrated in Fig. Physicians were encouraged to changes in individuals Salomon et al. but one that is qualitative in nature. such as In a distributed world. The laboratory of interaction with a technology. Using the difference is not merely one of quantitative pen or computer keyboard. We employed a pen-based EHR system. The DCI system the mediating role of technology on individual incorporates an extended version of the ICD-9 performance. This effect is research included a simulated doctor-patient Fig. drug mono- efficiently.ketabpezeshki. 4. physicians can change. The effects of technology refer laboratory-based research (Kushniruk et al. 2000). the effects with technology and the vocabulary standard (see Chap. and differen- stand how technologies promote enduring tial diagnoses. 1996). Kaufman they be computer-based or an artifact in another illustrated subsequently in the context of the medium. past history. In this capacity. such as diagnostic hypothesis generation We have studied the use of this EHR in both and evaluation. 4. the ordering by users while equipped with the technology. medical information graphs for medications. laboratory tests. when using an effective medical system also provides supporting reference information system. 7). and information on technologies may alleviate some of the cogni. of tests. physicians should be able information in the form of an integrated elec- to gather information more systematically and tronic version of the Merck Manual. The system effects of technology. to enduring changes in general cognitive capaci. The For example.

The paper-based records appear to of this screen-driven strategy. she also elicited numerous resemblance to the organization of information in irrelevant findings and pursued incorrect hypoth. reflecting the query structure of both effective as well as counter-productive uses the interface. In this particular case. using ten pairs of records provide guidance for asking the patient ques. the second physician using the EHR system and paper-based strategy involves the use of the EHR display to patient records. The results of one study replicated and cal cognition. In the screen-driven strategy. Perhaps.ketabpezeshki. after having used the screen-driven strategy. The study considered the fol. sicians entered significantly more information lowing questions (1) How do physicians manage about the patient’s chief complaint using the EHR. appear to have both strategies in their repertoire. We observed medical history. In addition. of EHRs on paper-based records on represented tive consequences of using the same EHR system (recorded) patient information. (2000) conducted a series of related The use of a screen-driven strategy is evidence studies with physicians in the same diabetes of the ways in which technology transforms clini. medium. information flow when using an EHR system? (2) Similarly. tain more information about the patient’s past age a certain sense of complacency. sion making. clinic. too reliant on the technology and had difficulty This finding is consistent with the enduring imposing her own set of working hypotheses to effects of technology even in absence of the par- guide the information-gathering and diagnostic. the physician was asked to conduct his in which they appear on the display to elicit next five patient interviews using only hand-writ- information. reasoning processes. the novice elicited almost system for 6 months. matched for variables such as patient age and tions. the user used it less discriminately.4 Cognitive Science and Biomedical Informatics 145 interview. For example. as evidenced in clinicians’ patterns extended the results of the single subject study of reasoning. The study first compared the contents which the subject pursues information from the and structure of patient records produced by the patient predicated on a hypothesis. one in the system. the structure and content imposed by information-gathering goals and allow of information was found to correspond to the the physician to allocate more cognitive resources structured representation of the particular toward testing hypotheses and rendering . information-gathering process. enduring effects of the use of EHR ences are likely to have an impact on clinical deci- systems on knowledge representations and clini. whereas. whereas a novice this is notably absent from the EHR. ticular system. On the other hand. All experienced users of this system ten paper records. and (3) Are there records. the structure and content of all of the relevant findings in a simulated patient the physician’s paper-based records bore a closer encounter. The results indicated that the EHRs contained In general. It's reasonable to assert that such differ- long-term. The authors also video-recorded and cal reasoning? One study focused on an in-depth analyzed 20 doctor-patient computer interactions characterization of changes in knowledge organi. Patel et al. phy- in a diabetes clinic. by 2 physicians with varying levels of expertise. the EHR than the paper-based records produced eses. a screen-driven strategy can enhance more information relevant to the diagnostic performance by reducing the cognitive load hypotheses. Patel and colleagues (2000) (reported above) regarding the differential effects extended this line of research to study the cogni. For example. (021) 66485438 66485457 www. the clinician problem type. A more experienced better preserve the integrity of the time course of user consciously used the strategy to structure the the evolution of the patient problem. After having used the system for 6 is using the structured list of findings in the order months. We have observed two distinct patterns zation in a single subject as a function of using of EHR usage in the interactive condition. EHRs were found to con- sions. However. this strategy can encour. physicians represented significantly What are the differences in the way physicians more information about the history of present organize and represent this information using illness and review of systems using paper-based paper-based and EHR systems. the subject became by the physician prior to exposure to the system. In employing this most striking finding is that.

146 V. afford turning or pushing order of information on the screen when asking downwards to open the door) or a water faucet. Although the expert user similarly used motion but actually need to be pushed to open the the EHR system to structure his questions. they render humans revealed that the less expert subject was more interactions with objects as effortless. the distributed resource model and analysis to a pate” in the process of cognition? The ecological provider order entry system. These include effects component (Hutchins 1995). one can often perceive the affordances of a the interface. screen (e. quence. and Harrison (2000) were among memory. door handles that appear to suggest a pulling 1996). then to identify areas of complexity that may The concept of affordance has gained substantial impede optimal recorded entries.R. This and processed by perceptual systems . demands on the user and the burden of working Wright. the use tions to produce new information. documented in a previous study (Kushniruk et al.g. of presented information on the EHR screen. indivisible information-processing system. uted cognition paradigm is that it can be used to The previously discussed research demon. & Harrison (2000) addresses the question such changes? What aspects of a display are more of “what information is required to carry out likely to facilitate efficient task performance and some task and where should it be located: as an what aspects are more likely to impede it? interface object or as something that is mentally Norman (1986) argued that well-designed arti. distributed resource model proposed by Wright. Fields. there are numerous artifacts This screen-driven strategy is similar to what we in which the affordances are less transparent (e.” The relative difference facts could reduce the need for users to remember in the distribution of representations (internal and large amounts of information.. In the distributed approach to HCI the first to develop an explicit model for coding research. For exam- strongly influenced by the structure and content of ple. The authors conclude that coordinating information from these representa- given these potentially enduring effects. the patient questions and recording the responses. When the affordances of an object The analysis of the physician-patient interactions are perceptually obvious.g. understand how properties of objects on the strates the ways in which information technolo. represented to the user. and effects of technology in the context of EHR most cognitive tasks have an internal and external use (Salomon et al. buttons) can serve as external gies can mediate cognition and even produce representations and reduce cognitive load. 1991). The research currency in human computer interaction.ketabpezeshki. External representations constitute affor- was much less bound to the order and sequence dances in that they can be picked up. cognition is viewed as a process of the kinds of resources available in the environ- coordinating distributed internal and external ment and the ways in which they are embodied representations and this in effect constitutes an on an interface. Kaufman One of the physicians was an intermediate-level enable individuals to know how to use them user of the EHR and the other was an expert user. body of research documented both effects with According to theories of distributed cognition.g. One the other hand. whereas poorly external) is central to determining the efficacy of designed artifacts increased the knowledge a system designed to support a complex task. Kaufman and Patel (2003a) applied How do artifacts in the external world “partici. (Rogers 2004). links. Horsky. analyzed. he was guided by the door handle (e.. Patel and D. The enduring changes in how one performs a task. The goal was to ana- approach of perceptual psychologist Gibson was lyze specific order entry tasks such as those based on the analysis of invariant structures in the involved in admitting a patient to a hospital and environment in relation to perception and action. the problem-solving process involves gathering strategies. It has consisted of two component analyses: a cognitive been used to refer to attributes of objects that walkthrough evaluation that was modified based (021) 66485438 66485457 www. of a particular EHR will almost certainly have a One of the appealing features of the distrib- direct effect on medical decision making. In particular. and as a conse- on knowledge-organization and information.L.. he door). What dimensions of an interface contribute to Fields.

The mon ground in clinical communication and can authors concluded that the redistribution and serve to update the state of the patient in a more reconfiguration of resources may yield guiding timely and accurate way. p.” The author has used the concept to as EHRs and patient safety. of system technology as well as adaptations that cially those who were new to the system. The study revealed barriers to the productive use sarily heavy cognitive demands on users. Distributed cognition analyses may also provide 2003. (2007. complexly configured displays) placed unneces. be deployed more effectively to establish com- terns of errors produced by clinicians.g. Following instrumental role in understanding and enhancing Hazlehurst et al. paper to harnessing the potential of cutting-edge tech- documents and blood pressure monitors) at a nologies in order to improve patient safety. remain largely uncharted.ketabpezeshki. port). principles and design solutions in the develop. These embodied in different individuals and inscribed are only a few of the cognitive challenges related in different media (e. a verbal utterance.. plays some functional role in a process within the Significant inroads have been made in areas such system. They can also serve an tion of system and human performance. very long menus. given point in time. These include under- cations of a surgical team in a heart room. standing how to capitalize on health information Kaufman and colleagues (2009) employed the technology without compromising patient safety concept of representational states to understand (particularly in providing adequate decision sup- nursing workflow in a complex technology. workflow and patient-centered decision making. the performance of clinicians and patients as they tional state is a particular configuration of an engage in a range of cognitive tasks related to information-bearing structure. displays. or a printed label. EHRs. They extended the sentations/graphical forms mediate reasoning in construct by introducing the concept of the “state biomedical informatics and how these represen- of the patient” as a kind of representational state tations can be used by patients and health con- that reflects the knowledge about the patient sumers with varying degrees of literacy. settings is constituted using shared resources and representations. (021) 66485438 66485457 www. The framework for distributed cognition is ment of complex interactive systems.. 2007) have drawn on this framework to a window into why technologies sometimes fail illuminate the ways in which work in healthcare to reduce errors or even contribute to them. The authors conducted a lated clinical ordering task performed by seven qualitative study of the ways in different media physicians. However.4 Cognitive Science and Biomedical Informatics 147 on the distributed resource model and a simu. Hazlehurst and colleagues (Hazlehurst et al. The potential scope of applied cognitive display.7 Conclusion and tools.g. such as a monitor health. It offers a novel and potentially powerful proved to be particularly useful in understanding approach for illuminating the kinds of difficulties the performance of teams or groups of individuals users encounter and finding ways to better struc- in a particular work setting (Hutchins 1995). It is comprises actors 4. together with shared understandings among actors that structure interactions in a work Theories and methods from the cognitive sci- setting. that research in biomedical informatics is very broad. understanding how various visual repre- mediated telehealth setting. information technologies. 540).com . there are explain the process of medication ordering in an promising areas of future cognitive research that intensive care unit and the coordinated communi. The activity system is the pri- mary explanatory construct. The circumvented such limitations. ture the interaction by redistributing the resources. Technologies can resources model was also used to account for pat. “a representa. still an emerging one in human-computer interac- The distributed cognition framework has tion. espe. The CW analysis revealed that the and communication practices shaped nursing configuration of resources (e. The “propagation of representational ences can shed light on a range of issues pertain- states through activity systems” is used to explain ing to the design and implementation of health cognitive behavior and investigate the organiza.

7. quantitative representations for lower Evans. formance. M. This relatively recent article summarizes new directions in decision-making research. Although diagrams and graphical Carayon. theories. interaction. F. & Shortliffe. D. R. Emerging paradigms of cognition in medical deci.. P. Yoskowitz.. illustrate quantitative information. Y. A. 1. (1993). siderations that need to be taken into works: toward a multidisciplinary science. A very readable and rela. and frame. J. What are the implications effects of technology and the effects with of these error types for system design? technology. use the system. Explain the difference between the interface flaws. medicine. The use of electronic health records only) book devoted to cognitive issues in medicine. The first (and 4. (2007). opment of representations and tools for tively comprehensive introduction to human-computer biomedical informatics. (2003).. HCI models. Cognitive science in literacy populations. acy patients. (2002). A. 42(1). Speculate about the Patel. Rogers. of Biomedical Informatics. Preece.R. they Mahwah: Lawrence Erlbaum Associates. A multifac- eted introduction to many of the issues related to also present challenges for low numer- human factors. it serves as a cloak for our ignorance” (Henriksen 2008).). 35. L. J. Discuss the meaning of this quote Questions for Discussion in the context of studies of patient safety. Discuss the issues involved E. V. Kaufman.148 V. How can the study of infor- design: beyond human-computer interaction (2nd mation visualization impact the devel- ed. cognitive effects of the system on per- Patel. Koppel and colleagues (2005) docu- the distribute cognition framework? mented two categories of errors in What implications does this approach clinicians’ use of CPOE systems: 1) have for the evaluation of electronic Information errors generated by frag- health records? mentation of data and 2) human-machine 2. & Sharp.. Discuss the various con- Carroll. H. L. Journal would take to study system usability. Arocha. Things that make us smart: defend. (2009). . MA: MIT Press. What are some of the assumptions of 8. & Patel. Co. A large urban hospital is planning to sion-making. How can each of these effects contribute to improving patient safety and reducing medical errors? (021) 66485438 66485457 www. (EHR) has been shown to differentially This multidisciplinary volume contains chapters by many of the leading figures in the field. Patel and D. J. This book that persist after the clinician ceases to addresses significant issues in human-computer inter- action in a very readable and entertaining fashion... A. H. Briefly characterize the effects ing human attributes in the age of the machine. L. V.. An edited volume on cognitive approaches to HCI. implement a provider order entry sys- 52–75..). D.ketabpezeshki. affect clinical reasoning relative to paper Norman. Handbook of human factors materials can be extremely useful to and ergonomics in health care and patient safety.L. (Ed. Discuss some of the ways in which of learning and cognition with a particular focus on external representations can amplify biomedical informatics. (2007). (1989). & Arocha. “When human error is viewed as a cause rather than a consequence. West Sussex: Wiley. F.. N. D. Cognitive and learning sciences in bio- medical and health instructional design: a review with and suggests some of the steps you lessons for biomedical informatics education. J. Journal of Biomedical Informatics. 176–97. 5. potential impact of EMRs on patient care. V. charts. Interaction cognition. The authors tem. You have been asked to advise them articulate a need for alternative paradigms for the on system usability and to study the study of medical decision making. they have on reasoning. A review 6. healthcare and patient safety. including those Reading: Addison-Wesley Pub. San account in the development of effective Francisco: Morgan Kaufmann. Kaufman Suggested Readings 3.

to communicate with colleagues and friends. As electronic institutions? health records are increasingly deployed. to search the medical lit- erature. Shortliffe. desktop computers. Computer Architectures for Health Care and Biomedicine 5 Jonathan C. federated answers to these questions: and hosted systems changing the way the • What are the components of computer Internet will be used for biomedical architectures? applications? • How are medical data stored and manipulated in a computer? • How can information be displayed clearly? 5. 5735 South Ellis Avenue. IL.1007/978-1-4471-4474-8_5. clini- • What are the maintenance advantages of cians use information systems to record medical Software-as-a-Service? observations. Computation Institute. NorthShore University HealthSystem.). • How can the confidentiality of data stored in and review test results. Computer architectures for health care • What advantages does using a database man. Cimino (eds. 1001 University Place.C. and Evanston C. you should know the • How are wireless. and mobile devices tions today? to access the Internet.H. J. 149 DOI 10. Physicians and patients distributed computer systems be protected? use personal computing environments such as • How is the Internet used for medical applica. This chapter is adapted from an earlier version in the third Chicago 60637. • How do local area networks facilitate data Health professionals and the general public sharing and communication within health care encounter computers constantly. I. and to do their clinical and administra- J.J. touching every aspect of human life. mobile devices. University of Chicago and Argonne National Laboratory. USA edition authored by GioWiederhold and Thomas e-mail: foster@uchicago. and biomedicine are the physical designs and agement system provide over storing and conceptual plans of computers and information manipulating your own data directly? systems that are used in biomedical collaborative. IL. • What is the technical basis and business model Architectures have both physical and conceptual of cloud computing? aspects. Foster. interactive. USA increasingly . Silverstein. © Springer-Verlag London 2014 (021) 66485438 66485457 www. In fact. order drugs and laboratory tests. MD. PhD Searle Chemistry Laboratory. Rindfleisch. computers are ubiquitous.ketabpezeshki. Biomedical Informatics. Foster After reading this chapter. and e-mail: jcs@uchicago.1 Computer Architectures • What are the functions of a computer’s operat- ing system? Architectures are the designs or plans of systems. Research Institute. Silverstein and Ian T. MS (*) tive work. laptops. E.

On the surface. pointing devices. or massive computer cluster. but the selection of appropriate hardware. that acquire a computer and to develop its software. and cost. display and style of usage computers greatly outweigh the differences. or even whole computers on one silicon recorder. Silverstein and I. computer hardware has become dra. such as network cuits on silicon chips resulted in dramatic interfaces. equipment) . in the ways that they are inter. Our aim is to give you the background form general computation necessary for understanding the underpinning • Computer memories that store programs and technical architecture of the applications dis. users. but layered hardware and software architectures. Figure 5. Later we will discuss assemblies of concepts related to computer hardware. We will describe the • Storage devices. and data • Graphics processing units (GPUs) that per- form graphic displays and other highly paral- 5. particularly genomic and imaging data.C. developed that remove much of the burden of connected. ware to “higher level” commands. in the way they are managed and shared. lap. encapsulation or abstraction of underlying soft- in the way they are secured. more reliable and encoded information between these subsystems personal. video displays. memory.2 Hardware lel computations • Input and output (I/O) devices. in the types of applications that they writing the infrastructure of applications via can run. touch screens.1 illustrates the configuration of a simple ing. Foster Individual computers differ in speed. and shar. digital video devices.150 J. computers for complex applications. security. The result is ple interact with them. indi- cate with the outside world whether desktop. Essentially all modern general-purpose com- bled is crucial to the success of computer applica. we will cover fundamental devices. Despite these differences. At the “chip” embedded in medical devices. ples expressed by John von Neumann in 1945. Extending this to mod- not been concerned with their internal workings ern computers. communications. largely distinguish different individual hardware In this chapter. Only large institutions could afford to trollers. and distributed systems (multiple computers General computer architectures follow princi- working together). We assume that you use computers but have von Neumann machine. and solid state drives. that connect computers to networks increases in computing power per dollar. simpler for individuals to program and to use. data that are being used actively by a CPU cussed in later chapters. Most computers are now manufactured with nication capabilities have increased so have data multiple CPUs on a single chip. facilitate user interaction and storage In the 1960s. such as hospital information systems. they are composed of one or more or how computers work together across the global • Central processing units (CPUs) that per- Internet. storage. that computers are increasingly complex in their ences among computers can be bewildering. and in some (021) 66485438 66485457 www. in the number of users that At the same time. con- operate. software packages have been they can support. The scale conceptual level. and the architecture under which they are assem. electrical pathways that transport matically smaller. the differ. however. laptops. • Communication equipment. their assembly into complex distributed systems that provide long-term storage for programs that enable biomedical applications.T. vidual personal computers. puters have similar base hardware (physical tions. processing. storage volumes. such as key- Early computers were expensive to purchase and boards. optical disks. As computation. This is generally true computers use the same basic mechanisms to whether they are large systems supporting many store and process information and to communi. such as magnetic disks and component parts that make up computers and tapes. the similarities among all these of computing. capacity. including data acquisition. • Data buses. faster.ketabpezeshki. gaming system. mobile top. software. and commu. and printers. the development of integrated cir. mobile device. software. and in the way peo. Since of computers that time.

I/O devices and communica. storage. The challenge then is for that make up the central component of a com- the software to distribute the computation across puter: the CPU.1 Central Processing Unit switch that can be set to either of two states stores a single bit value. Thus. For example. these units to gain a proportionate benefit. Each bit can assume one of two values: 0 or 1. combined to form basic units that can store letters 0101. and so on. the underlying principles are simple. turn are assembled into the larger functional units tions simultaneously). 0010.2. up to 1111. 5.1 The von Neumann model: the basic architecture of most modern computers. These registers in lel processing (performing multiple . A prime example is a processing unit itself. 0001. an area for memory.) These primitive units are be complex. add digits.5 Computer Architectures for Health Care and Biomedicine 151 Fig. switches can store 24. with one another. different combina- nization: primitive units (electronic switches) are tion of values: 0000. four structuring principle is that of hierarchical orga. as well as multiple layers and numbers. ing text and large numbers. The logical atomic element for all digital com- puters is the binary digit or bit. An electronic 5. 0100. 0110. The basic units are assembled tion interfaces. and compare values of memory. The computer comprises a single central processing unit (CPU).ketabpezeshki. or 16. 0011. 4 bits (021) 66485438 66485457 www. Multiple interconnected CPUs into registers capable of storing and manipulat- with shared memory layers further enable paral. (Think of a light switch that Although complete computer systems appear to can be either on or off. Here Sequences of bits (implemented as sequences of simple components can be carefully combined to switches) are used to represent larger numbers create systems with impressive capabilities. The and other kinds of information. the building blocks of computer systems. and a data bus for transferring data between the two cases multiple GPUs.

and comparison (021) 66485438 66485457 www. the sequence 0101 is the binary (base 2) processor) or for additional special characters representation of the decimal number 5—namely. spoken language. a convention for repre. The Delete and Arrow keys are only decimal integers but also fractional num. subtraction. (such as currency and mathematical symbols or 0 × 23 + 1 × 22 + 0 × 21 + 1 × 20 = 5. . and more complex data A standard called Unicode represents characters types such as pictures. and Function keys are used to modify case letters. memory. it can take on 28 or 256 values base standard does not cover its use. bers. The upper-case and lower-case alphabet. 5. which is the common way of transmitting cal functions that are basic to information pro- and storing these characters. and the Control. and common punctuation characters are shown here with their ASCII representations can represent any decimal value from 0 to 15.2 shows the ASCII is a small subset of Unicode. Silverstein and I.. a the CPU performs the mathematical and logi- byte. acters seen on a keyboard can be encoded and Groups of bits and bytes can represent not stored as ASCII. Not all char- (0–255). 7 bits are commonly placed into an 8-bit unit. general characters (upper-case and lower. the decimal digits.C. These By manipulating the contents of its registers. senting 95 common characters using 7 bits.T. American Standard Code for Information The CPU works on data that it retrieves from Interchange (ASCII).152 J. content of a medical record.g. Figure 5. The eighth bit may cessing: addition. be used for formatting information (as in a word e. and punctuation marks). placing them in working registers. A byte is a characters with diacritic marks).2 The American Standard Code for Information Interchange (ASCII) is a standard scheme for representing alphanumeric characters using 7 bits. digits. but the ASCII sequence of 8 bits. and the needed for foreign languages using up to 16 bits. instructions to the CPU.ketabpezeshki. other keys or to interact directly with programs. often dedicated to edit functions. Foster Fig.

Increasing powers of 210. or 1. it is limited in size.2. flash memory or solid state disks. and storage of data into memory or regis. and (2) archival storage is used to rent use. being specialized for fast read–write access. It is used to store a few crucial ger periods than in volatile memory are stored on programs that do not change and that must be long-term storage devices. and data that are in cur. long-term interface. It is also volatile: 5. is permanent and term storage. which usually is an faster computers had 32-bit and now 64-bit word operation that requires the retrieval. tions we need to store more information than can Working memory has two parts: read-only be held in memory.048.824 bytes. Some instructions can direct number. or address. CPU as a unit is called a word. and memory contents are not retained when The computer’s working memory stores the pro. optical disks.024 bytes. the medical record both read and written into.” “is equal to. word in memory by specifying the sequence gram in sequence.” “is less than”).741. a sharing of information. Such a transfer of flow control provides 5. the CPU also has registers that it uses to petabytes. the medical record of a patient who is the screen. the base is loaded from such storage into working mem- graphics that run the Macintosh computer ory whenever it is used. each of which provide persis- instructions that is executed each time the com.g. or results we place them into long- ROM. e. of its starting byte. gigabyte is 230. numbered in sequence. storage can be divided into two types: (1) active More familiar to computer programmers is storage is used to store data that may need to be RAM. tion. a computer is an instruction follower: a function of the computer’s design. It is used to store the of a patient who currently is being treated within programs. often just called memory. ROM also is used to store pro. Parallel flows may also be invoked.2 Memory its contents are changed when the next program runs. The needed information grams that must run quickly—e. or 109. a set of initial or magnetic tape. sizes that allow processing of larger chunks of tion. data written on (021) 66485438 66485457 www. In addition to registers that perform computa. RAM can be retrieved with little delay. data. Early com- it fetches an instruction from memory and then puters had word sizes of 8 or 16 bits. . the processor to begin fetching instructions from a different place in memory or point in the pro- gram. or 103.. unchanging. gram is the bootstrap sequence. Whereas memory is dedi- megabyte is 220.. but it cannot be Programs and data that must persist over lon- altered or erased. or 1. or 103. For many medical applica- grams and data currently being used by the CPU. and we want to save all that memory (ROM) and random-access memory information for a long time. computations and the images to be displayed on e. RAM is much larger than ROM. It can be read. we might speak of a 2 gigabyte memory Computer storage also provides a basis for the chip. newer. It also holds the intermediate results of store data for documentary or legal purposes. One such predefined pro. The CPU accesses each fetching and executing each instruction of a pro.073. To save valuable pro- (RAM). the hospital.2. executes the instruction. available at all times. are terabytes. information at a time. control values. and exabytes (1018). or 106. grams. The bytes of memory are ters. store instructions—a computer program is a set A sequence of bits that can be accessed by the of such instructions—and to control processing.g. The computer’s memory is relatively expensive. therefore.. A kilobyte is 210. Conceptually. tent storage for less cost per byte than memory puter is started. The processor often performs a simple loop.576 bytes. power is turned off. or fixed memory.g.3 Long-Term Storage flexibility in program execution. and are widely available. such as hard disks.5 Computer Architectures for Health Care and Biomedicine 153 (“is greater than. For deceased. and a cated to an executing program. The word size is In essence. manipula.ketabpezeshki. or 1.

6 Output Devices 5.ketabpezeshki. However. This is eas- used for general computation. Many systems compute informa- costly and awkward aspects of medical data pro. and a three-dimensional dis- simultaneous processing of many data elements play may provide feedback to the user. 5. When GPUs are review of the resulting text is needed.4 Graphics Processing Unit to type. they are referred to as general purpose GPUs or GPGPUs. rather than for ier for some users than typing. computer-controlled force or tactile feedback. is the complementary step in the processing of Data and user-command entry remain the most medical data. the primary surgical incision.g. Files and data. programming languages tion of sentence structures. tions simultaneously. redundant data entry can be mini. This ical virtual-reality environments also provide allows rapid manipulation of many memory loca. or of the output. Foster storage in file systems or in databases is avail. and use grammar rules to allow recogni- different. purpose for which they were invented. common instrument for data entry is the key- able to other users or processes that can access board on which the user types. tion that is transmitted to health care providers cessing. rections is convenient. electronic signals that can be transmitted to com.2. track pad. or frame three-dimensional pointing devices used in med- buffers. e. Certain data can be acquired automati. less mature. using optical capture.5 Input Devices The presentation of results. There are also three-dimensional pointing Graphics processing units are specialized for devices. and is displayed immediately on local personal cally. A cursor indi- the computer’s storage devices. and many diagnostic radiology on a display screen. where an . such as a mouse. cates the current position on the screen. video or graphics display. of a simulated needle being inserted general. or writers and readers need not be present at the touch screen.T. for computer graphics. crystal display (LCD) or light-emitting diode Furthermore. but such as Open CL. so that making insertions and cor- same time in order to share information. The speech input is and software architectures than CPUs making then stored as ASCII-coded text. is positioned in neously.2. many laboratory instruments provide computers or printed so that action can be taken.2. The most plays enables unlimited interaction with text. for venipuncture or a simulated scalpel making a cient than CPUs for video display. Silverstein and I. (LED) based displays of a personal computer mized if data are shared among computers over (PC). This technology them harder to use to date. voice signals captured through a microphone are such as comparing genomic sequences or image matched to the patterns of a vocabulary of known processing. They can be thought of as enabling of front of the screen. Some with tight coupling to their memory. digitized highly parallel computation on a single machine. essential to their power for so that a user can experience the resistance.154 J. GPUs generally require words. Most bases complement direct communication among programs allow the cursor to be moved with a computer users and have the advantage that the pointing device. is improving in flexibility and reliability. Most immediate output appears at its destination puters directly. The near-realistic quality of computer dis- networks or across direct interfaces. Newer languages.. Systems developers con- tinue to experiment with a variety of alternative input devices that minimize or eliminate the need 5. such as the flat-panel liquid instruments produce output in digital form. GPUs are much more effi. but with limited instruction sets. enable some computer codes to error rates remain sufficiently high that manual run over either CPUs or GPUs. own body. but leaves less flexibility in example. or just the user’s handling computation over many bytes simulta. (021) 66485438 66485457 www.C. For example. or for In automatic speech recognition.

g. Examples are ultrasound Data communication and integration are critical observations. For create objects. Multiple data electronically among applications and com- standard hardware interfaces. (see Sect. respectively. tions in the health care organization.7 Local Data Communications the image (Fig. and business offices. determines the contrast and color resolu. but the ments. for alerts. video. Transmitting onto a screen for group presentations.2. the hardware in the various areas will also Their spatial resolution is often better than that of diverge—e. and blue com. intensive care high-volume printers the night before scheduled units. Ink- memory represent the output for each pixel. fessionals in various areas. just as is done in a copier. (MRIs). instance. produce sound via digital-to-analog conversion giving the intensity for red. Graphical output is essential for summa. feed- depth. have lower resolution than laser printers and are ture elements called pixels. Ink-jet printers A graphics screen is divided into a grid of pic. laboratories. so printing is best cialized to fulfill the diverse needs of health pro- done in advance of need. Transmitting paper results in a produced in image formats that can be shown on much more passive type of information sharing. One or more bits in relatively slow. The number of pixels per square inch determines the spatial resolution of 5. For computer systems and may be used in medical color displays. relevant por. such as for Automatic tion of an image. Over time. For example. Color ink-jet printers are inexpensive. In a clinic. duce paper copies. growth and funding to accommodate change (021) 66485438 66485457 www. Even if their hardware visits. the number of bits per pixel.216 colors. In a jet printers that produce images of photographic black-and-white monitor. 5. LCD Information can be shared most effectively by color projectors are readily available so that the allowing access for all authorized participants output of a workstation can also be projected whenever and wherever they need it. imaging departments will require displays. back and instruction. output is deeply intertwined. their content will differ. Liquid rizing and presenting the information derived ink is sprayed on paper by a head that moves back from voluminous data. Three sets of multiple bits are External Defibrillators (AEDs). three sets of 8 bits per pixel provide 224 or 16. or bit applications.ketabpezeshki.. guish 22 or 4 display values per pixel: black. Here the base on the screen is associated with the level of inten. Demand for resolution of several thousand dots per inch). magnetic resonance images functions of health care information systems. Modern computing and communications are For portability and traditional filing. other (commercial typesetting equipment may have a areas will use more processor power. the value of each pixel quality are also readily available. especially at high resolution. and computed tomography (CT) scans. such as physicians’ tions of various patient records may be printed on offices. and forth for each line of pixels.777. Printing information is slower Computer systems used in health care are spe- than displaying it on a screen. allowing 600 dots (pixels) per inch more capable displays and larger storage. Much diagnostic information is ported on PCs.5 Computer Architectures for Health Care and Biomedicine 155 images.3). graphics terminals. which is then used to . ink cartridges raise the cost under high use. Laser printers use an electromechanically is identical. green. printed on paper. 5.3). particularly sound. A computer can necessary to specify the color of pixels on LCDs. colors are merged while being sprayed so that true sity. or gray scale. Both parameters determine the size requirements for storing images. Videoconferencing is also sup- (3DTVs). pharmacies. and two intermediate shades of gray. Other output mechanisms are available to white. There are also 3D printers which ponents of each pixel color. and some of controlled laser beam to generate an image on a that content must be shared with other applica- xerographic surface. color mixes are placed on the paper. 2 bits can distin. such as VGA and puter systems facilitates such sharing by mini- HDMI also enable computers to easily display to mizing delays and by supporting interactive high definition and even stereoscopic televisions collaborations. and interactive graphical ele.

Foster Fig.C.. (1982).3 Demonstration of how varying the number of Whitman) is easily discernible (Source: Reproduced with pixels and the number of bits per pixel affects the spatial permission from Price R.. Basic and contract resolution of a digital image. 8 bits per pixel. Orlando: WB Saunders) (021) 66485438 66485457 www.ketabpezeshki.E. et al.). Digital radiography: A focus on array of pixels. the subject (Walt clinical utility. In: the upper right corner was displayed using a 256 x 256 Price R.T.R.156 J. (Eds. The image in principles and instrumentation of digital radiography.R. & James A. Silverstein and . 5.

spe. infrared. the Internet. This design choice is rationalized Communication can occur via telephone by the assumption that most users receive more lines. digital tion without involving the telephone company or cable services and wireless services. in downloading cables. digital subscriber line (DSL) tech. but all subscribers network (LAN) allows local data communica- then share that capacity. It transmits available data transmission options for connect. installation of a local-area tives have a very high capacity. cable modems using coaxial cable (up to 30 For communication needs within an office. different con.ketabpezeshki. usable wireless bandwidth exceeds be implemented by hardware switches. These alterna. Asynchronous Transfer Mode (ATM) is nologies allow network communications using a protocol designed for sending streams of small. There are many wide-area networks (WANs). or radio waves (wireless). Also for DSL. In and displaying graphics. with relatively diverse systems bridges the differences in com. In fact. WiFi to establish access to other computers and cific communication needs and network capabili. and many personal health and fitness municate less reliably than a wireless smart devices with embedded computers now support phone with a strong wireless signal. bits) is encoded in the symbols. among computer components (such as wireless The overall bit rate of a digital communica. Thus. and video. For example. depending on the distance from the dis. low-speed service used to communicate back to puting environments. a wired computer in a busy dard for wireless communications). Mbps. while Bluetooth is another wireless ties and loads must always be considered when protocol generally used for communicating designing applications and implementing them. variable-length messages or packets of informa- ing local networks (such as in the home).5 Computer Architectures for Health Care and Biomedicine 157 occurs at different times. a Mbps) or direct satellite broadcast. so for example. conventional telephone wiring (twisted pairs). fixed-length cells of information over very high- These allow sharing of data and voice transmis. This assumption breaks down if users ventions or communication protocols must be generate large data objects on their personal obeyed. Home networking has been further However. ports multiple ATM circuits. while each case different communication interfaces typing relatively compact commands to make this must be enabled with the computer. tion service. the data source. In and reliability can be expected. Computers. symbols can be transmitted and the efficiency Frame Relay is a network protocol designed with which digital information (in the form of for sending digital information over shared. for example. expanded with the use of WiFi (IEEE 802. fiber-optic data than they send. The underlying optical transmis- (Mbps). tion efficiently and inexpensively over dedicated Integrated services digital network (ISDN) lines that may handle aggregate speeds up to 45 and. sion circuit sends cells synchronously and sup- tribution center. communication paths can reach capac. or a campus. Communication among speeds are often asymmetrical. Such a network is (021) 66485438 66485457 www. speed dedicated connections—most often digital sion ranging from 1 to 10 megabits per second optical circuits. dedicated or shared wires. and a different balance of performance machine that then have to be sent to other users. hotel where the network is overloaded may com.11 stan- ity. with a given ATM circuit are queued and pro- vices may be unavailable. images. building. cessed asynchronously with respect to each other tions industry is broadening access to digital in gaining access to the multiplexed (optical) services over wireless channels. in many transport medium. cell phones. happen. Because ATM is designed to countries. later. The cells associated In areas remote to wired lines. headsets to cell phones and wireless keyboards to tion link is a combination of the rate at which computers).com . informa- wired bandwidth in many areas. this case it may be more cost-effective in terms of wired connections are typically higher volume user time to purchase a symmetric communica- communication channels and more reliable. these digital ser. Transmission for tion bit rates over 10 gigabits per second (Gbps) rapid distribution of information can occur via are typical. transmission network access provider. but the communica.

between computers as though the machines were cally takes responsibility for implementing and on the same LAN. cial device that is connected to more than one work. Separate ever. and switches shuttle packets different networks. ences should not be apparent to the user. high-speed communications. network connection. Twisted-pair considerations. so the effective transmission speed seen by remote LANs may be connected by bridges. so in-house delivery of networking ser- among multiple . both shared and private. fiber-optic a term to refer to any relatively high bandwidth cable.158 J. LANs are be transmitted in a few seconds. another network. each user may be much lower. Important services provided by such net. high. Silverstein and I. satellite signal. side-branch hubs to bring service to small areas There are a variety of protocols and technolo. network and is equipped to forward packets that work administrators include integrated access to originate on one network segment to machines on WANs. An alternate medium. but data. infrared. twisted-pair wires (Cat-5 and better cation of these have many special trade-offs and quality) have become the standard. routers. LANs can be duplicated in parallel. workstation. With improved air by radio. vices to the desktop is still easier using twisted- ual workstations can retrieve data and programs pair wires. Multiple users limited to operating within a geographical area of and high-volume data transmissions such as at most a few miles and often are restricted to a video may congest a LAN and its servers. The information technology among these networks to allow sharing of data department of a health care organization typi. offers the highest bandwidth (over 1 bil. venient communication between machines on Gateways. wireless cellular telephone services. or switches (see below). Appli- however. although the differ. the entire contents of this book could nodes. service reli. or line-of-sight laser-beam transmission. resources—data. can also translate packet formats if the two con- ability. microwave. from network file servers: computers dedicated and WiFi are often used in a complementary to storing local files. Many LANs still operate at 10 Mbps. providing con. Splicing and connecting into optical cable together at high speeds to facilitate the sharing of is more difficult than into twisted-pair wire. each packet contains the data to be sent. to increase transmission speeds and distances by Wireless users of a hospital or clinic can use at least one order of magnitude over twisted-pair these radio signals from portable devices to (021) 66485438 66485457 www. Users working at individ. A router or a switch is a spe- linking multiple LANs to form an enterprise net. nected networks run different communication Early LANs used coaxial cables as the commu. how- specific building or a single department. such as cable service or lion bps or 1 Gbps) and a high degree of reliability third generation (3G) or fourth generation (4G) because it uses light waves to transmit informa. Even network addresses of the sending and receiving at 10 Mbps. but has been used more recently as 100 Mbps. Fiber-optic cable. communication signal–processing technologies. nication medium because they could deliver reli. When demand is routers. Foster dedicated to linking multiple computer nodes wire. protocols. which are tion signals and is not susceptible to electrical now widespread means of broadband Internet interference. Broadband has a specific techni- wiring is inexpensive and has a high bandwidth cal meaning related to parallelization of signal (capacity for information transmission) of at least transmission. specifically to the Internet (see later dis. twisted-pair wires. users can process information locally and then shared backbone of an enterprise network or save the results over the network to the file server LAN and twisted-pair wires extending out from or send output to a shared printer. Gateways perform routing and cussion on Internet communication).ketabpezeshki.T. and security. and twisted-pair wires or WiFi to the individual gies for implementing LANs. and other control information. the 100-Mbps networks are now cost-effective. software. Messages also can be transmitted through the able. Fiber-optic cable is used in LANs access and communication. The fashion—fiber-optic cable for the high-speed.C. Typically Rapid data transmission is supported by data are transmitted as messages or packets of LANs. and equipment— however.

0 representation A prominent theme of this book is that capturing and entering data into a computer manually is 11 difficult. the continuous signal tromagnetic interference when proximity of a recorded on the ECG strip shown in Fig.4).0 to instruments can overcome these problems. and electrocardio- gram (ECG)—into electrical signals. applications. for example. portable wireless devices (e. and so on. reinforced concrete walls. signals—signals that vary continuously. time-consuming. Typically. converted to one of four discrete levels (represented by 2 resentation for storage and processing. which are Fig.0 rate. each sampled value is pled periodically and are converted to digital rep.g. 5. which (ADC). cellular phones. . A digital computer stores and battery life had made portable wireless devices processes values in discrete values collected at weakly fit computers in medical applications discrete points and at discrete times. (3) more nearest discrete unit (Fig.4 Analog-to-digital conversion (ADC). pulse –2. and often have cessing of human voice input. were wholly ana- of cellular or similar radio technologies was also log devices. and (4) better to distinguish between two levels (e. 5. cell phone causes public address speakers to The computers with which we work are digi- make chattering sounds. These features and short tal computers. display (see. 19). and enter the data –1. on or off). data-acquisition and signal-processing tech- rized.0 manually. 01 Direct acquisition of data avoids the need for people to measure. and 1. Two parameters determine how closely the digital data represent the original analog signal: 5. they acquired an analog prohibited for a time for perhaps justified fear of signal (such as that measured by the ECG) and electromagnetic interference with telemetry or displayed its level on a dial or other continuous other delicate instruments.5 Computer Architectures for Health Care and Biomedicine 159 communicate with the Internet and.g. 19. Now. if you wish to discriminate among four levels. Similar services. Real-time acquisition of data from 10 the actual source by direct electrical connections 0.ketabpezeshki. Most naturally occurring signals are analog sion may not be reliable internally over long dis. You experience elec. You can think of ADC as sampling and decrease the radio transmission power when they rounding—the continuous value is observed have stronger wireless signals (also reducing the (sampled) at some instant and is rounded to the risk for electromagnetic interference). to servers that contain clinical data and niques are particularly important in patient- thus can gain entry to the LANs and associated monitoring settings (see Chap. when autho. analog signals must due to improvements in battery technology. (2) be converted to digital units. Before com- until recently.0 expensive. tab..4). error-prone. In fact. in many medical settings the use bedside monitors. In this example. for example. Hospitals have many instruments that techniques also apply to the acquisition and pro- generate electronic interference.3 Data Acquisition and Signal Analog Processing Considerations representation 2-bit digital 2. The signals are sam- discrete values. The conversion pro- smarter wireless radio use by the portable devices cess is called analog-to-digital conversion (e. mechanical movement. Sensors attached to a patient convert 00 biological signals—such as blood pressure.g. The first tances. tablet computers). with: (1) better battery life puter processing is possible. let computers) are exploding in use in hospital you need two bits (because 22 = 4). ADC is a technique for transforming continuous-valued signals to transmitted to the computer. so that radio transmis. You need one bit reliable hospital wireless networks. Automated bits) (021) 66485438 66485457 www. encode.

we find that the basic contraction repetition The sampling rate is the second parameter frequency is at most a few per second. 5.e.2 V will be undetectable if the need to sample at least twice as frequently as the instrument has been set to record changes between highest-frequency component that you need to −2. On the other hand. Note the degradation of the quality of the signal from (a) bration of the instruments. For instance. Precision also is est. Improper ranging will result in loss of informa. you between 0. the QRS signal which a signal changes value will produce a poor rises and falls within a much shorter interval than (021) 66485438 66485457 www.. Fig. that in (b) is lower.0 b –2. 5. When the sampling rate is very low (as in c). over amplification will produce clipped peaks and troughs c the precision with which the signal is recorded and the frequency with which the signal is sam- pled. Figure 5. the greater the number of underlying analog signal. QRS wave within each beat (see Sect.C. the more bits. for individual patient variation. either manually or to (c) automatically. the details of the signal may be lost if the signal is insuf- ficiently amplified. looking at an shows another example of improper ranging.0 10 1. 17. the more tal estimate and their correctness determines pre.6 The greater the sampling rate is. verts and transmits the signal. The sampling rate in (a) is high- levels that can be distinguished. A sam.1 and 0.5 V increments.5 Effect on precision of ranging.0 and 2. For example.0 . The number of bits used to encode the digi. for exam- ple.0 01 –1. is necessary for signals to be repre- sented with as much accuracy as possible. the results of the limited by the accuracy of the equipment that con. and that in (c) is the lowest. analog-to-digital conversion (ADC) can be misleading.0 in 0. Ranging and cali. The amplitude of signals from sensors must be ranged to account. representation (Fig. i. a change in a signal that varies ing and storing the data. but that the that affects the correspondence between an ana.5) con- log signal and its digital representation. As illustrated here. Silverstein and I. As a general rule.5 observe in a signal. The precision is the degree to which a digi- tal estimate of a signal matches the actual analog value. On the other hand.ketabpezeshki.T. oversampling increases the expense of process- tion. 5. ECG.6). tains useful frequency components on the order pling rate that is too low relative to the rate with of 150 cycles per second.160 J.0 00 Fig. closely the sampled observations will correspond to the cision. Foster 2-bit digital representation 11 Analog a representation 2.

The rate calculated by doubling the highest noise greatly.) Unwanted components per wires. must be transmitted to the computer.g. 3. each sures the difference. closer to the source the conversion occurs. a differential amplifier mea. poor contact between mon way to achieve such a conversion. for signal transmission eliminates of the signal are assumed to be noise and are interference from electrical . tude. Repetitive sig- and instruments carrying analog signals all nals. the article by Wiederhold and Clayton (1985) in difference should reflect the true signal. two several cycles. it a 1 value to a 0 value or vice versa. digital signals can be coded. 2. Shielding. The the Suggested Readings explains Fourier anal- use of glass fiber-optic cables. analog signals can be converted into caused by interference can be filtered out.. isolation. Usually. Often. Thus. they typically are processed to reduce (021) 66485438 66485457 www. such as an ECG. In this case. Digital noise include random fluctuations in a signal transmission of signals is inherently less detector or electrical or magnetic signals picked noise-sensitive than is analog transmission: up from nearby devices and power lines. posed into its individual components. a frequency-modulated representation. signal is sent through lines that pass near other permitting detection and correction of trans- equipment. often used in combination. Fourier analysis can transmit the ground voltage at the sensor. the patient). a signal is decom- the destination. ysis in greater detail. Filtering algorithms can be used to reduce Three techniques. development of digital signal processing bances from signals produced by processes other (DSP) chips—also used for computer voice than the one being studied (e. Once interference rarely is great enough to change the signal has been obtained from a sensor. Some noise (such as the 60-cycle because optical signals are not affected by interference caused by a building’s electrical relatively slow electrical or magnetic fields. Most types of interfer. thus. these algorithms minimize the amount of noise in a signal before are applied to the data once they have been its arrival in the computer: stored in memory. When the noise pattern differs one to carry the actual signal and the other to from the signal pattern. because interference directly frequency is called the Nyquist frequency. mail and other applications—facilitates feres with the ECG) are other common sources of such applications. The sensor and source (e. thus reducing the effects of twisted wires are used to transmit the signal— random noise. Often.ketabpezeshki. En route. and distur. Frequency modulation (FM) reduces ond. the effect of noise. near the signal source is now the most com- Inaccuracies in the sensors. the analog signals are mission errors. no loss of data will occur during one or more spatial dimensions. ideas of sampling and signal estimation apply As long as the interference does not cre- just as well to spatially varying signals (like ate amplitude changes near the high carrier images) with the temporal dimension replaced by frequency. circuitry) has a regular pattern. For robust transmission over long dis. the ECG data-sampling as changes of frequency rather than of ampli- rate should be at least 300 measurements per sec. Another aspect of signal quality is the amount Conversion of analog signals to digital form of noise in the signal—the component of the provides the most robust transmission. An Once the data have been acquired and cleaned FM signal represents changes of the signal up. At be used to filter the signal. eliminated. noise. with a distinct period and amplitude. the Furthermore. and grounding of cables its relatively random pattern. transmission. can be integrated over reduce electrical interference. The disturbs only the amplitude of the signal. Placing a microprocessor susceptible to electromagnetic interference.g. (The ence affect both wires equally. respiration inter. the portion of the signal that is known to be tances. The acquired data that is not due to the specific phe. A characteristic of noise is 1. Primary sources of the more reliable the data become. nomenon being measured. instead of cop..5 Computer Architectures for Health Care and Biomedicine 161 the basic heart beat.

Functional convention: Under the most com- sequence of numbers: a.d. The Internet is growing rapidly. they are organized hierarchi. may be assigned a the duration or intensity of the ST segment temporary address that persists just during a con- of an ECG. to allow automated ECG-based car. with one of the top-level domain-class identifi- work. common. Numeric IP addresses may have to . peri- ness to failure of individual links in the network. Internet is a WAN that is composed of many called name-servers.162 J. Data care of keeping the translation up to date. IP addresses.c. the block address assignment pro- comparing a waveform to models of known cess. Designated computers..b. The most fundamental is the protocol suite change. (021) 66485438 66485457 www. This transmission is always by structured packets. computer devices are exhausting the old 32-bit tion.. and has been moving slowly for more than a Further analysis is necessary to determine decade. While the changeover is complex with stored profiles of spoken words. a period in which expressed as dotted sequences of name segments.g.ketabpezeshki. or features. each ranging posing Internet names from segments: from 0 to 255— most often written as a dotted 1. names addresses are not assigned geographically (the are composed of hierarchical segments increas- way ZIP codes are). much work has gone into making this the meaning or importance of the signals— transition transparent to the user. a second identifying a subnet. but the logical name for a resource can referred to as the Transmission Control stay the same and the (updated) DNS can take Protocol/Internet Protocol (TCP/IP). An Internet address consist of a (ICANN).T. routing takes place based on only the tional links. or Computers that cago. convert a name into an IP regional and local networks interconnected by address before the message is placed on the net- long-range backbone links. Often. including interna. and overall process is governed today by the Internet all machines are identified by a standard for IP Corporation for Assigned Names and Numbers addresses. The Internet is in the process of the shape of the waveform by comparing changing to a protocol (IPv6) that supports 64-bit one part of a repetitive signal to another. with a first component identifying a net. Names are also most often Science Foundation in the 1980s.g. work. numbers of an IP address and the parts of a name. when they are used to designate a remote machine—by means of a hierarchical External routers link the users on a LAN to a name management system called the Domain regional network and then to the Internet. Multiple names may be used for a given computer that performs distinct services. to detect ECG beat irregularities. have names assigned. computers on the Internet also words recognized in a spoken input.C. Foster their volume and to abstract information are permanently linked into the Internet may have for use by interpretation programs. the voice signals can be compared address space. The computer also can analyze nected session. Because 32-bit (and 64-bit) numbers are diffi- diac diagnosis or to respond properly to the cult to remember. Although IP mon convention for the United States. e. whereas users whose the data are analyzed to extract important machines reach the Internet by making a wireless parameters. The Internet was begun by the National numeric IP address.g. and a third ers—e. ing in specificity from right to left. and proliferation of networked individual or institution. The names can be translated to IP 5..institution. beginning cally.. In speech recogni. a fixed IP address assigned. or by the Internet. therefore. Silverstein and I. connection only when needed. odic reorganizations of parts of the network are All Internet participants agree on many stan.4 Internet Communication addresses—e. computer. The Name System (DNS). because the worldwide expansion of e. various networking approaches were overtaken by but there is no correspondence between the four a common protocol designed and inspired by mili. Three conventions are in use for com- sequence of four 8-bit numbers..uchi- identifying a specific computer.class (ci.class (whitehouse.g. of the signal—e. tary considerations to enable scalability and robust.

g. Such a central site can also provide a have been added. so we restrict ourselves to topics important value pairs that specifies the components to health care.g. Geographic convention: Names are composed vent system and individual intrusions. this type of address is gener. whose capabilities may differ. It is ( or apple. and the ally more difficult for humans to understand Department of Energy. form communities. Gateways of various types connect all provide a service whereby all communications these networks. works and their users is such that constant gov). . some operated by demanding commercial Domain name) are explicitly labeled and may users. Attribute list address ( educational. and users get access to the regional the X. . come in any computer or individual. outside users from internal naming conventions international organizations. vigilance is needed at these service sites to pre- military.400) convention: all these activities is more than we can cover in this Names are composed of a sequence of attribute. The scope of 3. non. or Many countries outside of the United States another service— is defined in terms of sets of use a combination of these conventions. electronic mail. Such a scheme insulates (for commercial. many more classes loading.nih. and . incoming e-mails) go to a single address no longer possible to show the Internet map on a (e. tain ourselves. Aeronautics and Space Administration.abdn. The Web’s popularity and grow- although additional fields in a URL. tion try (cnri.paloalto. Country designations differ such as the Department of Defense. . the National as well. unsolicited and unwanted connections or mes- fix to name the World Wide Web (WWW) sages (spam) and perform research. as well as personal and group Note that domain names are case-insensitive. These worldwide inter-computer link- Department at the University of Aberdeen (an age conventions allow global sharing of informa- academic institution in the United Kingdom). which is equivalent to the address now provided by myriad commercial communica- csd.state. /ADMD turn gets WAN access through a network access for Administrative Management Domain provider (NAP). There are other WANs besides the .us or This convention derives from tions enter- may be . information can be transmitted to most computers An institution that has many computers may in the world. www. book. people.coun.. such protocols that govern how computers speak to as csd. (021) 66485438 66485457 www.g. communications. Even with this limitation. remote computer log in. World Wide Web. A D M D = B T / P R M D = AC / O = A b d n / Regional and national networking services are OU = csd/. of hierarchical segments increasing in and /PRMD for Private Management Internet. .va. such as ing services continues to change how we deal with file names used to locate content resources. and can allow dynamic machine profit. make purchases. and changes. file transfer. be it resource naming. government.ketabpezeshki. Nearly all countries have and has not been adopted broadly in the Internet their own networks connected to the Internet so that Computer Architectures for Health Care and Biomedicine 163 Initially the defined top-level domain classes tables are used to direct each message to the right were . The routing of packets of information between ficity from right to left and beginning with computers on the Internet is the basis for a rich a two-character top-level country domain array of information services. for the Computer Science each other. /C for Country name. and others by parts of the federal government. Other conventions have firewall—a means to attempt to keep viruses and evolved as well: www was often used as a and ISP organizations. selection for the service in order to distribute As Internet use has grown.g.. Each such service— identifier—e. who in elements (e.g.400 address standard that is used mainly networks through their institutions or privately by in Europe. we can needed to resolve the address— e../C = GB/ only scratch the surface of many topics. It has the advantage that the address paying an Internet service provider (ISP). and then local single diagram. The nature of attacks on net- services on a computer (e. respectively).

Assembly lan. 5. com- symbolically.1 Programming Languages In our discussion of the CPU in Sect. Machine-code instructions are the only instruc- tions that a computer can process directly. To increase Using a higher-level language. computation and information processing by the CPU.g. however. are difficult for people ters and memory and to perform primitive to understand and manipulate. results. creation of an assembly language. user-oriented symbolic- one-to-one correspondence between instructions programming languages. Instructions that tell the processor which operations to perform also are sequences of 0’s and 1’s.5.5 Software All the functions performed by the hardware of a computer system are directed by computer pro- grams. such as one of efficiency.164 J. the problems that the users guage replaces the sequences of bits of machine.2.1.T. a first step toward making paring characters. These binary patterns. 5. in assembly and machine languages. 5. of a computer wish to solve are real-world prob- language programs with words and abbreviations lems on a higher conceptual level. transfer of data and programs to and from working memory. A pro.ketabpezeshki. to monitor the status of hypertensive STORE it back into memory. To make gram called an assembler translates these communication with computers more under- instructions into binary machine-language repre. a programmer instructs be able to instruct the computer to perform tasks the computer to LOAD a word from memory. Foster 5. a programmer defines instructions into macros and thus reuse them. On the other hand. and to store the result. and so on. Silverstein and I. and handling all processor programming easier and less error prone was the exceptions (Fig. They want to meaningful to humans. or to order new medications. computer scientists sentation before execution of the code. data acquisition from input devices. without worrying about the details of how the ing the computer to transfer data between regis. or software (e. The details (021) 66485438 66485457 . hardware performs these operations. An variables to represent higher-level entities and assembly-language programmer must consider specifies arithmetic and symbolic operations problems on a hardware-specific level. such as to retrieve the latest trends of their test ADD an amount to the contents of a register. exchange of data across networks). 5. formatting and presentation of results via the GPU. patients.7 An assembly-language program and a corre- sponding machine-language program to add two numbers language or machine code or just code.C. Thus. such as incrementing registers.7). instruct. we can combine sets of assembly those listed in Table 5. There is a developed higher-level. we explained that a computer processes information by manipulating words of information in regis- ters. People think best operations. standable and less tedious. a binary representation called machine Fig.

declarations. An interpreter converts and describe how the statements. The syntactic rules instead of compiled. business Procedural Compiled Coercion By reference Formatted files Ada 1980 Math. MUMPS (M) is an interpreted language. the statements at one time. whereas a compiler translates all define the language’s grammatical structure. . is NNN. tion. Hundreds of languages have been devel- ment an operation that may translate to thousands oped—here we touch on only a few that are of machine instructions. business Procedural Compiled Strong By name Formatted files Standard ML 1989 Logic. The following sets of state- times. Semantics is the meaning given to the various gram. GIVING C (format file choice FORMAT MOVE C TO LN 6 "The value ("The value is" F5.1 Distinguishing features of 17 common programming languages Programming First Primary application Type Procedure Data management language year domain Type Operation checks call method method FORTRAN 1957 Mathematics Procedural Compiled Weak By reference Simple files COBOL 1962 Business Procedural Compiled Yes By name Formatted files Pascal 1978 Education Procedural Compiled Strong By name Record files Smalltalk 1976 Education Object Interpreted Yes By defined Object methods persistence PL/l 1965 Math.5 Computer Architectures for Health Care and Biomedicine 165 Table 5.FFF" ( PLUS A B )) no layout 6 c10 ADD A TO B. and and LISP) all have the same semantics: C := A + B C := A + B LN IS "The value (SETQ C PRINTF ( C ) WRITE10. math Functional Compiled Yes By value Stream files MUMPS (M) 1962 Data handling Procedural Interpreted No By reference Hierarchical files LISP 1964 Logic Functional Either No By value Data persistence C 1976 Data handling Procedural Compiled Weak By reference Stream files C++ 1986 Data handling Hybrid Compiled Strong By reference Object files Java 1995 Data display Object Either Strong By value Object classes JavaScript 1995 Interactive Web Object Interpreted Weak By value or Context-specific reference object classes Perl 1987 Text processing Hybrid Interpreted Dynamic By reference Stream files Python 1990 Scripting Hybrid Interpreted Dynamic By reference Stream files Erlang 1986 Real-time systems Functional. translate automatically a high-level program into Each statement of a language is characterized machine code. COBOL. Compiled Dynamic By reference Stream files concurrent of managing the hardware are hidden from the FORTRAN routinely is compiled before execu- programmer. which can subsequently be executed many syntactic constructs. who can specify with a single state.2 ) WRITE LN is ~ 5. A compiler is used to important from a practical or conceptual level. ments (written in Pascal. creating a binary pro. LISP may either be interpreted or compiled. Some languages are interpreted by syntax and semantics. and executes each statement before moving to the other language constructs are written—they next statement.ketabpezeshki. 2F" C) (021) 66485438 66485457 www.

there is a macro function (021) 66485438 66485457 www. and later invoking. par- ficiently flexible to deal with nearly any type of ticularly as multiple functions are being run over problem. to assign size these various features. in programming. versus power. but more reliable. the computer executes similar standard procedures in the form of functions and (perhaps exactly the same) instructions to manip. For example. Language (SQL) of database management systems smart programmers can instruct the computers to (discussed later in this section). scale they become increasingly important. but the underlying con- the result to variable C. A language meant for education and certain presentation format.ketabpezeshki. to build the language itself. in an intuitive manner.C. Tools exist to combine cific libraries takes more time than does learning related functions for specific tasks—e. cating which operations to perform.. If the user saves a safest method. some coming from libraries (such as for. gen. e. users can search for. their use—this is called type checking. An important distinction among onto other icons is a form of programming sup- languages is how those arguments are transmitted. actions performed for later reuse. Each language has a distinct syntax for indi. such as C or and subroutines create an environment in which Java have to be augmented with large collections users can perform complex operations by speci- of libraries of procedures. large databases using the Structured Query tion. These procedures are called with argu. fying single commands. Some systems allow such the procedure to go back to the source. regression analysis and correlation. Discussions about languages often empha- to add the values of variables A and B. to programs and also provide a basis for reuse of the record and manipulate data with formulas in the work by other programmers. decimal num. Other users cedures.166 J. specific types of computing problems. aware that they are programming per se.g. such as integers. and to write the result onto cern is nearly always the trade-off of protection a file.g. subroutines. highly reliable programs will include features to Specialized languages can be used directly by make it foolproof. Procedures enhance the clarity of larger may use a spreadsheet program. because such languages define additional proce- bers. users can possible in a more constraining language. Threads allow multiple execution units.. types of values. from. the medical record number of a language is so natural that it matches their needs patient—so that a procedure to retrieve a value. Foster Each set of statements instructs the computer wanted. such as Sequences of statements are grouped into pro. as shown or concurrency. routines that spawn other routines. although all these languages are suf. Languages also differ in a forms interface that displays retrieved data in a usability. the turn mainly sequences of invocations of such pro. such as SAS or R. if the ments—e. such as Excel. the same data simultaneously. and strings of characters. In each case. Such fea. Silverstein and I. then he or she ence (a pointer to where the value is stored) allows has created a program. Regardless of Programmers work in successively higher lev- the particular language in which a program is els of abstraction by writing. in the end. Large programs are in cells of a spreadsheet matrix. the access mechanisms are hidden from the user. Languages that focus on a flexible. and as systems in Table 5. Within these they may also have ulate sequences of 0’s and 1’s within its registers. With the help of perform some operations more efficiently than is statistical languages. ported by many layers of interpreters and Just giving the value in response to a request is the compiler-generated code. Without type checking. physical details of the data storage structures and cedures. Moving icons on a screen such as the patient’s age might be: age (medical and dragging and dropping them into boxes or record number). .T. match throughout dures for specialized tasks and hide yet more detail.. by way of checking that the nonprogrammers for well-understood tasks. Built-in functions eral computational infrastructure. and learning the spe. Giving the name provides the most script (a keystroke-by-keystroke record) of the information to the procedure. mat in LISP) and others written for the specific The end users of a computer may not even be application.g.1. and giving the refer. called Computer languages are tailored to handle threads. which can scripts to be viewed and edited for later updates be efficient but also allows changes that may not be and changes. and retrieve data tures may cause programs to be slower in execu. perform extensive statistical calculations.

Data become information Users ultimately interact with the computer when they are organized to affect decisions. and their findings during a visit. Data can. 47-882-365 Clark. extra layer. equivalent. additional . These names provide the links to 62-847-991 Barnes. users are able to concentrate on higher-level problems of informa- Table 5. Handling data is made easier if the language opening. strictly background. controls communication among hardware languages allow only internal structures to be components. Tanner F 07 Dec 1997 the user’s work from one session to another. These applications must included with a computer system and it manages deal with large volumes of varied data and manage the resources. 2). and learning (see Chap. The other interactions that users have with the OS. it allocates the resources of the ory to external. The devices for the user. tistical results of data collected from two related If the language does not directly support the best databases.2 Data Management 5. input devices to output devices. The same conceptual basis is supervises other programs running in the com- not available for data management facilities. speaking. It assigns the CPU to specific tasks. details of file management such as the creation. for persistence. including patients. Laurel F 10 May 1998 Deleting files that are no longer needed and 55-202-187 Davidson. or data can the challenge of incorporating multiple functions be viewed as a hierarchy. data structure to deal with an application. Clare F 11 Nov 1998 made persistent. reading. structure of a medical record. Some puter. modern layers of software coupled with programming must be done to construct the desired programming expertise now make such complex structure out of the available facilities. 5. matching packages handle these diverse tasks.2).3 Operating Systems Data provide the infrastructure for recording and sharing information. In shared systems. and closing of data files. It is easy to envision a sys. Such interoperation is not simple. Thus. a model that matches lates users from much of the complexity of han- dling these processes. Accessing and that supervises and controls the execution of all moving data from the points of collection to the other programs and that directs the operation of points of use are among the primary functions of the hardware. on external storage. for system among the competing users. we still face well with the rows of a table (Table 5. through an operating system (OS): a program actions. or by a ticker Even though many powerful languages and tape. and them. by on Macintosh computers via AppleScript. Data can also be viewed as records. tem where a Web browser provides access to sta. be viewed as a stream. writing.2 A simple patient data file containing records tion management. The resulting interactions routine in health information sys. supports moving structured data from internal mem. The OS insu- instance. the kernel of the mathematical facilities of computer languages are OS resides in memory at all times and runs in the based on common principles and are. external library pro. other fields of the record contain demographic Programmers can write application pro- information grams to automate routine operations that store (021) 66485438 66485457 www.5 Computer Architectures for Health Care and Biomedicine 167 available in the Microsoft Excel spreadsheet and well with data produced by some instruments. such as memory.5. The OS is software that is typically computing in medicine. TCP connections over the Internet. and handles the grams are used for handling storage. storage. manages the transfer of data from made persistent. matching well with the into a larger system. Once started. their visits. therefore money) and introduces inconsistencies among applications trying to share information. Travis M 10 Apr 2000 archiving those that should be kept securely are Note: The key field of each record contains the medical record number that uniquely identifies the patient. typically costs effort (and tems and our everyday lives. however. in that case. persistent storage. They do get involved in specify- for four pediatric patients ing which programs to run and in giving names to Record number Name Sex Date of Birth the directory structures and files that are to be 22-546-998 Adams.

programs and their data are active simultane- High individual demands are best allocated to dis. Protected switches resources rapidly among all the jobs that memory is only available to the program that allo- are running. Foster and organize data. 5. their jobs. Because computers need to perform a tate the integration and communication of infor. to facili. Because people work slowly compared cated it. monitor patient status. the OS spends resources for sarily imply having multiple processors. or to dedi. switching. Thus. simultaneous programs executing in one system whereas servers are multiuser systems. 5. several application pro- mation. Such shared resource access is important by the OS within a single computer system. several processors (CPUs) are used demands. When it however.8 Virtual-memory system. cially under multiprogramming. and re-queuing jobs.6). Note. another program may be receiving input then filed by the OS and are available to its users from external storage. to grams reside in main memory simultaneously.C. on the machine at the same time. below in Database Management Systems. queuing. If the total Memory may still be a scarce resource. Silverstein and I. the computer can respond to multiple ing system for smart phones and tablets) similarly users. ating results on a printer. the overhead increases dispro. while the CPU is executing one medical computing systems. thus where databases must be shared. In a . they may not all fit in the physical memory tributed systems (discussed in Sect. To solve this cated machines where all resources are prioritized problem. With the use of multiple PCs typically operate as single-user systems. When many portionately and slows the service for everyone. more than five times as much. Virtual memory provides users with the illusion that they have many more addressable memory locations than there are in real memory—in this case. Programs and data stored on peripheral disks are swapped into main memory when they are referenced. and data into pages. which can be kept in tempo- Fig. to perform all the functions provided by multiple devices. the OS will partition users’ programs to a primary user.ketabpezeshki. Thus. to perform bookkeeping functions. to aid in education—in Multiprogramming permits the effective use of short. all users have simultaneous access to not interfere with another or with the OS. as long as they do not make very heavy ing systems. and another may be gener- when needed.168 J. which protected memory comes in to play. These programs are program. ously. it becomes important to ensure one program does tiuser system. espe- demand is too high. that multiprogramming does not neces- is managing sharing. Web browsers and the Apple iOS (operat- with CPUs. seemingly at the same time. all users protect one process from another for security have the illusion that they have the full attention of (unless authorized by the user). as we discuss increasing the overall processing power.T. variety of services. logical addresses are translated automatically to physical addresses by the hardware (021) 66485438 66485457 www. users interact through the OS. to perform analyses. In multiprocess- the machine.

it is important that the data be com- memory (Fig. an isolated generally associated with the kernel of an OS. The meaning of data elements and the rela- tionships among those elements are captured in the structure of the database. The field is the most primi- Throughout this book. and cation programs. their meanings and their relationships to other A large collection of system programs are data must be stored.g. 5.ketabpezeshki. human’s body temperature in degrees Fahrenheit rity management.8). so data may be reorganized. Virtual memory can be assists users with data entry. however. compilers to handle programs and is linked to other data necessary to interpret written in higher-level languages. starting and stopping programs. Such a storage allocation is ports the integration and organization of data and called virtual memory. we must keep together clusters of ing on system status). computers will be replaced. information. so users can and retrieval. For example. of birth. Computers provide the primary means as the patient’s identification number. such as unless we know that that number represents a graphical user interface (GUI) routines.g. admission date. To avoid loss of descriptive files. Programming data management allocate many more pages than main memory can software is particularly difficult when multiple hold. Also individual programs and their data can users share data (and thus may try to access data use more memory than is available on a specific simultaneously).5 Computer Architectures for Health Care and Biomedicine 169 rary storage on disk and are brought into main are difficult to write. handle their point-and-click to the physical state of the patient. the observation was taken at a computer system. name. e. Databases are col- 5. patient who is identified by a unique medical ware. debuggers for its value—the value pertains to a particular newly created programs. When a memory page is ref. and when the relationships among data an address mapping from the virtual address of elements are complex. Modern software libraries related data throughout processing. and substantial libraries of certain time (02:35. an observation grams to perform complex mathematical func. voluminous data rapidly and at unpredictable each address referenced by the CPU goes through times. data element (e.. the number 99. Systems records. long-term storage.. and admitting the programs to manage the data are complex and diagnosis. This mapping is han. the database of a hospi- access to relevant and complete data from diverse tal’s registration system typically has fields such sources. . For example. Database technology sup- memory as needed. when they must search through computer. tive building block. instrument that he used to acquire the values.7) is useless These programs include utility programs. The programmers tend to move from needed page from storage. functions. date for organizing and accessing these data. tionships can be complex. the CPU need for long-term reliability makes it risky to creates space for it by swapping out a little-used entrust a medical database to locally written pro- page to secondary storage and bringing in the grams. secu. allow a variety of fonts. must be linked not only to the patient but also to tions and routines to present and manipulate the person recording the observation.4 Database Management lections of data. For health care the program to a physical address in main applications. to the graphical displays that access a variety of appli. we emphasize the impor. communication soft. These rela- include tools such as sorting programs and pro. Under virtual memory management.5. and so on. diagnostic programs to help maintain the record number. typically organized into fields. and the like. the total use of virtual memory must be limited to Not only the individual data values but also a level that permits the system to run efficiently. and check. Fields are usually grouped together to (021) 66485438 66485457 www. as well as descriptive metadata. dled automatically by the hardware on most and the organizational units that maintain the machines but still creates significant delays. gender. each field represents one data tance to good medical decision making of timely element. and files (see Table 5. 7 Feb 2000) in a certain way standard routines (such as for listing and viewing (orally). many times the size of real memory. project to project. plete and virtually error-free. the erenced that is not in physical storage.2).

Fig.g. the physician create communication transactions dles the details of managing and accessing data. data requests on the input screen of the report- ing data. all male hypertensive patients aged 45–64 years because new fields are added to a record—the for inclusion in a retrospective study. A record is uniquely identified by one or more key fields—e. for additions and reorganizations to the files and users must understand the contents and underly- hence the metadata. to retrieve the records of the structure of the database changes—e. 5. DBMS from changes in the way that data are Thus. to the same set of accounts.9 An example of a simple database query written inventory. The desire for data indepen. All these files in Structured Query Language (SQL). there will be continuing requirements Query formulation can be difficult. personnel and payroll. tions to database systems can help a user retrieve A database management system (DBMS) is information using a menu based on the schema. The report generator then pro- ranges of valid values. transactional applications. DBMSs often also provide an alternative.. an integrated set of programs that helps users to More often. When data are shows the syntax for such a query using SQL. will use a database sys- The conceptual (logical) view of a database pro. by searching and matching patient ID in the record. tem without the pharmacist or ordering physician vided by a DBMS allows users to specify what being aware of the other’s presence. and so on. ing structure of the database to construct a query dence—i. stored. fields are associated with lists or generator program... however. Users retrieve data from a database in either of The metadata also specifies where in the record the two ways. front-end applica- a database management system for shared data. treatments. with the pharmacy. 64 years and whose systolic blood pressure is greater than vices. The the results should be without worrying too much medication-order records placed in the database by about how they will be obtained.. Users specify their vides facilities for entering. the pharmacy application A crucial part of a database kept in a DBMS is creates the daily drug lists for the patient care units. editing. Programs are insulated by the financial reports generated for business offices. the DBMS han. Foster form records. Often.170 J. a record may be located thereby improving database integrity. containing the needed metadata. and retriev. Records that contain similar information are grouped in files.. name rather than by address. database programmers formu- users independent from changes made to applica. such as a store and manipulate data easily and efficiently.g. and many other topics. the data- base of a health care information system will have separate files containing information about charges and payments. keeping the applications of one set of correctly. Figure 5. late the requests for health professionals. then. the resource utilization reports used by the contents and organization of the records of all health care administrators and the end-of-month the data files. called report generation. because the programs access data by field simpler means for formulating such queries. tions by another group—is the key reason for using To support occasional use. drug order–entry system.ketabpezeshki. For instance.C. In addition to files about patients and their diag- noses. thus. 140 mmHg Metadata describes where in the record spe- cific data are stored. A DBMS also pro.9 metadata must be changed as well. Often. Some database queries are routine requests— A schema is the machine-readable definition of e. to the same ser.e. and drug therapies. patient identifica- tion number and observation time. When an ad hoc fashion—e. the DBMS can detect duces the actual query program using information (021) 66485438 66485457 www. to the same personnel.T. to be shared.g. The program will relate to one another: they may refer to the same retrieve the records of males whose age is between 45 and patients.g. Users can query the database directly digits representing the birth date are located and using a query language to extract information in how to convert the data to the current age. and how the right record can and request correction of some data-entry . be located. the schema. Silverstein and I.

To achieve both of reviewed in terms of its costs and benefits. Two key ideas make it possible to manage the tion processing (OLTP) systems. typically complexity of network software: network service designed for use by thousands of simultaneous stacks and network protocols. but complex generation should. Basic patient informa. often for research. applications programs. A reliable database DBMS with different data architectures.5 Computer Architectures for Health Care and Biomedicine 171 stored in the schema. However. Thus. and care settings. and people resources. a DBMS shared patient data (such as a hospital’s medical facilitates the integration of data from multiple records) or international databases (such as bib- sources and avoids the expense of creating and liographic databases of scientific literature or maintaining multiple files containing redundant genomics databases describing what is known information. systems will duplicate the data in two separate natural.5. each other over local and remote networks brings tion will be shared. will concepts that will allow you to understand net- enable health care institutions to attain the benefits work technology. tions around the world. we introduce the important Software for Network Communications). An introduc. The systems focus on use of DBMS differently. specialized applications and of large integrated databases. Most medical applications use stan. communications make it possible to share data Application-specific descriptions of a database are and resources among diverse users and institu- stored in such view schemas.ketabpezeshki. 5. At the same time. tion of multiple individual computer systems. be periodically queries.1 The Network Stack Database design and implementation has Network power is realized by means of a large become a highly specialized field. the internal preparation of the organization and use of databases in health data to be sent or received over the . the interfaces between the network data flow and dard products from established vendors. Internet hidden from groups that do not need them. one patient machine that service network communications. often at predetermined users doing simple repetitive queries. Wiederhold’s book (1981) discusses puter to the network. (2002). for report-generation programs can extract header querying across multiple patients simultaneously. the differing needs of multiple users. Data ware- intervals. but only the data relevant to the individual application area The ability of computers to communicate with are available to each group.5. it accommodates about the biomolecular basis of life and disease). these are called on line transac. Through the views. Thus. the existence of other data is tremendous power to computer users. these architectural goals. In this section. at a time. hospital information Reports that are not read are a waste of computer. both of independent. The data stored in a database have from a single source. For example. of the data.5. combined with communica. These strategies (021) 66485438 66485457 www. especially when both systems use data derived els. communicate with one another and to collabo- tions technology (see the following discussion on rate. Many DBMSs support multiple views.5 Software for Network clinical laboratory and the finance department Communications might use the same underlying database. the 5. Routine report typically by few users for infrequent. This software tion to the topic is provided by Garcia-Molina handles the physical connection of each com- et al. body of communications software. a single physical organization. The reports are formatted such that they housing or on line analytic processing (OLAP) can be distributed without modification. Thus will be able to provide needed and up-to-date computer architectures may require the coordina- information when that information is required. There are now tens of these databases and application architectures are millions of computers of different kinds on the inherently oriented toward the transactions needed Internet and hundreds of programs in each for the workflows of the applications. however. The use of Networks make it possible for remote users to database technology. Network users can access a DBMS controls access to data. information from the schema. yet different user groups can have different perspectives on the con- tents and structure of a database. or mod.

FTP File Transfer Protocol. The network stack serves to organize with only the layers directly above and below it and communications software within a . and make it possible that support electronic mail.172 J. Each level of the stack specifies a pro.ketabpezeshki. The four-level network stack for TCP/ put. topology of the network changes. The Network software is made modular by dividing network stack is machine. how to detect and correct errors. SMTP Simple Mail Transport sponding levels of the Open Systems Interconnection (OSI) Protocol. ensure that appli. Datagram Protocol. computer. and to manage the particular network hardware on the the means for detecting and correcting errors. are unaffected. Silverstein and I.10.3) are shared Application level. memory access. Applications that stack to the seven-level stack defined by the see a standard set of data-communication services International Standards Organization. Each layer communicates vices.10 TCP/IP network service level stack and corre. 5.and OS-dependent— the responsibilities for network communications because it has to run on particular hardware and to into different levels. UDP User Organization (ISO). cation programs are insulated from changes in The Application level is where programs run the network infrastructure. routing packets. and do not each have to worry about details such as At the lowest level—the Data Link and how to form proper packets of an acceptable size Physical Transport level—programs manage the for the network. IP Internet Protocol. a transaction-oriented User nications among machines—much as.). or if the addressing packets. But its layered design IP is shown in Fig. RARP Reverse Address Resolution Protocol allow communication to take place between any Datagram Protocol (UDP). DNS Domain Name Reference model developed by the International Standards System. and newer services two machines on the Internet. and con. such as real-time video. 5. Only the lower level Data Link and sions. file sharing and trans- for users to take advantage easily of the rapidly fer. Each level serves the Control Message Protocol. the applications trolling the timing and sequencing of transmis.C. from the level below it. The Transport level converts packet-level Network layers need to be updated. 5. many other services. and growing set of information resources and ser. If a computer changes its network con- The Network level implements the IP method of nection from a wired to a wireless network. Web posting. Foster Fig. TCP Transmission Control Protocol. machine. etc. which also compares serves the function of modularization. ARP Address Resolution level above and expects particular functions or services Protocol. browsing. how to route packets to the desired physical connection of the machine to the net. downloading. or how work. input–out- the levels. communications into several services for the Internet protocols(see Sect.T. does so through specific interface conventions. for two (021) 66485438 66485457 www. the physical-medium packet formats. with clear interfaces between deal with the OS on that machine (filing. ICMP Internet gressively higher level of abstraction. including a reliable serial conventions that serve to standardize commu- byte stream (TCP).

each layer communicates with only its tations or announcements to broad lists.2 Electronic Mail information is not secure in transit. but electronic mail etiquette conventions dictate within a machine. sound. and access to the World Wide Web. but is difficult to prevent. to exchange rout. encrypting attachments. and other optional header content (Multipurpose Internet Mail Extensions lines. is peer layer on the other machine. tion from providers to patients is potentially in violation of the HIPAA regulations in that the 5. stack is also supported by protocols. the subject line. Protocols are defined 1970. the attachment of multimedia ents. Spamming. FTP. however. and many more. Because secure email is generally not in available on the Internet: electronic mail. given that the patient has begun the (021) 66485438 66485457 www. JohnSmith@ Language (HTML)).5. pictures. Between vant. If the body of the e-mail message information. The body of the message contains free text. The conventions for representing data. and for replying to a requested to the RFC822 protocol. the can read its contents. which is sending e-mail solici- machines. As we said. By observing these is encoded according to the MIME standard it protocols. protocol. The messages travel rapidly: in email.5 Computer Architectures for Health Care and Biomedicine 173 people to communicate effectively. For example. other. The layering of the network ing it to a mailing list or a specific list-server. contents are readable by only the intended recipi- We briefly describe four of the basic services ents. A simple e-mail tronic mail. the address or domain name of the machine on which formatting of Web pages (Hypertext Markup the reader receives mail—e. using a defined annoying to recipients. and so on. Protocols for encrypted Network layer communicates with only peer e-mail. addresses of the recipi- sages (RFC822). a Windows machine. machines of different types can com. header contains information formatted according ing an action. For example. which controls the action. are also available. may also contain arbitrary multimedia informa- municate openly and can interoperate with each . such as drawings. their transmission is nearly they are using. if it is appropriate for physi- users via electronic mail. such as Privacy-Enhanced Mail (PEM) or Network layers. the SMTP application on Conventional e-mail is sent in clear text over the one machine communicates with only an SMTP network so that anyone observing network traffic application on a remote machine. each layer communicates with that such communications be focused and rele- only the layer directly above or below.g. the address of the sender. elec.ketabpezeshki. name or a personal alias followed by the IP connections to remote computers (SSH). downloaded to the addressee’s local computer for or a mainframe running VMS—they all appear reading. for request. conventions for e-mail addresses and text mes.. still called the ARPANET). the intended reader by giving the reader’s account transfer of files (File Transfer Protocol (FTP)). or video. Electronic mail was one of the versus conversation). for example. the delivery of e-mail messages The user addresses the e-mail directly to the (Simple Mail Transport Protocol (SMTP)). and Web access) and establish the message consists of a header and a body. the exchange of routing domain. communication of protected health informa- SSH. It remains Users send to and receive messages from other less clear. use. They ensure that the Internet Control Message Protocol (ICMP). the style of the interaction (lecture instantaneous. protocols define the format appearance of the date and time of the message. a procedure for handling first protocols invented for the Internet (around interruptions. It may either be read on the machine hold- the client does not have to know whether the ing the addressee’s account or it may be server is a UNIX machine. they must except for queuing delays at gateways and receiv- agree on the syntax and meaning of the words ing computers. When requesting a Web page from a server Mail is sent to the recipient using the SMTP stan- using the Hypertext Transfer Protocol (HTTP). but ing information or control information using the are not yet widely deployed. (MIME)). when what was to become the Internet was for every Internet service (such as routing. the same over the network if they adhere to the It is easy to broadcast electronic mail by send- HTTP protocol. mimicking use of the cians to answer direct questions from patients postal service. Similarly.

and the name 5.ketabpezeshki.T. Globus Online is a SaaS data movement communications can be encrypted to protect sen- solution (see Sect. Secure Shell allows a user to log in on a remote TLS) which is used by the HTTPS protocol (and computer securely over unsecured networks using generally shows a “lock” icon when the browser public-key encryption (discussed in next section). medical records. The Universal Resource Locator must authenticate to the same system) to over.g.3). font settings. papers with many figures or and other display specifications. distinct from the Java lan- Web browsing facilitates user access to remote guage (unfortunately similarly named) runs in the information resources made available by Web browser itself (much like protected memory). documents. Secure Shell enables complete com. called applets.5. thus (021) 66485438 66485457 www. (2) providing user iden. the user identification browser has been augmented with helpers or is by convention “anonymous” and the requestor’s plug-ins for the particular format used.4) that provides both secu. 5. (URL) is used to specify where a resource is come this. video. ing remote and untested software still represents a substantial security risk (see Sect. images.5. headings. and the HTML documents can also include small pro- user’s own machine becomes a relatively passive grams written in the Java language. The smoothness of such a which will execute on the user’s computer when terminal emulation varies depending on the dif. but download- extent the user’s account is authorized.5. The 5.4 SSH recently renamed to Transport Level . HTML supports con- programs and updates to programs.g. look like when displayed.7. but provides poor perfor. and (4) information—e. The e-mail address is used as the password.3 File Transfer Protocol (FTP) of the information resource within the remote FTP facilitates sending and retrieving large machine. patient information) from external view using protocols such as Secure Sockets Layer (SSL: 5. Silverstein and I.5.5 World Wide Web (WWW) JavaScript language. For open sharing of informa.5. Large health systems servers. and the like could be trans.. Foster insecure communication. complete ventional text. speech— that can be seen or heard if the tion by means of FTP sites. (1) accessing the remote computer using the IP This hypertext facility makes it possible to create address or domain name. Applets can provide animations and ferences between the local and remote computers. Secure Hyper Text Transfer Protocol (HTTP) is used FTP (SFTP) uses the same robust security mech. located in terms of the protocol to be used. (3) specifying the gated by the user.174 J. the domain name of the machine it is on.5. graphics. the user becomes a URL). also can compute summaries. to communicate between browser clients and anism as SSH (below).C. The Hyper Text Markup Language amounts of information—of a size that is uncom. merge information. 5. lists. The user interface is typically a Web have generally deployed secure communication browser that understands the World Wide Web portals over the Web (where both participants protocols. terminal in this context. (HTML) describes what the information should fortably large for electronic mail. highlighted buttons can be defined ferred via FTP. Such mance. a web of cross-referenced works that can be navi- tification to authorize access. tables. iary documents that contain other types of naming convention at the destination site. equations. sitive contents (e. credit card information or rity and high performance.. For instance. is securely communicating with the host in the If the log-in is successful. FTP access requires several steps: that point to other HTML documents or services. HTML can also refer to subsid- name of a file to be sent or fetched using the file. referenced. The Java language is designed such that important. Within HTML images for review. operations that might be destructive to the user’s mand line control of the remote system to the machine environment are blocked. transferring the data. Secure Shell replaced Telnet which was used for and interact with selected files on the user’s com- terminal emulation until network security became puter. fully qualified user of the remote system. servers and to retrieve HTML documents.

emo. cuts. sexually transmitted diseases. 10. by affecting our insurability. release of his or her personal health information to ent—generates requests and does complementary a care provider or information custodian under an processing (such as displaying HTML documents agreement that limits the further release or use of and images). the ability of a person to control the some protocol. and is becoming immunodeficiency virus (HIV) status. In general. health information is con- markup related to internal links. and safeguards. Security is the protection of pri- servers is database access. A common function provided by that information. but health Medical records contain much information about care systems are distinguished by having especially us. and notes regarding bouts with a health care information system serve five key the flu. or by lim- work toward more open. perimeter con- tional problems and psychiatric care. called eXtensible Markup Language to depend on electronic medical records for care. As the appearance of the document on a display to discussed in Chap. self-defining document iting our ability to get and hold a job. must be accurate and up to date.6 Data and System Security Concerns about. substance increasingly powerful. and to control access to these systems’ contents. they must be available whenever and wherever we XML further enhances the capabilities of the need care. accountability. The server provides infor. and methods to provide. secu- rity are part of most computer systems. enhancing substantially abuse. data also must be protected against loss. Orders for tests or treatments must be validated to ensure that they 5. Medical descriptions. 10. for issues from other types of markup. and access to such information must be semantic features. a more powerful markup frame. information. Confidentiality applies to information—in mation and computational services according to this context.6 Client-Server Interactions are issued by authorized providers. and to prejudice. (XML). Retrieved information vacy and confidentiality through a collection of is transferred to the client in response to requests. be mundane. Coupled with JavaScript. supports Privacy refers to the desire of a person to con- collaboration between the user of a local machine trol disclosure of personal health and other infor- and a remote computer. physical abuse. scripts.5 Computer Architectures for Health Care and Biomedicine 175 reducing this security risk. or mailed to other users. requirements touch on three separate concepts ing) machine and a server (responding) machine.ketabpezeshki. procedures. human . and the user’s computer—the cli. A client–server interaction. involved in protecting health care information. If we are work. These involving interactions between a client (request. The final results can be stored integrity and availability of information systems locally. measures enable an organization to maintain the yses on the data. about topics such as fertility and abortions. 5. 1997). and other fidential. (021) 66485438 66485457 www. by causing social embarrassment or more flexibility in terms of markup types. Some data are generally considered to as a computer platform. the security steps taken in blood pressures.5. and comprehensibility and behaviors. policies. has emerged. We discuss each of these functions in turn. sexual trol. namely availability.5. Within the HTML captures many aspects of document medical record. To separate appearance-related controlled because disclosure could harm us. in general. role-limited access. printed. and the information that they contain Web browser itself as a computer platform. Health privacy and confidentiality are discussed further in Chap. others highly sensitive. These documents and databases include data complex considerations for the use and release of ranging from height and weight measurements. The records A client–server interaction is a generalization of must also support administrative review and pro- the four interactions we have just discussed. and genetic predisposition the capabilities and capacity of the Web browser to diseases. to provide example. vide a basis for legal accountability. or broken bones to information functions (National Research Council. Security and then the client may perform specialized anal. there is much information about description from predefined markup related to which any given person may feel sensitive. mation.

Silverstein and I. a word-processing fectly . viruses. Technical means to ensure accountability A virus may be attached to an innocuous program include two additional functions: authentication or data file. Therefore.g. Backup copies sible for their access to. and so on are also subject to infection by systems must be prepared to operate even during viruses. It helps to ensure that users are respon- software to ensure availability. knowledge that tors to roll back to the earlier version of the soft. Authentication and authorization can be per- 5. watched. For short-term protection. are highly ethical. to 2.1 Availability and by producing large amounts of Internet traf- fic as it repeatedly distributes itself (what is called Availability ensures that accurate and up-to-date a denial of service attack). if a new version of a program damages Most people working in a medical environment the system’s database. to search through certain medi- 4. appropriate to his or her role in the health care ther distribute its copies. Such a macro can also at remote sites to protect the data in case of disas.C. several actions take place: 1. data records are being ware and database contents. however.ketabpezeshki. in the trails. In addition.g. form of viruses or worms—are also a threat. Critical medical tations. The authenticated user is authorized within spread itself to other computers. The viral code copies itself into other files and unique identification process. often to escape care. The user is authenticated through a positive 1. environmental disasters. detection. For example. information of long-term value is document may include macros that help in for- copied onto archival storage. A program installed by a virus may record formed most easily within an individual com- keystrokes with passwords or other sensitive puter system. through scanning of access audit Unauthorized software changes—e. It is primarily achieved to protect should be checked against known viral codes and against loss of data by ensuring redundancy— for unexpected changes in size or configuration. coordinate these access controls consistently A software virus causes havoc with computer across all the systems. Accountability for use of medical data can be network equipment. access to. (021) 66485438 66485457 www. etc. To protect against information is available when needed at appro. and use of. Enterprise-wide and operations. information provide a degree of protection against software based on a documented need and right to know. graphical presen- on duplicate storage devices.. all programs loaded onto the system priate places. It attaches these files to outgoing messages.176 J. when that program is executed and authorization. remote access-control standards and systems are age. so the document ter. such as residing in the computer. performing regular system backups. and use of. storage devices. name and password combination. been imported.T. and copies are kept matting the document. or data file is opened. by disturbing operations and system access available and are being deployed extensively. include viral codes. 2. addi- tional protection is conferred. because most institutions information. It is also essential to promoted both by surveillance and by technical maintain the integrity of the information system controls.2 Accountability important to provide secure housing and alterna- tive power sources for CPUs. system—e. and. even if it does not do disabling dam. The virus may collect email addresses to fur. Foster 5. The virus may install other programs to cal records of only patients under his or her destroy or modify other files. it is necessary to actions. Because It is not always obvious that a virus program has hardware and software systems will never be per. it is also 5. If one of the storage devices is attached to a remote processor. but. failures.6. data can be written can be infected. or perform other deleterious operate multiple computers. Spreadsheets.. the backups allow opera.6. serves as a strong impediment to abuse. the system to perform only certain actions 3.

3 Perimeter Definition so that their contents are accessible by unauthor- ized people. One is work runs and how users get outside access is kept secret. single-layer gies offer a powerful way to let bona fide users encryption with keys of 56-bit length (the length access information resources remotely. It is thus asymmetric and one end of to define the boundary: all sharable computers of the transaction can be proven to have been done the institution are located within the firewall. greater care must be taken to cryptography. Two kinds of cryptog- word combinations. For example. and keys of 128 and even 256 bits are machine and an enterprise server. receiver. Using a prescribed by the 1975 Data Encryption client–server approach. Most organizations install a firewall available. your users are and how they are accessing the Health systems are therefore requiring whole information system. one to encrypt trusted insider? Careful control of where the net. For trols depend on cryptographic technologies. Strong authentication and authorization con- mation system. over communication lines. Having passed this information. By using this twice (four Anyone who attempts to access a shared system keys). prise system administrators to monitor and ensure Under either scheme. If a key is lost. it is important for enter. physical access can be provided with a protecting data that are stored or are transmitted minimum of hassle using simple name and pass. where strong authentication and protocol services in addition to the exchange of sensitive access controls are in place. or if too many copies of secure the connection is. that increasingly portable computers (laptops. a key is needed to decode and make the informa- malicious intruder could install a new virus tion legible and suitable for processing. No matter how well. health care providers within a small physician Cryptographic encoding is a primary tool for practice. however. Even and nonrepudiation (to ensure that an action can- with a firewall in place. This arrangement leads to important wall. the information protects all communications and uses strong encrypted with the key is effectively lost as authentication to identify the user. (021) 66485438 66485457 www. both physically and logically. known to only the sender and information are not observed inappropriately. the key must be and that communications containing sensitive be kept secret. The other one can be made publicly necessary. unauthor- ultimately depends on responsible users and care ized people may gain access to information. an encrypted communi. by a specific entity. however. not be denied as having been done by the actor). content validation (to prove that the enabled services within the firewall (still limited contents of a message have not been changed). a that the firewall is not bypassed. sound security the key exist for them to be tracked. once data are encrypted. the user can then access authorship). the same key is used to encrypt ensure that the person is who he or she claims to and to decrypt information. the information and a second to decrypt it. This approach routine. In secret-key record. But intended receiver of information. such as digital signatures (to certify authentication step. machine encryption on portable devices. Because pow- software. two keys are used.5 Computer Architectures for Health Care and Biomedicine 177 5. If a clinician is traveling or raphy are in common use— secret-key cryptogra- at home and needs remote access to a medical phy and public-key cryptography. Thus. lions of candidate keys rapidly. If a key is stolen. It allows the system to con. install or use surreptitiously Keys of longer length provide more security. This is the single most common Perimeter definition requires that you know who mechanism by which HIPAA violations occur. Standard (DES)) are no longer considered cation link is negotiated between the user’s client secure.6. or load unauthorized because they are harder to guess. In public-key where is the boundary for being considered a . within the perimeter. one can certify both the sender and from the outside must first pass through the fire. or handheld devices) are not lost or stolen can provide some protection against loss.ketabpezeshki. by the applicable authorization controls). a wireless base station. trol the boundaries of trusted access to an infor. erful computers can help intruders to test mil- Virtual Private Network (VPN) technolo. Holding the keys in escrow by a trusted party tablets.

C.6. An audit trail contains records indicating what cific data within the record. It allows access for per. porting clinical documentation as required by tion may be encoded as digital certificates. and the IRB. ers and data stewards to understand whether data sonnel to only that information essential to the per. A central authentication or authorization server must itself be guarded and managed with extreme care. From a technological perspective. partitioned according to access criteria based upon and if possible some indication of the reason for use privileges. and reduces the possibil.T. or by As noted in Sect. Examples of can both provide a strong disincentive for valid access privileges include the following: improper access (because individuals will know • Patients: the contents of their own medical that accesses will be recorded) and allow respon- records sible parties to detect inadequate controls.ketabpezeshki. protected against improper deletion. A medical record is data was accessed. allows for rapid revocation ability Act (HIPAA) and as set by institutional of privileges as needed. which insurance companies then can be validated with a certification author. Many formance of their jobs and limits the real or hospitals also allow employees to review who has perceived temptation to access information beyond accessed their personal records. Centralizing authentication and mined by state and federal law such as provisions authorization functions simplifies the coordina. of the Health Insurance Portability and Account- tion of access control. • Insurance payers: justifications of charges ity and checked by the services so that the ser. The authorization informa. such records tion collected in the medical record. the many different collaborators in the access.5 Comprehensibility and information is not broken.6. If backed by strong authentication and health care all have diverse needs for the informa. policy following legal and ethical considerations ity of an intruder finding holes in the system. computer networks have waiver of authorization for patient groups become fundamental to both health care and bio- approved by an Institutional Review Board medical research.178 J. understand and have effective control over appro- priate aspects of information confidentiality and access. however. with sup- authorization as well. and patients can locked browser icon discussed earlier). . fur. by who. is being accessed correctly or incorrectly. as deter- themselves. so that the chain of access-control 5. when. Different types of information kept in medical vices do not need to check the authorizations records have different rules for release. • Community physicians: records of their patients • Specialty physicians: records of patients referred for consultations 5. Control tain the public certificates for major certificate authorities thus enabling them to check the verac. ther checks must be made to control access to spe. importance of patient privacy.5.4 Role-Limited Access authentication and access control are important control mechanisms. Architectures nicable diseases • Medical researchers: consented data.5. while 5.7 Distributed System • Public health agencies: incidences of commu. Even when overall system reassures employees and builds awareness of the access has been authorized and is protected. Comprehensibility and Control ensures that ity of Web sites using HTTPS (enabling the record owners. allowing for quasi-instantaneous (IRB) information sharing and communication within (021) 66485438 66485457 www. from where. data stewards. audit trails are perhaps the Role-limited access control is based on extensions most important means for allowing record own- of authorization schemes. Modern browsers con. Silverstein and I. This ability both a legitimate need. Foster Cryptographic tools can be used to control • Billing clerks: records of services.

information representations. policies. tion over many decades.ketabpezeshki.. with dif- within a single institution (e. ing access to computers located within its fire- ent hospitals). it must be possi- which despite much effort towards standardiza.4). The systems to be linked may live management systems (see Sect. in a clinical trial. exist. and in the absence of strong external control and/ or incentives. email. different meanings assigned to a clinical Distributed systems can be challenging to diagnosis). and Programming governance structures. 5. we see increased interest in not just To illustrate the types of problems that may enabling ad hoc access to remote data. mentation. 5. first of all. restrict- tronic medical record systems operated by differ. creation of large-scale distributed systems that As health care becomes more information span many institutions remains difficult. and a medical PACS system oper. • Hospitals HA and HB both have a governance erated queries (see Sect. seman- backup of its databases in a cloud provider. and governance.. videocon. 5. as a single infor. require solutions to a wide range of heterogeneity source expensive tasks to third-party providers. ble for a program running on one computer to (021) 66485438 66485457 www. A distributed system links multiple medical records maintained in databases DA and computer systems in such a way that they can DB at two hospitals. They are. and SSH for remote access to comput. so as to permit fed. The use of construct and operate.5) service provider). but in link.g. with personnel who are not HB employees to federate database systems. ers to the same patient ated by the Department of Radiology). We may find that: function. including syntactic ( . a vocabularies to represent patient information pathology database operated by the Department • Hospitals HA and HB assign different identifi- of Pathology. To build a distributed system.1 Distributed System tions. to some extent at least. appropriate distributed system technologies can worked systems. the with physical separation.7. but any to a wide range of failures.7. non-standards-based products. • Databases DA and DB use different database mation system. They ing information about its patients to be shared may be linked for a range of reasons: for example. Yet more seriously. to find records corresponding to computer and data systems ahead of imple- multiple admissions for a single patient at differ. HA and HB.4) different types (e. as problems. the devel- workflows. policy. This situation can ferencing. 5. • Databases DA and DB use different schemas and mation technology services organization. make even point-to-point integration of two sys- ers are all widely used to reduce barriers associated tems within a single hospital challenging. an electronic med. different for- when a medical center maintains an off-site mats used to represent data in databases). ferent access protocols and query languages ical record system operated by a hospital’s infor.g.. or to identify potential participants from the developers of the distributed system.g. to enable cross-institutional As this brief and partial list shows.. net.g. or within institutions of quite wall (see Sect..g. effective and robust solution must take into they commonly span administrative boundaries..5 Computer Architectures for Health Care and Biomedicine 179 and between clinical and research institutions. such as routing of infectious disease opment of a successful distributed system can data to a public health organization. consider a distributed system that is ing diverse information systems into distributed intended to provide a unified view of electronic systems. 5.g.7. or to out. tic (e. within • Hospital HA does not allow remote access to multiple institutions of similar type (e. These problems seem to be particularly prevalent within health care. but neither includes representation ent hospitals. different administrative units tend to adopt distinct and incompatible technical solu. account all aspects of the problem. and thus are inevitably subject assist with overcoming these problems.2) across different structure that reviews proposed changes to databases—e. elec. its electronic medical record system. driven. remains dominated by Applications such as the Web. a hospital and a third-party • Hospital HB interprets HIPAA rules as prevent- cloud (see Sect.

adoption within procedure that is to be called (e.g. /patients/{patient-id} tion and medical image access.5.5) to com- to hide heterogeneity across different platforms municate them.T. a ple. a brief description inter-company conflict (e. REpresentational State Transfer (REST). Foster issue a request that results in a program being lar concepts to those found in CORBA. where {patient-id} is the patient CORBAmed activity defined a range of specifi. GET or POST/patients/ Get/put profile for a ming technology. lab results specific patient (021) 66485438 66485457 www.5. communication typically occurs via the Internet including biomedical research. This approach is based on to discover what procedures are supported by a a small set of protocol standards and methodolo- particular remote system. and Amazon — has largely converged on about how the remote program is implemented.5) to encode remote pro- uted system programming problem. Table 5. while CORBA has had some success in certain Table 5. seeing increasingly widespread use in health care One distributed system programming as well. the programmer needing to know anything Twitter. Under the covers. for security credentials. These same methods are ular purpose. “find patients with influenza”). distributed research datasets and analytic procedures being system programming methods build on that made accessible over the network via Web base to allow a programmer to name a remote Service interfaces. specify Over the past 5 years. the following form.. due to their simplicity and the substantial approach on which much effort has been spent investment in relevant technologies occurring is CORBA. Facebook. Web Services have been pursued over the years to this distrib.5. TLS.. an numerous technology providers and adopters HTTP-based approach to distributed system formed the Object Management Group (OMG) architecture in which components are modeled as to define standards for describing remotely resources that are named by URLs and with accessible procedures (an Interface Definition which interaction occurs through standard HTTP Language: IDL) and for invoking those proce. the format of each request.0.g. On the left. a yet simpler architectural approach for defining Other related methods may allow a programmer interfaces to services. PUT. provide required consumer Web/Cloud market — as typified. manufacturing). These procedure call mechanism. Microsoft never Request Description adopted CORBA) had prevented it from having GET/patients Retrieve list of patients broad impact as a distributed system program. clinical systems remains modest for reasons base”). JSON and XML.3 illustrates a simple REST encoding wide range of other standards have been devel. and defines an IDL called Web via the introduction of a standardized remote Services Description Language (WSDL). In the field of health care. by the likes of Google. a new technology called GET/patients/{patient-id}/ Request lab results for a Web Services emerged that implemented simi.g. For exam- dures over a network. 5. of an interface to a medical record system. Various approaches simpler and more flexible form.. but in a run in another computer.5. the Common Object Request within industry. or alternatively what gies. All seek cedure calls and HTTP (Sect. Building on that base.3 Example REST representation of a patient industries (e.180 J.5. uses XML (Sect. and so forth—all without example. Unfortunately. specify arguments to that procedure listed above. namely REST and HTTP. However. Starting in the mid 1990s. identifier: cations for such things as personal identifica. 5.ketabpezeshki.g. of technical limitations in its specifications and on the right. a combination record interface. remote system should be contacted for a partic. and O Auth 2. the commercial and how results are to be returned. technologies have seen wide use in many areas.. with many protocols described in Sect. actions (POST. ( . At the core of this approach is Broker Architecture. 5. DELETE). “query data. GET.5. {patient-id} specific patient In the mid 2000s. This oped defining interaction patterns important for interface models patients as resources with the different fields. Silverstein and I.

5 Computer Architectures for Health Care and Biomedicine 181

Thus, to request all patients that we have per- These two approaches have various advan-
mission to see we send the following HTTP tages and disadvantages. The data warehouse
request: requires a potentially expensive ETL process,
GET/patients/ requires storage for a separate copy of all relevant
while to obtain the profile for a specific patient data, and may not be up to date with all source
named NAME we send this request: databases. However, it can permit highly efficient
GET/patients/NAME queries against the entirety of the data. The feder-
and to update that profile we send the request: ated query approach can provide access to the lat-
POST/patients/NAME est data from each source database, but requires
REST and HTTP define how to name resources potentially complex mediator technology.
(URLs) and messaging semantics, but not how to
encode message contents. Two primary message
encoding schemes are used: Java Script Object 5.7.3 Parallel Computing
Notation (JSON) and XML. The following is a
potential JSON encoding of a response to a GET/ Parallel computers combine many microproces-
patients/request. The response provides a list of sors and/or GPUs (see Sect. 5.2) to provide an
patient identifiers plus (because there may be many aggregate computing capacity greater than that of
more patients that can fit in a single response) a single workstation. The largest such systems
information regarding the number of patients available today have more than one million pro-
included in this response (X), an offset that can be cessing cores. While systems of that scale are not
provided in a subsequent request to get new patients used in medicine, parallel computers are becom-
(Y), and the number of patients remaining (Z): ing more commonly used in biomedical comput-
{"patients": ing as a means of performing large-scale
{"list":[list of patients], computational simulations and/or analyzing large
"count":X, quantities of data. In basic research, parallel
"offset":Y, computers are used for such purposes as mining
"remain":Z} clinical records, genome sequence analysis, pro-
} tein folding, simulation studies of cell mem-
branes, and modeling of blood flow. In
translational research, parallel computers are
5.7.2 Distributed Databases commonly used to compute parameters for com-
puter programs that are then used in clinical set-
Distributed databases are a special case of a dis- tings, for example, for computed aided diagnosis
tributed system. The problem here is to enable of mammograms.
queries against data located in multiple databases. When computing over large quantities of
Two different methods are commonly used. In a data, it is often useful to employ a parallel
data warehouse approach, an extract-transform- database management system. These systems
load (ETL) process is used to extract data from support the same SQL query language as
the various sources, transform it as required to fit sequential database management systems (see
the schema and semantics used by the data ware- Sect. 5.5.4.) but can run queries faster when
house, and then load the transformed data into the using multiple processors. Another increasingly
data warehouse. In a federated query approach, a popular approach to parallel data analysis is to
query is dispatched to the different databases, use the MapReduce model, popularized by
applied to each of them independently, and then Google and widely available via the free
the results combined to get the complete answer. Hadoop software. MapReduce programs may
Intermediate components called mediators may be less efficient than equivalent SQL programs
be used to convert between different syntaxes and but do not require that data be loaded into a
semantics used in different systems. database prior to processing.

(021) 66485438 66485457

182 J.C. Silverstein and I.T. Foster

5.7.4 Grid Computing high-speed Internet, vastly increased demand
from e-commerce, powerful lightweight Web
Grid computing technologies allow for the fed- protocols, and an effective business model—
eration of many computers and/or data resources cloud computing has achieved large-scale adop-
in such a way that they can be used in an inte- tion in ways that previous efforts had not. The
grated manner. Grid computing is the foundation, implications of these developments for medical
for example, for the worldwide distributed sys- informatics will surely be profound.
tem that analyzes data from the Large Hadron The National Institutes of Standards and
Collider (LHC) in Geneva, Switzerland. The 10 Technology (NIST) defines cloud computing as
or more petabytes (10 × 1015 bytes) produced per “a model for enabling ubiquitous, convenient,
year at the LHC is distributed to several hundred on-demand network access to a shared pool of
institutions worldwide for analysis. Each institu- configurable computing resources (e.g., net-
tion that participates in this worldwide system works, servers, storage, applications, and ser-
has its own local computer system administration vices) that can be rapidly provisioned and
team, user authentication system, accounting sys- released with minimal management effort or ser-
tem, and so forth. Grid technologies bridge these vice provider interaction” (Mell and Grance,
institutional barriers, allowing a user to authenti- 2011). They distinguish between three distinct
cate once (to “the grid”), and then submit jobs for types of cloud service (see Fig. 5.11):
execution at any or all computers in the grid. • Software as a Service (SaaS) allows the con-
Grid computing is used in many academic sumer to use the provider’s applications run-
campuses to link small and large computer clus- ning on a cloud infrastructure. Examples
ters and even idle desktop computers for parallel include Google mail, Google Docs, Salesforce.
computing applications. But its biggest use in com customer relationship management, and a
research is to enable sharing of large quantities of growing number of electronic medical record
data across institutional boundaries. By address- and practice management systems.
ing the challenges of authentication, access con- • Platform as a Service (PaaS) allows the con-
trol, and high-speed data movement, grid sumer to deploy consumer-created or acquired
computing technologies make it possible, for applications onto the cloud infrastructure, cre-
example, to acquire genome sequence from a ated using programming languages, libraries,
commercial sequencing provider, transport that services, and tools supported by the provider.
data over the network to a cloud computing pro- Google’s App Engine and Salesforce’s Force.
vider (see Sect. 5.7.5), perform analysis there, com are examples of such platforms.
and then load results into a database at a research- • Infrastructure as a Service (IaaS) allows the
er’s home institution. consumer to provision processing, storage,
networks, and other fundamental computing
resources on which the consumer is able to
5.7.5 Cloud Computing deploy and run arbitrary software. Amazon
Web Services and Microsoft Azure are popu-
The late 2000s saw the emergence of success- lar IaaS providers.
ful commercial providers of on-demand com- Each level of the stack can, and often does, build
puting and software services. This concept is on services provided by the level below. For exam-
certainly not new: for example, McCarthy first ple, Google Mail is a SaaS service that runs on
referred to “utility computing” in 1961, various compute and storage infrastructure services oper-
time sharing services provided computing over ated by Google; Globus Online is a SaaS research
the network in the 1970s and 1980s, and grid data management service that runs on infrastruc-
computing provided such services in the 1990s ture services operated by Amazon Web Services.
and 2000s. However, it is clear that—perhaps NIST further distinguishes between public
driven by a combination of quasi-ubiquitous cloud providers, which deliver such capabilities

(021) 66485438 66485457

5 Computer Architectures for Health Care and Biomedicine 183

growing number of both commercial and non-
profit SaaS offerings designed to accelerate com-
mon research tasks. For example, Mendeley
organizes bibliographic information, while
Globus Online provides research data manage-
ment services.
Similarly, in health care, we see many inde-
pendent physicians and smaller practices adopt-
ing SaaS electronic medical record systems. The
relatively high costs and specialized expertise
required to operate in-house systems, plus a per-
ception that SaaS providers do a good job of
addressing usability and security concerns, seem
to be major drivers of adoption. Similarly, a
growing number of biomedical researchers are
using IaaS for data- and compute-intensive
Fig. 5.11 The NIST taxonomy of cloud providers. SaaS, research. Meanwhile, some cloud providers (e.g.,
PaaS, and IaaS providers each offer different types of ser- Microsoft) are prepared to adhere to security and
vices to their clients. Cloud services are distinguished by privacy provisions defined in HIPAA and the
their Web 2.0 interfaces, which can be accessed either via
Web browsers or (for access from other programs or HITECH act. Nevertheless, while some hospitals
scripts) via simple APIs are using IaaS for remote backup (e.g., by storing
encrypted database dumps), there is not yet any
significant move to outsource major hospital
to anyone, and private cloud providers, which information systems.
provide such on-demand services for consumers
within an organization.
Benefits claimed for cloud computing include 5.8 Summary
increased reliability, higher usability, and reduced
cost relative to equivalent software deployed and As we have discussed in this chapter, the synthe-
operated within the consumer’s organization, due sis of large-scale information systems is accom-
to expert operations and economies of scale. plished through the careful construction of
(IaaS providers such as Amazon charge for com- hierarchies of hardware and software. Each suc-
puting and storage on a per-usage basis.) Potential cessive layer is more abstract and hides many of
drawbacks include security challenges associated the details of the preceding layer. Simple meth-
with remote operations, lock-in to a remote cloud ods for storing and manipulating data ultimately
provider, and potentially higher costs if usage produce complex information systems that have
becomes large. powerful capabilities. Communication links that
Outside health care, cloud computing has connect local and remote computers in arbitrary
proven particularly popular among smaller busi- configurations, and the security mechanisms that
nesses, who find that they can outsource essen- span these systems, transcend the basic hardware
tially all routine information technology functions and software hierarchies. Thus, without worrying
(e.g., email, Web presence, accounting, billing, about the technical details, users can access a
customer relationship management) to SaaS pro- wealth of computational resources and can per-
viders. Many companies also make considerable form complex information management tasks,
use of IaaS from the likes of Amazon Web such as storing and retrieving, communicating,
Services and Microsoft Azure for compute- and authorizing, and processing information. As the
data-intensive computations that exceed local technology landscape evolves and computer
capacity. In research, we see the emergence of a architectures continuously increase in complexity,

(021) 66485438 66485457

184 J.C. Silverstein and I.T. Foster

they will also increasingly, necessarily, hide that works up to network applications, using real-world
complexity from the user. Therefore, it is para- example networks to illustrate key principles. Covers
applications and services such as email, the domain
mount that systems designers, planners, and name system, the World Wide Web, voice over IP, and
implementers remain sufficiently knowledgeable video conferencing.
about the underlying mechanisms and distin- Teorey, T., Lightstone, S., Nadeau, T., & Jagadish, H.
guishing features of computing architectures so (2011). Database modeling and design: Logical
design (5th ed.). San Francisco: Elsevier. This text
as to make optimal technology choices. provides and excellent and compact coverage of mul-
tiple topics regarding database architectures, including
core concepts, universal modeling language, normal-
Suggested Readings ization, entity-relationship diagrams, SQL, and data
Wiederhold, G., & Clayton, P.D. (1985). Processing bio-
Council, N.R. (1997). For the record: Protecting elec- logical data in real time. M.D. Computing, 2(6),
tronic health information. Washington, DC: National 16–25. This article discusses the principles and prob-
Academy Press. This report documents an extensive lems of acquiring and processing biological data in
study of current security practices in US health care real time. It covers much of the material discussed in
settings and recommends significant changes. It sets the signal-processing section of this chapter and it pro-
guidelines for policies, technical protections, and legal vides more detailed explanations of analog-to-digital
standards for acceptable access to, and use of, health conversion and Fourier analysis.
care information. It is well suited for lay, medical, and
technical readers who are interested in an overview of
this complex topic.
Garcia-Molina, H., Ullman, J.D., & Widom, J.D. (2008). Questions for Discussion
Database systems: The complete book (2nd ed.). 1. What are four considerations in decid-
Englewood Cliffs: Prentice-Hall. The first half of the
book provides in-depth coverage of databases from the
ing whether to keep data in active versus
point of view of the database designer, user, and appli- archival storage?
cation programmer. It covers the latest database stan- 2. Explain how operating systems and
dards SQL:1999, SQL/PSM, SQL/CLI, JDBC, ODL, cloud architectures insulate users from
and XML, with broader coverage of SQL than most
other texts. The second half of the book provides in-
hardware changes.
depth coverage of databases from the point of view of 3. Discuss what characteristics determine
the DBMS implementer. It focuses on storage struc- whether computer clusters or cloud
tures, query processing, and transaction management. architectures are better for scaling a
The book covers the main techniques in these areas
with broader coverage of query optimization than
given computational problem.
most other texts, along with advanced topics including 4. Explain how grid computing facilitates
multidimensional and bitmap indexes, distributed federation of resources.
transactions, and information-integration techniques. 5. Describe the architectural advantages
Hennessy, J.L., & Patterson, D.A. (2011). Computer
architecture: a quantitative approach (5th ed.). San
and disadvantages of different comput-
Francisco: Morgan Kaufmann. This technical book ing environments.
provides an in-depth explanation of the physical and 6. Explain how REST and XML enable
conceptual underpinnings of computer hardware and flexibility, modularity, and scale.
its operation. It is suitable for technically oriented
readers who want to understand the details of com-
7. How can you prevent inappropriate
puter architecture. access to electronic medical record
Mell, P. and Grance, T. (2011). The NIST definition of information? How can you detect that
cloud computing. NIST Special Publication 800–145, such inappropriate access might have
National Institute of Standards and Technology. This
brief document provides a concise and clear definition
of cloud computing. 8. You are asked whether a medical prac-
Tanenbaum, A., & Wetherall, D. (2010). Computer net- tice should outsource its information
works (5th ed.). Englewood Cliffs: Prentice-Hall. The technology functions to third party
heavily revised edition of a classic textbook on com-
puter communications, this book is well organized,
cloud providers. What factors would
clearly written, and easy to understand. It first enter into your recommendation?
describes the physical layer of networking and then

(021) 66485438 66485457

Software Engineering for Health
Care and Biomedicine 6
Adam B. Wilcox, Scott P. Narus,
and David K. Vawdrey

After reading this chapter, you should know the applications and components of health care infor-
answers to these questions: mation systems, and describe how they are used
• What key functions do software applications and applied to support health care delivery. We
perform in health care? give examples of some basic functions that may
• How are the components of the software be performed by health information systems, and
development life cycle applied to health care? discuss important considerations in how the soft-
• What are the trade-offs between purchasing ware may be acquired, implemented and used.
commercial, off-the-shelf systems and devel- This understanding of how a system gets put to
oping custom applications? use in health care settings will help as you read
• What are important considerations in compar- about the various specific applications in the
ing commercial software products? chapters that follow.
• Why do systems in health care, both internally- Health care is an information-intensive field.
developed and commercially-purchased, require Clinicians are constantly collecting, gathering,
continued software development? reviewing, analyzing and communicating infor-
mation from many sources to make decisions.
Humans are complex, and individuals have many
6.1 How Can a Computer System different characteristics that are relevant to health
Help in Health Care? care and that need to be considered in decision-
making. Health care is complex, with a huge
Chapter 5 discusses basic concepts related to body of existing knowledge that is expanding at
computer and communications hardware and ever-increasing rates. Health care information
software. In this chapter, we focus on the software software is intended to facilitate the use of this
information at various points in the delivery pro-
cess. Software defines how data are obtained,
A.B. Wilcox, PhD (*) • S.P. Narus, PhD organized and processed to yield information.
Department of Medical Informatics,
Software, in terms of design, development, acqui-
Intermountain Healthcare,
5171 South Cottonwood St, sition, configuration and maintenance, is there-
Murray, UT 84107, USA fore a major component of the field. Here we
D.K. Vawdrey, PhD
Department of Biomedical Informatics, The authors gratefully acknowledge the co-authors of the
Columbia University, 622 W. 168th Street, VC-5, previous chapter edition titled “System Design and
New York, NY10032, USA Engineering in Health Care,” GioWiederhold and Edward
e-mail: H. Shortliffe.

E.H. Shortliffe, J.J. Cimino (eds.), Biomedical Informatics, 185
DOI 10.1007/978-1-4471-4474-8_6, © Springer-Verlag London 2014

(021) 66485438 66485457

186 A.B. Wilcox et al.

provide an introduction to the practical consider-
ations regarding health information software. his family. James has a primary care physi-
This includes both understanding of general soft- cian, Linda Stark, who practices at a clinic
ware engineering principles, and then specifi- that is part of a larger health delivery net-
cally how these principles are applied to health work, Generation Healthcare System
care settings. (GHS). GHS includes a physician group,
To this aim, we first describe the major soft- primary and specialty care clinics, a ter-
ware functions within a health care environment tiary care hospital and an affiliated health
or health information system. While not all func- insurance plan.
tions can be covered in detail, some specific James needs to make an appointment
examples are given to indicate the breadth of soft- with Dr. Stark. He logs into the GHS patient
ware applications as well as to provide an under- portal and uses an online scheduling appli-
standing of their relevance. We also describe the cation to request an appointment. While in
software development life cycle, with specific the patient portal, James also reviews results
applications to health care. We then describe from his most recent visit and prints a copy
important considerations and strategies for acquir- of his current medication list in order to dis-
ing and implementing software in health care set- cuss the addition of an over-the-counter
tings. Finally, we discuss emerging software supplement he recently started taking.
engineering influences and issues and their impact Before James arrives for his visit, the
on health information systems. Each system can clinic’s scheduling system has already
be considered in regard to what it would take to alerted the staff of James’s appointment
make it functional in a health care system, and and the need to collect information related
what advantages and disadvantage the software to his diabetes. Upon his arrival, Dr. Stark’s
may have, based on how it was created and imple- nurse gathers the requested diabetes infor-
mented. Understanding this will help you identify mation and other vital signs data and enters
the risks and benefits of various applications, so these into the electronic health record
that you can identify how to optimize the positive (EHR). In the exam room, Dr. Stark reviews
impact of health information systems. James’s history, the new information gath-
ered today, and recommendations and
reminders provided by the EHR on a report
6.2 Software Functions in tailored to her patient’s medical history.
Health Care They both go over James’s medication list
and Dr. Stark notes that, according to the
6.2.1 Cases Study of Health Care EHR’s drug interaction tool, the supple-
Software ment he is taking may have an interaction
with one of his diabetes medications. One of
The following case study illustrates many impor- the reminders suggests that James is due for
tant functions of health care software. an HbA1c test and Dr. Stark orders this in
the EHR. Dr. Stark’s nurse, who has been
alerted to the lab test order, draws a blood
sample from James. Before the appointment
James Johnson is a 42-year old man living ends, Dr. Stark completes and signs his
in a medium-sized western U.S. city. He is progress note and forwards a visit summary
married and has two children. He has Type- for James to review on the patient portal.
II diabetes, but it is currently well-controlled A few days after his appointment, James
and he has no other health concerns. There receives an email from GHS that alerts him
is some history of cardiovascular disease in to an important piece of new information in

(021) 66485438 66485457

6 Software Engineering for Health Care and Biomedicine 187

his patient record. Logging into the patient his electronic record through a Web-based inter-
portal application, James sees that his HbA1c face to the clinic EHR. His medication and lab
test is back. The test indicates that the result is history, as well as notes from Dr. Stark and the
elevated. Dr. Stark has added a note to the care manager, help them quickly assess his con-
result saying that she has reviewed the lab and dition and develop a plan. James is admitted as
would like to refer James to the GHS Diabetes an inpatient for overnight observation and,
Specialty Clinic for additional follow-up. again, doctors and nurses on the ward are able
James uses the messaging feature in the to access his full record and record new observa-
patient portal to respond to Dr. Stark and tions and treatments, which are automatically
arrange for an appointment. James also clicks shared with the outpatient EHR. They are also
on an infobutton next to the lab result to able to reconcile his outpatient prescriptions
obtain more information about the abnormal with his inpatient medications to ensure continu-
value. He is linked to patient-focused material ity. James is stabilized by the next day. He
about HbA1c testing, common causes for high receives new discharge medications, which
results, and common ways this might be simultaneously discontinue his existing orders.
addressed. Lastly, James reviews the visit Because Dr. Stark is listed as James’s pri-
summary note from his appointment with Dr. mary care physician, she is notified both at
Stark to remind him about suggestions she admission and discharge of his current status.
had for replacing his supplement. She is able to review his discharge summary in
At his appointment with the Diabetes the EHR. She instructs her staff to send a mes-
Specialty Clinic, James notes that they have sage through the patient portal to James to let
access to all the information in his record. him know she had reviewed his inpatient record
A diabetes care manager reviews the important and to schedule a follow-up appointment.
aspects of James’s medical history. She sug- The GMS EHR is also part of a statewide
gests more frequent monitoring of his labora- health information exchange (HIE), which
tory test results to see if he is able to control his allows medical records to be easily shared
diabetes without changes to his medications. with health care providers outside a patient’s
She highlights diet and exercise suggestions in primary care provider. This means that if
his patient portal record that have been shown James should need to visit a hospital, emer-
to help. The care manager sends a summary of gency department or specialty care clinic out-
the visit to Dr. Stark so that Dr. Stark knows side the GMS network, his record would be
that James did follow-up with the Clinic. available for review and any information
A year later, James is experiencing greater entered by these outside providers would be
difficulty controlling his diabetes. Dr. Stark and available to Dr. Stark and the rest of the GMS
the Diabetes Care Manager have continued to network. In James’s state, the local and state
actively monitor his HbA1c and other labora- health departments are also linked to the HIE.
tory test results, and occasionally make changes This allows clinics, hospitals and labs to elec-
to his treatment regimen. They are able to use the tronically submit information to the health
EHR to track and graph laboratory test results departments for disease surveillance and case
and correlate them with changes in medications. reporting purposes.
Due to family problems, James struggles with Back at home, James’s wife, Gina, is able
adherence to his medication regimen, and he is to view his record on the GHS patient portal
not maintaining a healthy diet. As a result, his because he has granted her proxy access to
blood sugar has become seriously unstable and his account. This allows her to see the note
he is taken to the GHS hospital emergency from Dr. Stark and schedule the follow-up
department. Doctors in the ED are able to access appointment. Gina also views the discharge

(021) 66485438 66485457

188 A.B. Wilcox et al.

instructions that were electronically sent to about the study. Growing interested in the
James’s patient record. As she looks deeper possible benefits of the research, James
into information about diabetes that GHS had enrolls electronically in the study and is later
automatically linked to James’s record, Gina contacted by a study coordinator. Because
sees a note about a research study into genetic GHS researchers are conducting the study,
links with diabetes. Concerned about their relevant parts of James’s EHR can be easily
two children, Gina discusses the study with shared with the research data tracking
James, and he reviews the on-line material system.

This fictional case study highlights many of 6.2.2 Acquiring and Storing Data
the current goals for improving health care deliv-
ery, including: improved access to care, increased The amount of data needed to describe the state
patient engagement, shared patient-provider of even a single person is huge. Health profes-
decision-making, better care management, medi- sionals require assistance with data acquisition to
cation reconciliation, improved transitions of deal with the data that must be collected and
care, and research recruitment. In the case study, processed. One of the first uses of computers in a
each of these goals required software to make medical setting was the automatic analysis of
health information accessible to the correct indi- specimens of blood and other body fluids by
viduals at the proper time. instruments that measure chemical concentra-
In today’s health care system, few individu- tions or that count cells and organisms. These
als enjoy the interaction with software depicted systems generated printed or electronic results to
in the case study with James Johnson. Although health care workers and identified values that
the functions described in the scenario exist were outside normal limits. Computer-based
at varying levels of maturity, most health care patient monitoring that collected physiological
delivery institutions have not connected all data directly from patients were another early
the functions together as described. The cur- application of computing technology (see Chap.
rent role of software engineering in health care 19). These systems provided frequent, consistent
is therefore twofold: to design and implement collection of vital signs, electrocardiograms
software applications that provide required (ECGs), and other indicators of patient status.
functions, and to connect these functions in a More recently, researchers have developed medi-
seamless experience for both the clinicians and cal imaging applications as described in Chaps. 9
the patients. and 20, including computed tomography (CT),
The case study also highlights the usefulness magnetic resonance imaging (MRI), and digital
of several functions provided by health care soft- subtraction angiography. The calculations for
ware applications for clinicians, patients, and these computationally intensive applications can-
administrators. Some of these functions include: not be performed manually; computers are
1. Acquiring and storing data required to collect and manipulate millions of
2. Summarizing and displaying data individual observations.
3. Facilitating communication and information Early computer-based medical instruments
exchange and measurement devices provided results only
4. Generating alerts, reminders, and other forms to human beings. Today, most instruments can
of decision support transmit data directly into the EHR, although the
5. Supporting educational, research, and public interfaces are still awkward and poorly standard-
health initiatives ized (see Chaps. 4 and 7). Computer-based

(021) 66485438 66485457

6 Software Engineering for Health Care and Biomedicine 189

systems that acquire information, such as one’s EHR system can be easily retrieved. Here the
health history, from patients are also data- variety of users must be considered. Getting
acquisition systems; they free health profession- cogent recent information about a patient enter-
als from the need to collect and enter routine ing the office differs from the needs that a
demographic and history information. researcher will have in accessing the same data.
Various departments within a hospital use com- The query interfaces provided by EHRs and clin-
puter systems to store clinical data. For instance, ical research systems assist researchers in retriev-
clinical laboratories use information systems to ing pertinent records from the huge volume of
keep track of orders and specimens and to report patient information. As discussed in Chap. 21,
test results; most pharmacy and radiology depart- bibliographic retrieval systems are an essential
ments use computers to perform analogous func- component of health information services.
tions. Their systems may connect to outside services
(e.g., pharmacy systems are typically connected to
one or more drug distributors so that ordering and 6.2.4 Facilitating Communication
delivery are rapid and local inventories can be kept and Information Exchange
small). By automating processing in areas such as
these, health care facilities are able to speed up ser- In hospitals and other large-scale health care
vices, reduce direct labor costs, and minimize the institutions, myriad data are collected by multi-
number of errors. ple health professionals who work in a variety of
settings; each patient receives care from a host of
providers—nurses, physicians, technicians, phar-
6.2.3 Summarizing and macists, and so on. Communication among the
Displaying Data members of the team is essential for effective
health care delivery. Data must be available to
Computers are well suited to performing tedious decision makers when and where they are needed,
and repetitive data-processing tasks, such as col- independent of when and where they were
lecting and tabulating data, combining related obtained. Computers help by storing, transmit-
data, and formatting and producing reports. They ting, sharing, and displaying those data. As
are particularly useful for processing large vol- described in Chaps. 2 and 12, the patient record is
umes of data. the primary vehicle for communication of clini-
Raw data as acquired by computer systems are cal information. The limitation of the traditional
detailed and voluminous. Data analysis systems paper-based patient record is the concentration of
must aid decision makers by reducing and pre- information in a single location, which prohibits
senting the intrinsic information in a clear and simultaneous entry and access by multiple peo-
understandable form. Presentations should use ple. Hospital information systems (HISs; see
graphs to facilitate trend analysis and compute Chap. 13) and EHR systems (Chap. 12) allow
secondary parameters (means, standard devia- distribution of many activities, such as admis-
tions, rates of change, etc.) to help spot abnor- sion, appointment, and resource scheduling;
malities. Clinical research systems have modules review of laboratory test results; and inspection
for performing powerful statistical analyses over of patient records to the appropriate sites.
large sets of patient data. The researcher, how- Information necessary for specific decision-
ever, should have insight into the methods being making tasks is rarely available within a single
used. For clinicians, graphical displays are useful computer system. Clinical systems are installed
for interpreting data and identifying trends. and updated when needed, available, and
Fast retrieval of information is essential to all affordable. Furthermore, in many institutions,
computer systems. Data must be well organized inpatient, outpatient, and financial activities are
and indexed so that information recorded in an supported by separate organizational units.

(021) 66485438 66485457

190 A.B. Wilcox et al.

Patient treatment decisions require inpatient may be available but are easily overlooked by
and outpatient information. Hospital adminis- overloaded health professionals. Surveillance
trators must integrate clinical and financial and monitoring systems can help people cope
information to analyze costs and to evaluate the with all the data relevant to patient management
efficiency of health care delivery. Similarly, cli- by calling attention to significant events or situa-
nicians may need to review data collected at tions, for example, by reminding doctors of the
other health care institutions, or they may wish need to order screening tests and other preventive
to consult published biomedical information. measures (see Chaps. 12 and 22) or by warning
Communication networks that permit sharing them when a dangerous event or constellation of
of information among independent computers events has occurred.
and geographically distributed sites are now Laboratory systems routinely identify and flag
widely available. Actual integration of the abnormal test results. Similarly, when patient-
information they contain requires additional monitoring systems in intensive care units detect
software, adherence to standards, and opera- abnormalities in patient status, they sound alarms
tional staff to keep it all working as technology to alert nurses and physicians to potentially dan-
and systems evolve. gerous changes. A pharmacy system that main-
tains computer-based drug-profile records for
patients can screen incoming drug orders and
6.2.5 Generating Alerts, Reminders, warn physicians who order a drug that interacts
and Other Forms of Decision with another drug that the patient is receiving or
Support a drug to which the patient has a known allergy or
sensitivity. By correlating data from multiple
In the end, all the functions of storing, displaying sources, an integrated clinical information sys-
and transmitting data support decision making by tem can monitor for complex events, such as
health professionals, patients, and their caregiv- interactions among patient diagnosis, drug regi-
ers. The distinction between decision-support men, and physiological status (indicated by labo-
systems and systems that monitor events and ratory test results). For instance, a change in
issue alerts is not clear-cut; the two differ primar- cholesterol level can be due to prednisone given
ily in the degree to which they interpret data and to an arthritic patient and may not indicate a
recommend patient-specific action. Perhaps the dietary problem.
best-known examples of decision-support sys-
tems are the clinical consultation systems or
event-monitoring systems that use population 6.2.6 Supporting Educational,
statistics or encode expert knowledge to assist Research, and Public Health
physicians in diagnosis and treatment planning Initiatives
(see Chap. 22).