You are on page 1of 3

Machine Learning for the Improvement of

Healthcare
Sebastián Gutiérrez Jaramillo
Faculty of Engineering, Universidad de Antioquia
Medellín, Colombia
sebastian.gutierrezj@udea.edu.co

Abstract— As the population continues to grow, medical considerably improve their hopes of recovery. Now, talking about
care becomes increasingly difficult. Therefore, a lot of research how this tool works, one can find an article from the University of
is done to improve the quality of the healthcare system around Ljubljana in Slovenia, by Igor Kononenko [2] where he talks about
the world, but there is one approach that promises more than the ability of machine learning to learn through symbols. Working
the others. With Machine Learning, some processes of the with decision trees and decision rules, a program can analyze large
medical attention could be not only automatized but improved volumes of data and compare final diagnoses with medical
in terms of efficiency and accuracy. The major advantages that examinations to achieve the precision of an experienced
can provide this technology are related with analyzing patterns professional but in a hundredth of the time.
and linking medical histories with databases. In this way,
collaboration between healthcare professionals and software To arrive at a brief explanation of what machine learning is, it
engineers is an exceptional option to begin the development of is necessary to consider the contrast of it with traditional
future healthcare. programming. As Isaac Kohane [1] shows, traditional
programming requires an engineer to write in a specific language
Keywords: machine learning, artificial intelligence, analysis, giving specific instructions for a program to run correctly, that is,
healthcare, improvement. they need rules established by a person. In terms of medicine, it
would be incredibly difficult to develop a useful program in a
I. JUSTIFICATION traditional way because medicine is a huge and complex field that
cannot be reduced or compacted into rules. However, Machine
Today, everyone has heard of at least one case of ineffective Learning solves several of these drawbacks: first, it is not
healthcare, especially in less developed countries. The use of necessary for the programmer to write all the code that governs the
artificial intelligence tools such as Machine Learning could make software, he only has to establish limits and from there the
processes cheaper and improve time efficiency. And with program can be self-sufficient; Second, Machine Learning is not
population growth overcoming medical services and staff, many based on established rules as mentioned above, on the other hand,
lives would be saved by implementing automated diagnostics and it can be said that it works as a model that learns through
smart databases. Therefore, it is a much-needed research that could examples.
improve the quality of life of countless people worldwide and
signify the next step for humanity in terms of technology. Kononenko [2] describes in his work some examples of
algorithms that have currently been tested to measure their
II. OBJECTIVES capabilities as applications for complementary medicine. From the
A. General naive Bayesian classifier that uses a specific formula that serves to
determine the probability of a certain event occurring given some
Show the wide possibilities that Machine learning applications previous conditions, to the R-wizard system that bases its
have in medicine. operation on decision trees, learning with the test method and
error. These and several other programs were compared on
B. Specific features such as performance, transparency, and detailed
• Determine the consequences of updating traditional processes explanations, showing that each has unique advantages but that all
are excellent ways to improve the clinical diagnostic process.
and if they are worth assuming.
• Estimate the costs of implementation and training of the A particularly important aspect related to the implementation
of this technology is its consequences. While it is natural to think
artificial intelligence tool. that this could mean a reduction in the total amount of work
• Present current examples of methods using Machine available to new healthcare professionals, despite this, Dr. Jeffrey
Dean [1] talks about how Machine Learning can increase the work
Learning in medicine. of the clinicians. Because this tool can take advantage of the
enormous amount of data and patterns found in the files of many
patients, it can also make predictions with greater precision. In this
III. STATE OF ART
way, many more people will be able to treat their diseases in time,
There has been some previous research related to machine which would be much more difficult without this new technology.
learning for medical purposes. In each of them the disadvantages Machine Learning is a tool that allows us to see how artificial
of traditional healthcare are exposed, mainly in terms of intelligence could mean improvements in different areas of science
inefficiency of time. For example, in an article [1] published by and health, and that it is completely possible to implement it
Stanford University for the New England Journal of Medicine, the without compromising the employment of professionals in these
contrast is made of how a patient has to wait for multiple areas.
specialists from different areas to have the time to attend to him
and therefore during that period his illness inevitably progresses.
As explained later in the same article, the possibility of having an
intensive specialist always available to the patient could
IV. THEORETICAL FRAMEWORK patients. Next, the model must be trained in its ability to interpret
the data it has collected by performing processes that can be
The application of machine learning has its potential mainly in roughly described as trial and error [1] in order to correct its
the diagnostic stage of health care. First, it is necessary to think of interpretations based on parameters set by the software architect in
this tool as a model, which "can learn the patterns of the health charge. Finally, the results obtained by the machine learning model
trajectories of a large number of patients." And this can make it are compared with a proven field reference and thus can ensure
easier in the future to anticipate events related to the health of a how accurate the program is.
given patient. In addition, thanks to this pattern analysis, it will
also be possible to approximate details such as: how likely it is that Why is it important to consider Machine Learning for the
the patient will return to work after recovery or how quickly his or medicine of the future? Well, as stated in the Stanford University
her illness could progress. Now that the concept is understood as a article [1] it is concluded that a diagnostic error will occur in
model that interprets patterns, it can be separated into two main practically all patients' lives at least once. Furthermore, there will
types: supervised learning and unsupervised learning. be cases in which it will be critical to give the correct diagnosis at
the right time in order to facilitate treatment and improve life
Supervised Learning expectancy in general. Currently, hospital databases are already
Supervised learning begins with the objective of predicting a mostly digital and tend to continue in this way, thus contributing,
known outcome or goal. Recurring supervised learning problems hopefully and unexpectedly, to the growth of databases that are
include handwriting recognition (such as handwriting recognition), optimal for training the models of Machine-Learning already
image classification of objects (e.g., is it a cat or a dog?), and mentioned. And because the accuracy of these algorithms is
document classification (e.g., is it a heart failure clinical trial or a directly proportional to the amount of data they have to train and
financial report?). Remarkably, these are all tasks that a trained the number of times their training function is executed, the first
person can do well and therefore the computer often tries to tool is already there waiting for research to continue so that the
approximate human performance. Supervised learning focuses on number of times these programs are launched to improve medical
classification, which involves choosing between subgroups to diagnosis, both in accuracy and in time efficiency, also grows.
better describe a new instance of data, and prediction, which
involves estimating an unknown parameter (such as the As Benjamin Glicksberg [3] says in his article entitled
temperature in San Francisco tomorrow afternoon). Automatic Learning in Cardiovascular Medicine: are we there yet?
What might be some examples of supervised learning in There is a need for this tool in cardiovascular medicine, because it
medicine? Perhaps the most common example seen by a is a field of medicine in which predictions can be especially far-
cardiologist is automated ECG (electrocardiography) reaching. And with Machine Learning, data can be analyzed in
interpretation, where pattern recognition is performed to select much less than half the time, and with the help of specialized
from a limited set of diagnoses. In radiology, automated detection physicians, data developers, and data scientists, you can also
of a lung nodule from a chest X-ray would also represent achieve the quality that health professionals would. According to
supervised learning. In both cases, the computer approximates Kononenko [2], the best way to implement these Machine
what a trained physician is already capable of doing with great Learning tools in today's healthcare system is as a complement to
accuracy. medical care. In this way, physicians would continue to care for
patients and reach out to their own conditions, while, in addition,
artificial intelligence tools routinely analyze data in search of
Unsupervised Learning
patterns in the data corresponding to patients with similar
In contrast, in unsupervised learning, there are no outcomes to characteristics and can predict through unsupervised learning
predict. Instead, we are looking for patterns or groupings that diseases or complications that have some significant probability of
occur naturally within the data sets. This is a more difficult task to occurring in the future.
judge and often the value of the outcomes achieved through
unsupervised learning is assessed by asking: Are these new Clinician Workflow
patterns useful in any way? How could this kind of learning be
used in medicine? Well, given the immense amount of variation Just as many artificial intelligence tools are used in the
that exists in the most common diseases, predicting their behavior entertainment field to make it easier for consumers to find content
is very difficult for supervised learning algorithms. However, since they like; they can also be applied in physicians' routines to
these are mostly heterogeneous groupings of data, unsupervised facilitate the conditions necessary to treat more patients in a timely
learning could be used to look for patterns among the thousands of and efficient manner. In what way? In addition to improving the
data. diagnostic process as mentioned above, machine learning tools can
One concrete application, as stated in an article by the also help expose relevant information in a patient's medical record
American Heart Association [4], would be to apply unsupervised to a physician without the need for multiple clicks. Data entry from
learning to heart disease, taking a heterogeneous condition such as forms and text fields can be improved with the use of machine
myocarditis. One can start with a large group of similar individuals learning techniques such as predictive typing, voice dictation, and
with unexplained heart failure. One can then perform biopsies on automatic summarization.
them and characterize the cellular composition of each sample. For
example, a count of T lymphocytes, neutrophils, macrophages, Not only does this increase the speed at which healthcare
eosinophils, etc. would be done. It would then be possible to see if professionals can perform their tasks, it also means that they can
there are recurrent patterns of cell composition, which, in turn, provide more work in less time and will make it possible to
could suggest a mechanism and guide the therapies to be explored. employ new professionals within healthcare institutions. Also
noteworthy are the broad forms of integration offered by Machine
Now, whether the model is monitored or not, there are some Learning tools for the world of medicine and information
steps that must be taken. First, it is necessary to prepare the technology, moving towards a more efficient and therefore more
construction of the model by establishing the main task for the accessible service for all vulnerable populations in the poorest
algorithm, and then select the database from which the algorithm parts of the world.
will do the data collection. For example, if the objective of the
model is to determine the patterns of a common disease, the
database could be the medical history of a group of diverse
REFERENCES
[1] Rajkomar, A., Dean, J. and Kohane, I., 2019. Machine Learning In
Medicine. [online] Med.stanford.edu. Available at:
<https://med.stanford.edu/content/dam/sm/shvca/documents/july-31-
ai-homework3.pdf> [Accessed 29 September 2020].
[2] Kononenko, I., n.d. Machine Learning For Medical Diagnosis:
History, State Of The Art And Perspective. [online]
Citeseerx.ist.psu.edu. Available at:
<http://citeseerx.ist.psu.edu/viewdoc/download?
doi=10.1.1.96.184&rep=rep1&type=pdf> [Accessed 29 September
2020].
[3] K. Shameer, K. Johnson, B. Glicksberg, J. Dudley and P. Sengupta.
Machine learning in cardiovascular medicine: are we there yet?
Heart.bmj.com, 2018. [online]. Available:
https://heart.bmj.com/content/heartjnl/104/14/1156.full.pdf.
[Accessed: 20- Oct- 2020].
[4] R. Deo. Machine Learning in Medicine. Ahajournals.org, 2015.
[online]. Available:
https://www.ahajournals.org/doi/epub/10.1161/CIRCULATIONAHA.
115.001593. [Accessed: 20- Oct- 2020].

You might also like