This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2016.2614991, IEEE Journal of
Biomedical and Health Informatics
JBHI-00268-2016.R1
A Shared Decision Making System for Diabetes
Medication Choice Utilizing Electronic Health
Record Data
Yu Wang, Peng-fei Li, Yu Tian, Jing-jing Ren, Jing-song Li*
 Organization (WHO), 9% of adults worldwide have diabetes,
Abstract—The use of a shared decision making (SDM) process
in antihyperglycemic medication strategy decisions is necessary
due to the complexity of the conditions of diabetes patients.
Knowledge of guidelines is used as decision aids in clinical
situations, and during this process, no patient health conditions
are considered. In this manuscript, we propose a SDM system
framework for type 2 diabetes mellitus (T2DM) patients that not
only contains knowledge abstracted from guidelines but also
employs a multi-label classification model that uses
class-imbalanced electronic health record (EHR) data and that
aims to provide a recommended list of available
antihyperglycemic medications to help physicians and patients
have a SDM conversation. The use of EHR data to serve as a
decision support component in decision aids helps physicians and
patients to reach a more intuitive understanding of current health
conditions and allows the tailoring of the available knowledge to
each patient, leading to more effective SDM. Real-world data
from 2,542 T2DM inpatient EHRs were substituted by 77 features
and eight output labels, i.e., eight antihyperglycemic medications,
and these data were utilized to build and validate the
recommendation model. The multi-label recommendation model
exhibited stable performance in every single-label classification
and showed the ability to predict minority positive cases in which
the average recall value of the eight classes was 0.9898 . As a whole
multi-label classifier, the recommendation model demonstrated
outstanding performance, with scores of 0.0941 for Hamming Loss,
0.7611 for Accuracyexam, 0.9664 for Recallexam, and 0.8269 for Fexam.
Index Terms—Decision aids, electronic health records,
multi-label classification, shared decision making
I. INTRODUCTION
D IABETES is one of the most influential noncommunicable
diseases worldwide. According to the World Health
This work was supported by Chinese National High-tech R&D Program
(No.2013AA041201 & No.2015AA020109) and the Fundamental Research
Funds for the Central Universities of China.
Yu Wang, Peng-fei Li, and Yu Tian are with Engineering Research Center
of EMR and Intelligent Expert System, Ministry of Education, Collaborative
Innovation Center for Diagnosis and Treatment of Infectious Diseases, College
of Biomedical Engineering and Instrument Science, Zhejiang University,
Hangzhou, China (e-mail: daisy222542@zju.edu.cn; 21115055@zju.edu.cn;
ty.1987823@163.com).
Jing-jing Ren is with General Practice Department, The First Affiliated
Hospital, Zhejiang University, Hangzhou, China (e-mail: lisarjj@126.com).
*Jing-song Li is with Engineering Research Center of EMR and Intelligent
Expert System, Ministry of Education, Collaborative Innovation Center for
Diagnosis and Treatment of Infectious Diseases, College of Biomedical
Engineering and Instrument Science, Zhejiang University, Hangzhou, China;
(e-mail: ljs@ zju.edu.cn).
2168-2194 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1
and 1.5 million deaths are directly attributed to diabetes each
year [1]. In the report China Noncommunicable Disease
Surveillance 2010, 9.7% of adults older than 18 years old were
estimated to have diabetes using the 1999 WHO diabetes
diagnostic criteria, whereas 11.7% were estimated to have
diabetes using the 2010 American Diabetes Association (ADA)
criteria, including a glycated hemoglobin A1c (HbA1c)
concentration of 6.5% or higher [2], thus making China the
country with the largest diabetes population in the world. Of the
Chinese diabetes population, more than 95% have type 2
diabetes mellitus (T2DM). Complications are common in
T2DM patients, and antihyperglycemic medications are
prescribed to manage glycemic control and to reduce the risks
of developing complications. The antihyperglycemic agents
commonly prescribed for T2DM patients include metformin,
sulfonylureas, meglitinides, thiazolidinediones, and insulin [3].
Surveys have found that the adverse effects of and burdens
imposed by medications might be greater than those of future
diabetes complications [4]; thus, the trade-off between the
benefits and harms of medications must be considered.
Therefore, physicians face the challenge of choosing the right
medication strategy for patients and often feel confused and
unsupported [5]. In recent years, new antihyperglycemic agents
have joined the pharmacological armamentarium, further
increasing the complexity of the T2DM medication strategy. In
addition to diabetes guidelines, determining a specific
medication strategy for certain T2DM patients to achieve the
glycemic control target is also attributed to the patients’ values
and preferences, in which the shared decision-making (SDM)
process is necessary.
The best decision is often difficult to make in clinical
situations, and a trade-off is necessary between known benefits
and harms. Patient-centered care is introduced as “the care that
is respectful of and responsive to individual patient preferences,
needs, and values” [6] and has been found to improve patient
outcomes [7]. SDM is a core strategy of promoting
patient-centered care [8]. In SDM, patients and physicians
collaborate to make decisions using the best available evidence,
and patients are encouraged to consider available treatment
options and the likely benefits and harms of each option [9].
Decision aids (DAs), the interventions that support SDM,
prepare patients to participate in making specific choices
among available treatment options that they prefer and can be
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2016.2614991, IEEE Journal of
Biomedical and Health Informatics
JBHI-00268-2016.R1
used as a supplement to physicians’ consultations during
clinical activities [10]-[12]. Because patient-centered care is
ubiquitous in modern healthcare, DAs can be used before,
during, or after the clinical encounter in a variety of carriers
such as cards, brochures, or videos.
DAs are developed in various clinical situations [13]-[15]. In
SDM for diabetes, Breslin et al. designed the Diabetes
Medication Choice [16] for patients with T2DM to lessen
decisional conflict and increase the patients’ knowledge,
communication, and involvement in the decision-making
process [17]-[19]. Diabetes Medication Choice integrates
HbA1c control effects, daily routine, and side effects, including
those regarding weight gain and the hypoglycemia risks of five
classes of antihyperglycemic medications, using issue cards
[16]. In addition to Diabetes Medication Choice, the latest
Cochrane Decision Aid Inventory cites six additional DAs for
T2DM [20], three of which are public. In the decision-making
process, a patient always expects to know what decisions other
patients similar to him/herself would make; however, all of the
above-mentioned T2DM DAs only prepare patients with
common T2DM knowledge based on guidelines and random
trials, which are derived to address this balancing act. Although
the inclusion of knowledge regarding guidelines is essential for
helping patients make decisions, it is too much information for
patients, most of whom are not professionals, to understand and
utilize. The most appropriate approach for giving patients
personalized access to the available options of diabetes
medications is to tailor that knowledge. Therefore, we included
electronic health record (EHR) data in the process used for the
development of T2DM DAs that would help patients
understand their specific clinical situation and the specific
knowledge they might need.
In the past decade, patients’ clinical histories and values have
been given increasing emphasis in clinical interactions and
have been well documented in electronic health records (EHRs)
as a result of the spread of health information technologies
(HIT) [21]. The results from reusing EHR data have been used
in more and more decision support situations [22]-[24], which
allows patients undergoing the decision-making process to
review the decisions of similar patients. In this study, we
proposed a SDM system framework connected to EHRs for
T2DM patients by combining data mining results and relevant
knowledge to provide decision support and to provide patients
and physicians with tailored knowledge and choices about
antihyperglycemic medications. We developed an app for
Android smart terminals as the prototype of this framework in
which the patients’ clinical histories and values are extracted
from EHRs and a multi-label classification method was applied
to these data to achieve a recommendation among 8 classes of
available antihyperglycemic medications. The SDM system we
proposed can help both patients and physicians review patients’
conditions more comprehensively and guide the consultations
more specifically in adherence to patients’ current conditions.
2168-2194 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
2
II. METHODS
A. Design framework and system structure
International Patient Decision Aids Standards (IPDAS)
criteria [10], [12] and an SDM model proposed by Elwyn [25]
were utilized when designing and developing our proposed
SDM system. The DA development process is not specifically
provided by IPDAS criteria, but certain attributes of the
decision-making process are defined. For this T2DM SDM
system, we considered the following IPDAS criteria: 1)
providing information about health conditions by organizing
patients’ histories and current values relevant to T2DM; 2)
helping patients understand that an antihyperglycemic
medication choice should be made; 3) providing available
antihyperglycemic medication options and their benefits and
harms; and 4) providing an intelligent recommendation of
antihyperglycemic medications to help patients understand how
their specific values will affect their decision.
The T2DM SDM system was designed to be
physician-guided, and it would be suitable in no matter
inpatient setting or outpatient setting from the view of usage.
Fig. 1. Overview of the T2DM SDM system framework.
An overview of the T2DM SDM system framework is shown in
Fig. 1. EHR data are extracted by the T2DM SDM system, and
a recommendation model for T2DM medication choice is built
utilizing EHR data. The evidence-based knowledge base
provides preliminary knowledge about options, as well as the
recommendation model and certain raw data in the EHR
constitute the decision support basis of the T2DM SDM system.
In this study, relevant antihyperglycemic medication
knowledge has been abstracted from the 2015 ADA Diabetes
Guidelines [5] and the 2013 Chinese Diabetes Association
(CDA) Diabetes Guidelines. The detailed algorithm and
validation of the recommendation model will be described
later in this paper.
Following Elwyn’s SDM model, three phases were designed
in our T2DM SDM system: choice talk, option talk, and
decision talk (Fig. 2). When an antihyperglycemic medication
prescription is required by an inpatient, the choice talk will
begin with a review of the patient’s health conditions, including
demographic characteristics, clinical histories, lab test results,
and current vital sign values with all of the above health
conditions obtained from the EHR via a clinical history review
panel. During the comprehensive review of health conditions,
physicians will help patients recognize what condition they are
in and the necessity of an antihyperglycemic medication
strategy. In the option talk phase, knowledge of available
antihyperglycemic medications based on current guidelines and
meta-analyses will be provided to prepare patients with a
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2016.2614991, IEEE Journal of
Biomedical and Health Informatics
JBHI-00268-2016.R1
Fig. 2. T2DM SDM system decision making process.
preliminary understanding of their options by presenting
benefits and harms. Costs of antihyperglycemic medications
are extracted directly from EHRs. Using the recommendation
model, a medication strategy will additionally be provided to
patients, making available options more concise to understand.
In the decision talk phase, physicians will guide patients to
form their preferences and make a decision based on an
understanding of their available options. In the proposed DAs,
the option talk and decision talk phases will share the same
interface.
B. Model of T2DM medication recommendation
1) Problem transformation
According to ADA hyperglycemia management criteria,
T2DM medication therapy consists of several phases: initial
drug therapy, advancing to dual combination and triple
combination therapy, and combination injectable therapy [5].
An antihyperglycemic medication strategy might be more
complicated than that described in the guidelines because
T2DM patients are always accompanied by various
complications and their conditions are highly varied. An
analysis of real-world EHR data revealed that a T2DM
inpatient could be prescribed three or more antihyperglycemic
medications simultaneously; therefore, the output of a T2DM
medication recommendation strategy is a list of appropriate
medications rather than one single medication. Thus, the
recommendation problem is transformed to a multi-label
classification problem.
We will now describe the development of the T2DM
medication recommendation model using a multi-label
classification method. The input of the model will be the
features of T2DM patients documented in their EHRs,
including demographic data, concomitant diseases, lab test
results and vital sign values. The output of the model will be
one or more out of the following eight classes of
antihyperglycemic medications: alpha-glucosidase inhibitors,
DPP-4 inhibitors, meglitinides, GLP-1 receptor inhibitors,
insulin, metformin, sulfonylureas, and thiazolidinediones. In
this manuscript, we denote ‘labels’ as the predicted medication
classes, which are the output of the recommendation model.
2) Multi-label medication recommendation model
In the multi-label learning problem, one case could be
classified into different categories/classes, and the multi-label
learning problem could be divided into two groups: 1) problem
transformation methods that transform the problem into one or
more single-label classification problems; or 2) algorithm
2168-2194 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
3
Fig. 3. Flow chart of the multi-label classification algorithm.
adaption methods that extend specific learning algorithms to
address multi-label data directly.[26] In reality, very often the
training and/or test data are class-imbalanced (the number of
observations belonging to each class is not balanced). The
class-imbalance phenomenon usually produces classifiers that
have poor predictive recall, particularly when the positive label
is the minority class [27]. In this study, an ensemble learning
algorithm is built to work on both the multi-label classification
problem and the class-imbalanced problem.
Fig. 3 shows the flow chart of the ensemble learning
algorithm. We integrated Binary Relevance (BR) [28],
Synthetic Minority Over-sampling Technology (SMOTE) [29],
and Random Forest [30]. First, BR was employed as the
building block of the recommendation model to decompose the
multi-label learning problem into q independent binary
classification problems, where q is the number of labels. We
hypothesized a lack of potential correlations among labels.
Then, SMOTE was applied to decrease the effect of the
class-imbalanced problem. Finally, Random Forest was used as
a meta-binary-classifier to predict whether a label should be
assigned to the case. The above ensemble learning algorithm is
called BSR for short in the sections below.
3) Model validation
To build the multi-label recommendation model,
hospitalization data from a large comprehensive hospital in
Beijing, China were utilized. We retrieved data from 2,582
patients older than 18 years old with a primary diagnosis of
T2DM (not including T2DM during pregnancy) who were
consuming one or more types of antihyperglycemic
medications and were admitted between June 2007 and April
2014. Their demographic data, concomitant diseases, lab test
results, and vital sign values were collected from the hospital
EHR.
For SMOTE, if the number of cases in the minority class of
training data is N, then N “synthetic” examples are created to
over-sample the minority class, and 2×N examples of the
majority class are selected to form the final training set. For
Random Forest, 500 trees are constructed, and eight features
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2016.2614991, IEEE Journal of
Biomedical and Health Informatics
JBHI-00268-2016.R1 4
TABLE I
are randomly selected when building a single tree. A 10-fold STATISTICAL CHARACTERISTICS OF THE FINAL MULTI-LABEL DATASET
cross validation was performed to test the accuracy of the Dataset statistics measures Count (n%)
multi-label recommendation model. To validate the Total 2542
performance of the ensemble algorithm, two other popular Number of patient features 77
multi-label classifiers were compared.
Label cardinality 1.552
Multi-Label k-Nearest Neighbor (ML-kNN): k-nearest
neighbor techniques were adapted to address multi-label data in Input features Count (n%) Mean (±SD)
which a maximum a posteriori (MAP) rule was utilized to make Gender
a prediction by reasoning with the labeling information
Male 1668 (0.656)
embodied in the neighbors, proposed by Zhang et al [31]. The
number of neighbors, k, was set to 10, and the smoothing factor Female 874 (0.344)
was set to 1. Age 66.46 (±13.81)
• Random k-Labelsets (RAkeL): a small random subset of (18, 45] 173 (0.068)
labels and learning of a single-label classifier was considered (45, 70] 1280 (0.504)
for the prediction of each element in the powerset of this subset >70 1089 (0.428)
[32]. Random Forest was used as the base binary learner. Concomitant
The performance measures for multi-label classification diseases (CD)
problems can be categorized into two groups: example-based With no CD 927 (0.365)
metrics and label-based metrics [26]. In this study, we With one CD 1127 (0443)
employed four example-based measures: Hamming Loss,
With 2 CDs 362 (0.142)
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦𝑒𝑥𝑎𝑚 , 𝑅𝑒𝑐𝑎𝑙𝑙𝑒𝑥𝑎𝑚 , and 𝐹𝑒𝑥𝑎𝑚 .
With 3 or more
Hamming Loss evaluates the fraction of misclassified 126 (0.050)
CDs
instance-label pairs, in which is a relevant label is missed or an Vital sign
irrelevant label is predicted; defined as Equation (1), in which h values
denotes the multi-label classifier; 𝑝 is the number of test sets; 𝑌𝑖 Systolic blood
132.58 (±17.71)
is the true set of test data; ℎ(𝑥𝑖 ) is the predicted set of test data; pressure (mmHg)
and ∆ stands for the symmetric difference between two sets. Diastolic blood
1 𝑝 76.78 (±11.11)
pressure (mmHg)
Hloss(h) = ∑𝑖=1 |ℎ(𝑥𝑖 )∆𝑌𝑖 | (1)
𝑝
And 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦𝑒𝑥𝑎𝑚 , 𝑅𝑒𝑐𝑎𝑙𝑙𝑒𝑥𝑎𝑚 , and 𝐹𝑒𝑥𝑎𝑚 are defined as Temperature (℃) 36.26 (±2.21)
Equation (2)-(4).
1 𝑝 |𝑌 ⋂ℎ(𝑥𝑖 )| Pulse rate (/min) 75.30 (±9.52)
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦𝑒𝑥𝑎𝑚 (h) = ∑𝑖=1 𝑖 (2)
𝑝 |𝑌𝑖 ⋃ℎ(𝑥𝑖 )|
1 𝑝 |𝑌𝑖 ⋂ℎ(𝑥𝑖 )| Output labels Count
𝑅𝑒𝑐𝑎𝑙𝑙𝑒𝑥𝑎𝑚 (h) = ∑ (3)
𝑝 𝑖=1 |𝑌 | 𝑖
2×|𝑌𝑖 ⋂ℎ(𝑥𝑖 )| Alpha-glucosidase inhibitors 953
𝐹𝑒𝑥𝑎𝑚 (h) = |𝑌𝑖 |+|ℎ(𝑥𝑖 )|
(4)
DPP-4 inhibitors 24
Meglitinides 163
III. RESULTS GLP-1 receptor inhibitors 11
A. Recommendation model performance and validation Insulin 1525
After data cleaning, 2,542 cases were included with eight Metformin 788
medication labels and 77 patient attributes that contained two Sulfonylureas 415
demographic variables, 11 concomitant diseases, five vital sign Thiazolidinediones 66
values, and 57 lab test results. The gender and concomitant
diseases variables were binary, the age and vital sign value antihyperglycemic medications were prescribed to each patient
variables were continuous and regularized, and the lab test on average, was defined as Equation (5), where 𝐷 is the number
results variables were nominal because the structured lab test of cases and 𝑌𝑖 represents the label of the 𝑖 case.
1 |𝐷|
results included three statuses in the hospital EHR: H (test LC(D) = ∑𝑖=1 |𝑌𝑖 | (5)
|𝐷|
results were higher than the upper limit of the normal range); N
As shown in Table I, a class-imbalance phenomenon was
(test results fell in the normal range); and L (test results were
evident after decomposition of the multi-label dataset into eight
lower than the lower limit of the normal range). Table I lists the
single-label datasets. Seventy percent of the 2542 cases were
statistical characteristics of the patients in the final dataset. In
included in the training set, and the others were included in the
total, 65.6% of the patients were male, and 93.2% of the
test set. Fig. 4 illustrates the average performances of 10-fold
patients were older than 45 years old. Of the eight output
cross validation single-label Random Forest classifiers for each
medication classes, insulin was prescribed to the majority of
antihyperglycemic medication label with or without SMOTE.
patients whereas GLP-1 receptor inhibitors were prescribed to
If SMOTE was not applied before Random Forest, the classifier
the least number of patients (only 11). The Label Cardinality
accuracy was high in medication labels where class-imbalance
(LC) value was 1.552, which means that 1.552 classes of
was severe, such as DPP-4 inhibitors (0.9906) and GLP-1

2168-2194 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2016.2614991, IEEE Journal of
Biomedical and Health Informatics
JBHI-00268-2016.R1 5

receptor inhibitors (0.9958), and relatively low in medication
labels where positive cases and negative cases were relatively
balanced, such as α-glycosidase inhibitors (0.6603) and insulin
(0.6648). After applying SMOTE, the classifier accuracy for
each medical label was found to be stable. Recall measure
indicates the classifier’s ability to predict positive cases. Fewer
positive cases in the dataset resulted in a more severe
class-imbalance and lower recall. Without SMOTE, the recall
other two methods.
B. System implementation
According to the above designed framework and SDM
process, a T2DM SDM system prototype was developed that
can run on Android smart devices. EHR data were extracted
using an Ensemble platform (Intersystems, Cambridge, MA,
United States.), and the BSR model was built using R language.
Fig. 5 shows two screenshots of the T2DM SDM system.
During the choice talk phase, physicians can help patients
understand their specific conditions and review their clinical
history using the clinical history review panel (Fig. 5a). In the
option talk and decision talk phases, physicians use the same
interface (Fig. 5b). Both panels in Fig. 5 are mainly divided into
three parts: basic patient information (red region in Figs. 5a and
5b), current vital sign values (purple region in Figs 5a and 5b),
and the clinical history review as a timeline (Fig. 5a) or a list of

Fig. 4. Performance for each antihyperglycemic medication label with or

without SMOTE.

was zero on all of the medication labels with the exception of

α-glycosidase inhibitors (0.315), insulin (0.8436), and
metformin (0.2211). SMOTE improved the classifier’s ability
to predict positive cases; all of the positive cases in six
medication labels were predicted, and the recall for the other
two medications was also high (0.9217 for insulin and 0.9970
for thiazolidinediones). The F-measure provides a
comprehensive assessment of the classifier’s ability to predict
both negative and positive cases. As shown in Fig. 4c, the
F-measure was higher after applying SMOTE for every
TABLE II
PERFORMANCE COMPARISON WITH ML-KNN AND RAKEL

Algorithm Hamming Loss ↓ Accuracyexam↑ Recallexam↑ Fexam↑

BSR 0.0918 0.7611 0.9664 0.8269

ML-kNN 0.1662 0.4221 0.4373 0.4597
RAKEL 0.1608 0.4881 0.5450 0.5473
For each evaluation criterion, ↓ indicates “the smaller the better” whereas ↑
indicates “the larger the better”.

medication label. In severe class-imbalanced labels, the
F-measure was relatively lower than that in class-balanced
labels, which was caused by false positive prediction.
To validate the overall performance of the multi-label
recommendation model, the prediction results were assembled
and compared to two popular multi-label classification
algorithms: ML-kNN and RAKEL. The example-based
measure of the three methods is shown in Table II. The average
accuracy after 10-fold cross validations was 0.7611, the recall
was 0.9664, and the F-measure was 0.8269, which was much
higher than the results of the other two methods. For Hamming
Loss, the BSR was 0.0918; thus, it performed better than the
Fig. 5. Screenshots of the T2DM SDM system interface.
antihyperglycemic medications with their corresponding
benefits and harms (Fig. 5b, orange region).
The region describing the patient’s basic information
includes basic demographic information and current admission
administration information. The patient’s temperature, blood
pressure, pulse rate, weight, HbA1c, and blood sugar values are
displayed in the patient’s current vital sign values region. In the
pink region shown in Fig. 5a, the clinical history of the patient
is displayed as a timeline. If a lab test report or an examination
report is available, the detailed report will be obtained by
clicking on the report name (Fig. 5a, blue region). Physicians

2168-2194 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2016.2614991, IEEE Journal of
Biomedical and Health Informatics
JBHI-00268-2016.R1 6

can use the button in the blue circle shown in Fig. 5a to navigate tailored the option information by giving patients and
to option and decision talk panel (Fig. 5b). Similarly, if physicians an intuitive and clear sense of what decisions may
physicians want to help the patient review his/her clinical be appropriate based on the patient values and guiding them to
history and detailed clinical information during the option talk focus on the specific knowledge that they may need the most.
or decision talk phases, they can click the button in the blue Because glycemic control is an issue that is repeated in every
circle shown in Fig. 5b. The orange region (Fig. 5b) is the main encounter, and changes in a patient’s lifestyle and other
component of the DA tool and includes the following: names of conditions may affect the choice of mellitus medications,
the eight types of antihyperglycemic medications, HbA1c diabetes medication choice should not be established only once
control effects, daily routine, weight effects, hypoglycemia or only depend on “static” knowledge. Compared with a
risks, estimated costs per day, and other side effects. The
well-known T2DM DA, Diabetes Medication Choice [14], the
estimated costs per day are extracted by the Ensemble platform
EHR data utilization in our T2DM SDM system placed more
from the current hospital EHR price list and listed by specific
emphasis on the patient’s current condition, and this
medication, for example “acarbose: 4.92-7.41”. Additionally,
the recommended medication resulting from the BSR is decision-making process could be repeated easily and
highlighted in green. effectively.
DAs are being developed primarily in North America,
IV. DISCUSSION Europe, and Australia. Policy makers in these regions, such as
the National Health Service in the United Kingdom [35] and the
In this study, a T2DM SDM system was developed that
Institute of Medicine in the United States [4], have been making
followed IPDAS criteria and Elwyn’s SDM model to help
efforts to involve patients in decision making. Moreover,
patients and their physicians have a conversation about
Malaysia is a pioneer of SDM promotion in Asia where
antihyperglycemic medications. In addition to knowledge
physicians are aware of SDM but few DAs are used in practice
retrieved from guidelines telling patients the benefits and harms
[36]. In China, the concept of SDM remains unfamiliar to
of each mellitus medication, we utilized patient information in
clinical professionals, let alone DA implementations. Few
hospital EHRs to make a medication recommendation list to
studies have investigated SDM in Chinese patients, and none
provide decision support. We applied an ensemble multi-label
involve patients in Mainland China [37]. Due to the scarce
classification method, BSR, to the real-world EHR data to
healthcare resources in China, a physician may treat more than
achieve the reuse of EHR data and outstanding classifier
100 patients per day; therefore, the communication time is
performance, particularly in the prediction of the minority
limited. SDM with tailored knowledge is an effective method to
positive cases. Although the T2DM SDM system described in
improve the communication between physicians and patients.
this manuscript was designed to be physician-guided, patients
Therefore, the promotion of SDM in China may be an
can still play an active role in the SDM process. Through the
opportunity to enhance the trust between physicians and
clinical history review panel, patients can use the DAs to
patients and make the consultation more pertinent. However,
review their health conditions and realize that they need an
evidence medicine and HIT have been undergoing rapid in
antihyperglycemic medication. In the option talk phase, the DA
recent years, and they have laid the foundation for the
system lists all of the medication options with their usage
promotion of SDM and DA tools. To the best of our knowledge,
guidelines, benefits and harms, whereas in the decision talk
this study constitutes the first attempt for the development of a
phase, patients can talk to their physicians more effectively
DA that was designed and developed for patients in Mainland
according to the specific medications recommended by the
China. When designing the system content and user interfaces,
system and therefore aid the decision-making process. This
Chinese T2DM guidelines and real-world Chinese hospital
SDM process is interactive and would not be successful in the
EHR data were taken into account, as were the HIT utilization
absence of either the physician or patient.
customs in today’s Chinese healthcare.
The relative dependent system framework, compared to the
The T2DM SDM system and the process used in its
DAs embedded into EHRs, has the advantages of feasibility
development have several limitations, and certain work is
and maintenance. When bringing this framework into practice
needed in the future. First, the model was built only from EHR
in the clinic workflow, hospital information technology
data for hospitalized T2DM, and the outcome of the medication
engineers only need to provide certain data interfaces rather
use was not taken into account. However, it is known that
than system redevelopment. Once the model or knowledge
antihyperglycemic medication decisions are also commonly
required for decision support is changed or outdated,
made in an outpatient setting and that the T2DM medication
modification of the T2DM SDM system will not affect the
strategy decision is often based on previous outcomes of
normal operation of the hospital’s EHR system.
medication use. With the future development of HIT in China,
Idiomatically, physicians help patients tailor knowledge in
medical records for outpatient settings and clinical outcome
clinical encounters according to their conditions. However, in
could be better recorded, and the method proposed in this study
recent decades, the “Meaningful Use” Regulation for EHRs
could be easily extended. Second, the proposed
have not only promoted the adoption of EHRs worldwide but
recommendation model and clinical guidelines serve as
has also resulted in a large amount of high-quality EHR data;
decision support basis separately in the T2DM SDM system.
thus, the reuse of EHR data has become an alternative method
Physicians need to combine the knowledge of guidelines and
[33], [34]. The recommendation model applied in this study
the results of the recommendation model to guide the patients

2168-2194 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2016.2614991, IEEE Journal of
Biomedical and Health Informatics
JBHI-00268-2016.R1
make decisions. Although the workflow of the DA, in which
the physicians make the final decision and the crowd wisdom
theory [38] can avoid the risk of wrong decisions to some extent,
the inclusion of a guideline check in the recommendation
model in the future would be an effective way to avoid the risk
of wrong decisions due to wrong predictions from the
recommendation model. Third, the current model can only be
evaluated using statistic methods. Effective implementation of
SDM in China demands urgent realization. To transform this
study from an experimental research project to a practical
application, further efforts are necessary to explore how the
SDM implementation pattern can be adapted to Chinese
conditions. In addition, in the future, an analysis of the T2DM
SDM system proposed in this manuscript after its use in
practical clinical settings would be valuable and promote SDM
development in China.
V. CONCLUSION
This work focused on achieving a direct pathway between
the SDM process and patient data and proposed a SDM system
framework about antihyperglycemic medication choice. A
prototype of this framework was developed, in which
guidelines and evidence studies prepared patients with
preliminary knowledge, and real-world EHR data were utilized
to build a multi-label classification method to provide a
medication recommendation list based on patients’ current
conditions. The applied multi-label classifier performed
outstanding in severe class-imbalanced data compared to
ML-kNN and RAKEL. The T2DM SDM system with decision
support based on reused EHR data provides a more intuitive
understanding of patient conditions and tailors knowledge
about diabetes medications, leading to more effective SDM.
ACKNOWLEDGMENT
The authors would like to thank the editors and the
anonymous reviewers for their help in improving the paper.
REFERENCES
[1] "Diatetes Programme," Jan. 17, 2016; http://www.who.int/diabetes/en/.
[2] Y. Xu, L. Wang, J. He, and et al., “PRevalence and control of diabetes in
chinese adults,” JAMA, vol. 310, no. 9, pp. 948-959, 2013.
[3] S. E. Inzucchi, R. M. Bergenstal, J. B. Buse, M. Diamant, E. Ferrannini,
M. Nauck, A. L. Peters, A. Tsapas, R. Wender, and D. R. Matthews,
“Management of Hyperglycemia in Type 2 Diabetes, 2015: A
Patient-Centered Approach: Update to a Position Statement of the
American Diabetes Association and the European Association for the
Study of Diabetes,” Diabetes Care, vol. 38, no. 1, pp. 140-149, January 1,
2015, 2015.
[4] E. S. Huang, S. E. S. Brown, B. G. Ewigman, E. C. Foley, and D. O.
Meltzer, “Patient Perceptions of Quality of Life With Diabetes-Related
Complications and Treatments,” Diabetes Care, vol. 30, no. 10, pp.
2478-2483, October 1, 2007, 2007.
[5] A. C. Larme, and J. A. Pugh, “Attitudes of primary care providers toward
diabetes: barriers to guideline implementation,” Diabetes care, vol. 21, no.
9, pp. 1391-1396, 1998.
[6] J. G. Carroll, “Crossing the Quality Chasm: A New Health System for the
21st Century,” Quality Management in Healthcare, vol. 10, no. 4, pp.
60-61, 2002.
2168-2194 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
7
[7] C. Rathert, M. D. Wyrwich, and S. A. Boren, “Patient-Centered Care and
Outcomes: A Systematic Review of the Literature,” Medical Care
Research and Review, vol. 70, no. 4, pp. 351-379, August 1, 2013, 2013.
[8] M. J. Barry, and S. Edgman-Levitan, “Shared decision making—the
pinnacle of patient-centered care,” N Engl J Med, vol. 366, no. 9, pp.
780-781, 2012.
[9] G. Elwyn, C. Dehlendorf, R. M. Epstein, K. Marrin, J. White, and D. L.
Frosch, “Shared decision making and motivational interviewing:
achieving patient-centered care across the spectrum of health care
problems,” Ann Fam Med, vol. 12, no. 3, pp. 270-275, May 1, 2014,
2014.
[10] A. M. O'Connor, C. L. Bennett, D. Stacey, M. Barry, N. F. Col, K. B.
Eden, V. A. Entwistle, V. Fiset, M. Holmes-Rovner, S. Khangura, H.
Llewellyn-Thomas, and D. Rovner, “Decision aids for people facing
health treatment or screening decisions,” Cochrane Database Syst Rev, no.
3, pp. CD001431, 2009.
[11] G. Elwyn, A. O'Connor, D. Stacey, R. Volk, A. Edwards, A. Coulter, R.
Thomson, A. Barratt, M. Barry, S. Bernstein, P. Butow, A. Clarke, V.
Entwistle, D. Feldman-Stewart, M. Holmes-Rovner, H.
Llewellyn-Thomas, N. Moumjid, A. Mulley, C. Ruland, K. Sepucha, A.
Sykes, and T. Whelan, “Developing a quality criteria framework for
patient decision aids: online international Delphi consensus process,”
BMJ, vol. 333, no. 7565, pp. 417, 2006-08-24 22:00:11, 2006.
[12] G. Elwyn, A. M. O'Connor, C. Bennett, R. G. Newcombe, M. Politi, M.-A.
Durand, E. Drake, N. Joseph-Williams, S. Khangura, A. Saarimaki, S.
Sivell, M. Stiel, S. J. Bernstein, N. Col, A. Coulter, K. Eden, M. Härter, M.
H. Rovner, N. Moumjid, D. Stacey, R. Thomson, T. Whelan, T. van der
Weijden, and A. Edwards, “Assessing the quality of decision support
technologies using the International Patient Decision Aid Standards
instrument (IPDASi),” PLoS ONE, vol. 4, no. 3, pp. e4705, 2009.
[13] M. Coylewright, K. Shepel, A. LeBlanc, L. Pencille, E. Hess, N. Shah, V.
M. Montori, and H. H. Ting, “Shared Decision Making in Patients with
Stable Coronary Artery Disease: PCI Choice,” PLoS ONE, vol. 7, no. 11,
pp. e49827, 2012.
[14] K. Kaiser, W. Y. Cheng, S. Jensen, M. L. Clayman, A. Thappa, F.
Schwiep, A. Chawla, J. J. Goldberger, N. Col, and J. Schein,
“Development of a shared decision-making tool to assist patients and
clinicians with decisions on oral anticoagulant treatment for atrial
fibrillation,” Current Medical Research and Opinion, vol. 31, no. 12, pp.
2261-2272, 2015/12/02, 2015.
[15] E. M. DeWitt, E. A. Lipstein, K. Staun, L. Scherer, J. Taylor, C. M.
Lannon, and W. B. Brinkman, “A178: development of tools to facilitate
shared decision making about medications for juvenile idiopathic
arthritis—a project of the pediatric rheumatology care and outcomes
improvement network,” Arthritis Rheumatol, vol. 66, pp. S232-S233,
2014.
[16] M. Breslin, R. J. Mullan, and V. M. Montori, “The design of a decision
aid about diabetes medications for use during the consultation with
patients with type 2 diabetes,” Patient Educ Couns, vol. 73, 2008.
[17] S. Tamhane, R. Rodriguez-Gutierrez, I. Hargraves, and V. Montori,
“Shared decision-making in diabetes care,” Curr Diab Rep, vol. 15, no. 12,
pp. 1-10, 2015/10/12, 2015.
[18] R. J. Mullan, V. M. Montori, N. D. Shah, T. J. Christianson, S. C. Bryant,
G. H. Guyatt, L. I. Perestelo-Perez, R. J. Stroebel, B. P. Yawn, and V.
Yapuncich, “The diabetes mellitus medication choice decision aid: a
randomized trial,” Arch Intern Med, vol. 169, 2009.
[19] A. LeBlanc, K. L. Ruud, M. E. Branda, K. Tiedje, K. R. Boehmer, L. J.
Pencille, H. Van Houten, M. Matthews, N. D. Shah, C. R. May, B. P.
Yawn, and V. M. Montori, “The impact of decision aids to enhance
shared decision making for diabetes (the DAD study): protocol of a
cluster randomized trial,” BMC Health Serv Res, vol. 12, no. 1, pp. 1-9,
2012.
[20] S. E. Brossette, A. P. Sprague, J. M. Hardin, K. B. Waites, W. T. Jones,
and S. A. Moser, “Association Rules and Data Mining in Hospital
Infection Control and Public Health Surveillance,” J Am Med Inform
Assoc, vol. 5, no. 4, pp. 373-381, 1998.
[21] E. H. Shortliffe, and A. Abu-Hanna, “Clinical prognostic methods: trends
and developments,” J Biomed Inform, vol. 48, pp. 1-4, Apr, 2014.
[22] P. A. Nguyen, S. Syed-Abdul, U. Iqbal, M.-H. Hsu, C.-L. Huang, H.-C. Li,
D. L. Clinciu, W.-S. Jian, and Y.-C. J. Li, “A Probabilistic Model for
Reducing Medication Errors,” PLoS ONE, vol. 8, no. 12, pp. e82401,
2013.
[23] J. S. Mathias, A. Agrawal, J. Feinglass, A. J. Cooper, D. W. Baker, and A.
Choudhary2, “Development of a 5 year life expectancy index in older
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2016.2614991, IEEE Journal of
Biomedical and Health Informatics
JBHI-00268-2016.R1
adults using predictive mining of electronic health record data,” J Am
Med Inform Assoc, vol. 20, no. e1, pp. e118-e124, 2013.
[24] I. Batal, D. Fradkin, J. Harrison, F. Moerchen, and M. Hauskrecht,
“Mining recent temporal patterns for event detection in multivariate time
series data,” in Proceedings of the 18th ACM SIGKDD international
conference on Knowledge discovery and data mining, Beijing, China,
2012, pp. 280-288.
[25] G. Elwyn, D. Frosch, R. Thomson, N. Joseph-Williams, A. Lloyd, P.
Kinnersley, E. Cording, D. Tomson, C. Dodd, S. Rollnick, A. Edwards,
and M. Barry, “Shared decision making: a model for clinical practice,” J
Gen Intern Med, vol. 27, no. 10, pp. 1361-1367, 2012/10/01, 2012.
[26] Z. Min-Ling, and Z. Zhi-Hua, “A review on multi-label learning
algorithms,” IEEE Trans Knowl Data Eng, vol. 26, no. 8, pp. 1819-1837,
2014.
[27] S. Daskalaki, I. Kopanas, and N. Avouris, “Evaluation of classifiers for an
uneven class distribution problem,” Appl Artif Intell, vol. 20, no. 5, pp.
381-417, 2006/06/01, 2006.
[28] M. R. Boutell, J. Luo, X. Shen, and C. M. Brown, “Learning multi-label
scene classification,” Pattern Recognition, vol. 37, no. 9, pp. 1757-1771,
9//, 2004.
[29] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer,
“SMOTE: synthetic minority over-sampling technique,” Journal of
artificial intelligence research, vol. 16, pp. 321-357, 2002.
[30] L. Breiman, “Random forests,” Mach Learn, vol. 45, no. 1, pp. 5-32.
[31] M.-L. Zhang, and Z.-H. Zhou, “ML-KNN: A lazy learning approach to
multi-label learning,” Pattern Recognition, vol. 40, no. 7, pp. 2038-2048,
7//, 2007.
2168-2194 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
8
[32] G. Tsoumakas, and I. Vlahavas, "Random k-Labelsets: An Ensemble
Method for Multilabel Classification," Machine Learning: ECML 2007,
Lecture Notes in Computer Science J. Kok, J. Koronacki, R. Mantaras, S.
Matwin, D. Mladenič and A. Skowron, eds., pp. 406-417: Springer Berlin
Heidelberg, 2007.
[33] Z. Huang, W. Dong, L. Ji, and H. Duan, “Outcome Prediction in Clinical
Treatment Processes,” Journal of Medical Systems, vol. 40, no. 1, pp.
1-13, 2015.
[34] K. Ng, A. Ghoting, S. R. Steinhubl, W. F. Stewart, B. Malin, and J. M.
Sun, “PARAMO: A PARAllel predictive MOdeling platform for
healthcare analytic research using electronic health records,” Journal Of
Biomedical Informatics, vol. 48, pp. 160-170, Apr, 2014.
[35] G. Elwyn, S. Laitner, A. Coulter, E. Walker, P. Watson, and R. Thomson,
“Implementing shared decision making in the NHS,” BMJ, vol. 341,
2010-10-14 23:09:55, 2010.
[36] C.-J. Ng, P.-Y. Lee, Y.-K. Lee, B.-H. Chew, J. P. Engkasan, Z.-I. Irmi,
N.-S. Hanafi, and S.-F. Tong, “An overview of patient involvement in
healthcare decision-making: a situational analysis of the Malaysian
context,” BMC Health Serv Res, vol. 13, no. 1, pp. 1-7, 2013.
[37] R. Huang, M. R. Gionfriddo, L. Zhang, A. L. Leppin, H. H. Ting, and V.
M. Montori, “Shared decision-making in the People’s Republic of China:
current status and future directions,” Patient Prefer Adherence, vol. 9, pp.
1129-1141, 08/06, 2015.
[38] J. Surowiecki, and M. P. Silverman, “The wisdom of crowds,” American
Journal of Physics, vol. 75, no. 2, pp. 190-192, 2007.