POPULATION HEALTH MANAGEMENT

Volume 22, Number 3, 2019

Mary Ann Liebert, Inc.
DOI: 10.1089/pop.2018.0129

Transforming Diabetes Care Through Artificial Intelligence:

The Future Is Here

Irene Dankwa-Mullan, MD, MPH,1 Marc Rivo, MD, MPH,2 Marisol Sepulveda, DO, MPH,3
Yoonyoung Park, ScD,4 Jane Snowdon, PhD,5 and Kyu Rhee, MD, MPP6

Abstract

An estimated 425 million people globally have diabetes, accounting for 12% of the world’s health expen-
ditures, and yet 1 in 2 persons remain undiagnosed and untreated. Applications of artificial intelligence (AI) and
cognitive computing offer promise in diabetes care. The purpose of this article is to better understand what AI
advances may be relevant today to persons with diabetes (PWDs), their clinicians, family, and caregivers. The
authors conducted a predefined, online PubMed search of publicly available sources of information from 2009
onward using the search terms ‘‘diabetes’’ and ‘‘artificial intelligence.’’ The study included clinically-relevant,
high-impact articles, and excluded articles whose purpose was technical in nature. A total of 450 published
diabetes and AI articles met the inclusion criteria. The studies represent a diverse and complex set of innovative
approaches that aim to transform diabetes care in 4 main areas: automated retinal screening, clinical decision
support, predictive population risk stratification, and patient self-management tools. Many of these new AI-
powered retinal imaging systems, predictive modeling programs, glucose sensors, insulin pumps, smartphone
applications, and other decision-support aids are on the market today with more on the way. AI applications have
the potential to transform diabetes care and help millions of PWDs to achieve better blood glucose control,
reduce hypoglycemic episodes, and reduce diabetes comorbidities and complications. AI applications offer
greater accuracy, efficiency, ease of use, and satisfaction for PWDs, their clinicians, family, and caregivers.

Keywords: diabetes care, artificial intelligence, cognitive computing, artificial pancreas, retinal imaging,
glucose monitoring

Background amputations, adult-onset blindness, and almost doubles the

risk of heart attack and all-cause mortality, leading to hos-
pitalization, long-term complications, and higher costs.3
D iabetes is a global pandemic. An estimated 425
million people worldwide have diabetes, accounting for
12% of the world’s health expenditures, and yet 1 in 2 persons
Decades of well-designed studies have established that
intensive therapy effectively delays the onset and slows the
remain undiagnosed and untreated.1 Type 2 diabetes is driven progression of diabetes-related complications, such as reti-
by the global obesity epidemic and a sedentary lifestyle that nopathy, nephropathy, and neuropathy.4 Yet, a recent study
overwhelms the body’s internal glucose control requiring of 300,000 patients with type 2 diabetes who were started on
exogenous insulin.2 Millions of newborns are born to medical therapy found that after 3 months, 31% of patients
mothers with gestational diabetes. Children born with type 1 had discontinued their diabetes medications altogether: this
diabetes mellitus, in which the body cannot produce insulin, increased to 44% by 6 months, and to 58% by 1 year. Only
require life-long insulin therapy. In the United States, dia- 40% eventually restarted diabetes medications.5 Optimal
betes is the leading cause of kidney failure, lower limb care for persons with diabetes (PWDs) often is hampered by
1
IBM Corporation, Watson Health, Bethesda, Maryland.
2
Population Health Innovations, Inc., Miami Beach, Florida.
3
Occupational Medicine, University of Iowa, Iowa City, Iowa.
4
IBM Corporation, IBM Research, Cambridge, Massachusetts.
5
IBM Corporation, Watson Health, Yorktown Heights, New York.
6
IBM Corporation, Watson Health, Cambridge, Massachusetts.
ª Irene Dankwa-Mullan et al. 2018; Published by Mary Ann Liebert, Inc. This Open Access article is distributed under the terms of the
Creative Commons License ([Link] which permits unrestricted use, distribution, and reproduction in
any medium, provided the original work is properly cited.

229
230 DANKWA-MULLAN ET AL.

the absence of real-time, key health information necessary systematic review of clinically relevant diabetes AI applica-
to make informed choices associated with intensive therapy tions is missing. The purpose of this article is to better un-
and tight diabetes control. Although advances in technology derstand what meaningful AI advances may be relevant today
offer unprecedented and inexpensive access to essential in- to PWDs, their primary care clinicians, endocrinologists,
formation for many individuals in many fields, its impact in health professionals, family, and caregivers.
the care of patients with diabetes seems rather limited. The
challenges of real-time diabetes care information are com- Methods
pounded by the rapid expansion of medical knowledge. The
The study team conducted a predefined, online PubMed
index of biomedical literature contains more than 28 million
search of publicly available sources of information using the
articles as of June 2018 and is growing at a rate of more than
search terms ‘‘diabetes’’ and ‘‘artificial intelligence (AI).’’
850,000 new citations each year.6 Each person will generate
To identify articles with clinically-relevant, high-impact
more than 1 million gigabytes of health-related data in his or
diabetes AI applications, the team excluded manuscripts
her lifetime, the equivalent of about 300 million books. An
with publication dates before 2009 and those whose purpose
estimated 80% of health data is unstructured. This in-
was primarily technical in nature (eg, focused solely on AI
cludes clinician notes, clinical trials, hospital records and
algorithm development). The first-pass search identified a
discharge summaries, imaging and laboratory reports, and
total of 763 clinically-relevant abstracts. Additional review
nonclinical data sources, including device and sensor data
excluded 313 as duplicative or primarily technical. The
(often referred to as Internet of Things data), genomic
second-pass review yielded a total of 450 unique, clinically-
data, and social determinants of health data.7 Ninety per-
relevant articles researching the direct application of AI in
cent of a person’s health outcomes may be attributed to
diabetes prevention, diagnosis, and treatment. The infor-
genomics and exogenous data, underscoring the impor-
mation was then collated and classified. The research was
tance of PWDs and their clinicians collecting and
conducted between March and May of 2018.
leveraging these data to make informed health choices.8
Rapid advances in artificial intelligence (AI) offer the
Results
promise of making both real-time structured and unstruc-
tured health data available for the care of PWDs. The Turing The PubMed search yielded a total of 450 clinically-relevant
Archive for the History of Computing defines AI as ‘‘the and high-impact articles published in the last decade related to
science of making computers do things that require intelli- the field of applied AI in diabetes care. The AI applications
gence when done by humans.’’9 AI covers a broad range of aimed to improve a broad spectrum of diabetes care, from
approaches to simulating human intelligence and perform- diabetes screening and detection to monitoring and treatment,
ing various reasoning tasks, such as visual perception, speech and included apps, devices, and systems that aid patients, cli-
recognition, analytics, decision making, and translation be- nicians, and health systems. The published articles included in
tween languages. Cognitive systems employ the spectrum of this search were of high clinical impact in that they sought to
AI approaches to extend and scale human knowledge and produce and test AI approaches that may impact diabetes care
expertise by enabling humans to leverage vast knowledge significantly in the areas of access, accuracy, efficiency, af-
sources rapidly to solve problems. fordability, speed, and satisfaction of patients, clinicians, and
Today, AI is harnessing massive amounts of vital infor- caregivers. A review of the high-impact articles suggests
mation to meet consumer demand in every business, including that AI applications are aiming to transform diabetes care in
health care. A 2017 survey found that 68% of mobile health 4 main areas: automated retinal screening, clinical decision
app developers and publishers believe that diabetes continues support, predictive population risk stratification, and patient
to be the single most important health care field with the best self-management tools, as summarized in Table 1.
market potential for digital health solutions within the near A diverse and complex set of AI approaches and cogni-
future, and that 61% see AI as the most disruptive technology tive computing systems were employed in these studies.
shaping the digital health sector.10 Although advances in AI Table 2 defines the more common AI approaches described
for health care are being reported in the literature11 and new in the research and lists their clinical applications in diabetes
AI-powered devices are being approved for diabetes care,12 a care.

Table 1. Categorization of Artificial Intelligence and Diabetes Care

Category Number of articles Most common clinical AI applications
Automated Retinal Screening 96 Detection of diabetic retinopathy, maculopathy, exudates,
and other abnormalities from normal findings
Clinical Decision Support 126 Detection and monitoring of diabetes and comorbidities
such as neuropathy, nephropathy and wounds
Predictive Population Risk Stratification 135 Identification of diabetes subpopulations at higher risk for
complications, hospitalization, and readmissions
Patient Self-Management Tools 94 AI-improved glucose sensors, artificial pancreas, activity
and dietary tracking devices
TOTAL 450
AI, artificial intelligence.
 Table 2. Common Artificial Intelligence Approaches Used in Diabetes Care
Method How it works Strengths Limitations Application area
Multilayer Composed of neurons in input layer, output Can model complex nonlinear Greater number of parameters Prediction models, patient
perceptron layer, and multiple hidden layers. Neurons relationship have to be estimated self-management tools
in each layer are connected to all neurons without convolution
in the next layer, making each layer fully Less effective than many
connected to the next. other deep models
Learns by ‘‘backpropagation’’ method
Convolutional neural Composed of multiple layers of neurons with Can model complex nonlinear Require a large amount of Retinal screening
network (CNN) the convolution layer having neurons that relationship data to train
look at small patches of the input image at Ideal for image, audio, video Computationally intensive
a time, like a filter, and are convolved Many parameters require fine
across the whole input image and share tuning while training the
parameters across the image. model
Learns by ‘‘backpropagation’’ method
Each layer of the CNN detects the presence
of specific features across space, detecting
more high-level features as moving
forward
Random forest Creates an ensemble of decision trees Easy to fit, generally produces Can be slow in prediction Retinal screening, decision
In each tree, a random set of features are good performance Only applicable for discrete support, prediction models,
considered for determining root nodes and Can be used for both outcomes - if outcome is patient self-management

231
splits classification and continuous, it must be tools
regression problems categorized
Can measure relative Difficult to interpret
importance of features
easily
Robust to outliers and avoids
overfitting, given sufficient
sample size
Fuzzy logic/fuzzy Provides a probability value between 0 and 1 Resembles human reasoning Requires an expert curation of Retinal screening, decision
system rather than deterministic decision (0 or 1) High interpretability rules support, sensors and
for membership in a certain class Easy to modify rules artifical pancreas
Does not require large data
Support vector Classification method for binary outcomes Performs well in nonlinear Does not scale well to large Retinal screening, decision
machine (SVM) (not often used for multiclass problems, decision boundaries data support, prediction models,
but techniques for multiclass SVM exist) Robust to overfitting Hard to interpret patient self-management
Works by adding data to a high-dimensional tools
space and finds a hyperplane that separates
2 classes best (that maximizes the distance
between the plane and nearby data points,
or margin)
(continued)
 Table 2. (Continued)
Method How it works Strengths Limitations Application area
Logistic regression Classification method for binary outcome Easy to fit, efficient, and Only binary classification Prediction models
Predicts the probability of an outcome (0 or scalable Can be sensitive to outliers
1) based on the features Supported by most standard Requires transformation of
Learns the model coefficients by maximum software nonlinear features
likelihood estimation Can obtain probability of the
Finds a line or hyperplane that best outcome, which can be
represents the data points useful
Natural language Computational tools and methods to process, Critical in building intelligent Usually requires a large Prediction models
processing analyze, and perform inference of human machine and human– amount of human-
languages computer interactions annotated records to train

232
Can process and analyze free-
text information such as
electronic physician notes
K-nearest neighbors Categorizes input data into several classes Does not make assumptions Computationally intensive Retinal screening, decision
algorithm using its k nearest neighbors about underlying Sensitive to outliers or support, prediction models,
distribution localized data patient self-management
Can be used for both tools
classification and
regression problems
Easy to understand and
implement
 Table 3. Summary of Selected Key Diabetes Artificial Intellegence Studies and Description of Outcomes
Training data/validation Model performance
Author, date Title Learning model data/features Testing data/features Study outcomes application
Retinal Screening
Gulshan V. Development and Deep CNN Data set: 128,175 retinal images Test data: EyePACS-1 data Sensitivity: 97.5% for EyePACS-1, Deep machine learning
201618 Validation of a Deep Ground truth: images graded 3 to set (n = 9963 images from 96.1% for Messidor-2 algorithm had high
Learning Algorithm for 7 times for DR, diabetic 4997 patients) Specificity: 93.4% for EyePACS-1, sensitivity and
Detection of Diabetic macular edema, and image Messidor-2 data set (n = 1748 93.9% for Messidor-2 specificity for
Retinopathy in Retinal gradability by a panel of 54 images from 874 patients) ROC: 0.991 (95% CI, 0.988– detecting referable
Fundus Photographs US licensed ophthalmologists 0.993) for EyePACS-1, 0.990 DR.
(95% CI, 0.986–0.995) for
Messidor-2
Rahim S. 201420 Detection of Diabetic Fuzzy image Dataset: public data sets Test data: remaining 10% For k-NN, polynomial SVM, RBF Fuzzy image
Retinopathy and processing, ML (DIARETDB0, DIARETDB1, data SVM, and NB, respectively processing together
Maculopathy classifiers (1- MESSIDOR, DRIVE, STARE, Sensitivity: 0.87, 0.45, 0.92, 0.91 with the retinal
Using Fuzzy Image nearest REVIEW, ROC) Specificity: 1.00, 0.98, 0.94, 0.75 structure extraction
Processing neighbour, NB, New data set of 600 images from Accuracy: 0.93, 0.7, 0.93, 0.75 in DR screening can
SVM) 300 patients in Malaysia - help produce a more
oversampled to a total of 990 reliable and efficient
images screening system
Training data: 90% of data set

233
Lam C. 201821 Retinal Lesion Detection CNN (GoogLeNet) Data set: manually created image Test data: public image data In validation using the patch Regionally trained
With Deep Learning patches from public image set (eOphta) (n = 463) images: CNNs can detect
Using Image Patches data set (Kaggle retinopathy Ground truth: labeled by 2 Accuracy 98%, ROC 99% with and distinguish
data subset, n = 243) ophthalmologist at pixel GoogLeNet between subtle
Ground truth: labeled by 2 level (3 classes) For microaneurysm, exudates pathologic lesions
ophthalmologists respectively in test data: with only a few
Training data:1050 patches Pixel-level ROC: 0.94, 0.95 hundred training
Validation: 274 patches Precision and recall AUC for examples per lesion.
detection task: 0.86, 0.64
Keel S. 201823 Feasibility Deep CNN Data set: public data set Test data: data from 96 Sensitivity: 92.3% AI-based DR screening
and Patient (Inception v3) (LabelMe, n = 66,790) participants who agreed to Specificity: 93.7% appears to be
Acceptability of a Ground truth: Images graded by receive both retinal 96% of participants reported that feasible, accurate,
Novel Artificial 21 ophthalmologists screening approaches and they were either satisfied or very and well accepted
Intelligence-based Training data: 58,790/66,790 complete a questionnaire satisfied with the automated by patients attending
Screening Validation: 8000/66,790 Ground truth: compared to screening model endocrinology
Model for Diabetic manual screening result by 78% reported that they preferred outpatient settings.
Retinopathy at ophthalmologist the automated model over
Endocrinology manual.
Outpatient Services
(continued)
 Table 3. (Continued)
Training data/validation Model performance
Author, date Title Learning model data/features Testing data/features Study outcomes application
Predictive Modeling and Risk
Stratification
Han L. 201525 Rule Extraction from Ensemble learning Data set: China Health and Test data: remaining 10% of For positive cases: The proposed hybrid
Support Vector using SVM and Nutrition Survey data data Precision: 89.6% system can provide
Machines Using RF rule (n = 7913, 646 diabetic) Recall: 44.3% a tool for the
Ensemble Learning extraction Training data: 90% of dataset F-score: 0.593 diagnosis of
Approach: An Validation: 10-fold cross- For all cases: diabetes from
Application for validation for model Weighted average precision: population-based
Diagnosis of Diabetes parameter selection 94.2% nutritional surveys,
15 features selected using Weighted average recall: 93.9% and it supports a
univariate LR, chi-square second opinion for
tests, information gain-based lay users
method, and RF
Shankaracharya. Computational Mixture of expert Data set: 1415 subjects (947 Test data: 311/1415 Best result achieved The proposed tool for
201226 Intelligence-based system based on diabetic) Sensitivity: 99.5% identifying
Diagnosis Tool for the MLP Training data: 1104/1415 Specificity: 99.07% individuals with
Detection of Accuracy: 99.36% prediabetes,
Prediabetes and Type 2 diabetes, and
Diabetes in India nondiabetes is
highly accurate and
may be used for
large-scale diabetic
screenings.

234
Wei WQ. 201027 A High Throughput NLP, SVM, and Data set: 57,707 electronic notes No separate test data were F-score for cases: 0.956 The proposed approach
Semantic Concept semantic from 1600 DM patients and specified F-score for controls: 0.957 is accurate and
Frequency Based knowledge 1600 control patients in Mayo Precision for cases: 0.968. responsive to the
Approach for Patient Clinic Semantic knowledge: varying urgent need to
Identification: A Case Validation: 10-fold cross- degrees of F-score, precision, develop a general
Study Using Type 2 validation for model selection and recall values reported. automatic approach
Diabetes Mellitus Features: Semantic concept units for diabetic patient
Clinical Notes extracted from notes and case-finding and
classified into semantic type characterization.
groups
Corey KE. Development and LR with adaptive Data set: electronic medical Test data: randomly selected Specificity; 91% The NAFLD
201628 Validation of an LASSO records from 620 patient 611 high-risk patients Sensitivity: 51% classification
Algorithm to Identify randomly selected from the identified by classification PPV: 89% algorithm is superior
Nonalcoholic Fatty high-risk patients in Partners algorithm NPV: 56% to ICD-9 billing data
Liver Disease Healthcare Additional validation: AUC: 0.85 (compared to 0.75 alone. This
(NAFLD) in the Ground truth: compared to chart independent test set of using ICD-9 billing codes only, approach is simple
Electronic Medical review by a hepatologist 314,292 patients P < 0.0001) to develop, deploy,
Record Features: laboratory Ground truth: 100 random and can be applied
measurements, diagnosis positive case record review across different
codes, and concepts extracted institutions to create
from medical notes EMR-based cohorts
of individuals with
NAFLD.
(continued)
 Table 3. (Continued)
Training data/validation Model performance
Author, date Title Learning model data/features Testing data/features Study outcomes application
Neves J. 201529 A Soft Computing Logic Program- Data set: data from 558 total Test data: remaining 1/3 of ANN performance in test data set The proposed model
Approach to Kidney ming, ANN patients (175 diagnosed with data Sensitivity: 93.19% showed good
Diseases Evaluation CKD) Specificity: 91.9% performance in
Training data: 2/3 of data set PPV: 84.4% predicting the
Clinical information about CKD NPV: 96.6% likelihood of a CKD
as rewritten into Logic diagnosis
Programming algorithms, and
its terms as training and test
sets of ANN
Features: 24 variables grouped
into 5 categories
Rau HH. 201630 Development of a Web- ANN, LR Data set: data from 2060 diabetic Test data: 618/2060 ANN performance was superior to Data mining systems
based Liver Cancer patients in the National Health that of LR for predicting enable clinicians to
Prediction Model for Insurance Research Database diabetics who will be diagnosed predict those
Type II Diabetes (NHIRD) of Taiwan with liver cancer in the next 6 diabetics at greater
Patients by Using an Training data: 1442/2060 years. risk for the
Artificial Neural Sensitivity: 0.757 development of liver
Network Specificity: 0.755 cancer.
AUC: 0.873
Vyas R. 201632 Building and Analysis of SVM Training data: positive and Test data: 129 proteins Accuracy: 78.20% This integrated
Protein-Protein negative proteins from PDB extracted via text mining Precision: 68.26% approach has a
Interactions Related to and UniProt databases from literature AUC: 0.788 potential to identify

235
Diabetes Mellitus (n = 2653) disease-related
Using Support Vector proteins, functional
Machine, Biomedical annotation, and
Text Mining and other proteomics
Network Analysis studies.
López B. 201833 Single Nucleotide Random forest, k- Data set: data from 677 subjects Test data: 10-fold cross- For risk prediction RF is a useful method
Polymorphism (SNP) NN (248 diabetic), each containing validation used. No AUC: 0.89 for learning
Relevance Learning 96 SNPs regarding type 2 separate test data were RF outperformed SVM and LR in predictive models to
with Random Forests diabetes specified terms of prediction accuracy help physicians to
for Type 2 Diabetes Features: SNP data, clinical and stability of the estimated identify the relevant
Risk Prediction information, SNP value relevance SNPs associated
relevance with and predictive
of type 2 diabetes.
Clinical Decision Support
Lo-Ciganic WH. Using Machine Learning Random survival Data set: 33,130 non-dual- Test data: remaining 10% The adherence thresholds most Machine learning
201534 to Examine Medication forests, survival eligible Medicaid enrollees data discriminating for risk of all- approaches hold
Adherence Thresholds trees models with type 2 diabetes cause hospitalization promise as an
and Risk of Training data: 90% of data set varied from 46% to 94% - the intuitive and
Hospitalization Features: sociodemographics, widely used 80% adherence powerful approach
measures of service use, health threshold is not optimal for for customizing
status, diabetes treatment predicting risk of hospitalization interventions in
intensity medication
adherence in
diabetics and
optimizing health
outcomes.
(continued)
 Table 3. (Continued)
Training data/validation Model performance
Author, date Title Learning model data/features Testing data/features Study outcomes application
Shu T. 201735 An Extensive Analysis of k-NN, SVM with 8 Data set: 284 diabetes mellitus Test data: 10-fold cross- The best texture feature Compared with
Various Texture image extractor and 231 healthy samples validation used. No extractor, Image Gray-scale traditional
Feature Extractors to methods separate test data were Histogram (bin n = 256), diagnostic methods
Detect Diabetes specified combined with SVM that rely on blood
Mellitus Using Facial Sensitivity: 99.64% samples, the Image
Specific Regions Specificity: 98.26% Gray-scale
Accuracy: 99.02% Histogram is a
highly accurate,
non-invasive way to
diagnose diabetes
using facial and
tongue features.
Katigari KM. Fuzzy Expert System for Fuzzy expert Data set: diagnostic parameters Test data: 213 medical For diagnosis and severity of The fuzzy expert
201736 Diagnosing Diabetic system and their importance records of patients diabetic neuropathy system can help
Neuropathy developed by specialists used diagnosed with diabetic Sensitivity: 89% diagnose and
to develop fuzzy expert neuropathy Specificity: 98% determine the
system Accuracy 93% severity of diabetic
neuropathy.
Wang L. 201737 Area Determination of Two-stage SVM Data set: 100 foot ulcer images Test data: cross-validation Sensitivity: 73.3% Computer-based

236
Diabetic Foot Ulcer with simple from 15 patients used. No separate test data Specificity = 94.6% systems provide
Images Using a linear iterative were specified high performance
Cascaded Two-Stage clustering and rates for measuring
SVM-Based conditional diabetic wounds and
Classification random fields monitoring wound
healing status, and
are sufficiently
efficient for
smartphone-based
image analysis.
Glucose Sensors and Artificial Pancreas
Mauseth R. Testing of an Artificial Fuzzy Logic N/A Total 17 meal, 13 exercise FLC v2.1 showed improvements in Stress testing the AP
201538 Pancreas System With Controller studies in 10 subjects with mean blood glucose after pizza system followed by
Pizza and Exercise systems (FLC) type 1 diabetes (T1D) consumption, after exercise adjustments to the
Leads to Improvements FLC v2.0 test: 9 meal and 4 testing, in reducing dosing matrix
in the System’s Fuzzy exercise studies with FLC hyperglycemia, and percentage significantly
Logic Controller v2.0, followed by interim time spent in euglycemic range improved FLC
analysis performance when
FLC v2.1 test: remaining 8 retested for mean
meal and 9 exercise studies blood glucose, high
using updated FLC blood glucose, and
normal blood
glucose
(continued)
 Table 3. (Continued)
Training data/validation Model performance
Author, date Title Learning model data/features Testing data/features Study outcomes application
Ling SH. 201239 Natural Occurrence of Fuzzy reasoning Data set: 16 type 1 diabetic Test data: remaining 269 data Advanced noctural hypoglycemic The proposed system
Nocturnal model with patients points from 8/16 patients episode detection offers a noninvasive
Hypoglycemia hybrid particle Training data: 320 data points Sensitivity: 85.7% means to detect
Detection Using Hybrid swarm from 8/16 patients Specificity: 79.8% Hypoglycemic hypoglycemic
Particle Swarm optimization episodes detection episodes in type 1
Optimized Fuzzy with wavelet Sensitivity: 80.0% diabetic patients.
Reasoning Model mutation Specificity: 55.1%
Herrero P. Advanced Insulin Bolus Combination of N/A In silico testing using Using CBR(R2R), mean blood The proposed
40
2015 Advisor Based on R2R and CBR commercial type 1 diabetes glucose improved in both adult smartphone system
Run-To-Run Control simulator generated 1- and adolescent populations and keeps the simplicity
and Case-Based month data for 10 adults hypoglycemia was completely of a standard bolus
Reasoning and 10 adolescents eliminated (R2R alone was not calculator while
scenarios able to do it in the adolescent enhancing its
population) performance by
providing more
adaptability and
flexibility.
DeJournett L. In Silico Testing of an Knowledge-based N/A In silico analysis: 126 000 On average, time in control range An AI-based closed-
201641 Artificial-Intelligence- system unique 5-day simulations was 94.2%, time in range 70– loop glucose

237
Based Artificial resulting in 107 million 140 mg/dl was 97.8%, time in controller may be
Pancreas Designed for glucose values hyperglycemic range was 2.1%, able to improve on
Use in the Intensive time in hypoglycemic range was the results achieved
Care Unit Setting 0.09% by currently existing
Average coefficient of variation: ICU-based PID/
11.1% MPC controllers
Patient Diabetes Self-Management Tools
Zhang W. 2015. ‘‘Snap-n-Eat’’: Food SVM Data set: 2000 food images Test data: 5-fold cross Accuracy: 85% The proposed
44
Recognition and comprising 15 predefined validation smartphone mobile
Nutrition Estimation on categories system can
a Smartphone Ground truth: manual annotation recognize food
items present on a
plate and estimates
their calorific and
nutrition content,
automatically
helping diabetic
patients make more
informed food
choice decisions.
(continued)
 Table 3. (Continued)
Training data/validation Model performance
Author, date Title Learning model data/features Testing data/features Study outcomes application
Cvetković B. Activity Recognition for Ensemble of Data set: average 11 hours of Test data: second week of Best result achieved by Multi- Smartphone sensors
201645 Diabetic Patients Using models (SVM, phone and 7.5 hours of ECG recordings Classifier Adaptive Training using machine
a Smartphone J48, random recordings per day for 2 weeks (MCAT) method learning and
forest, Jrip, from 9 healthy volunteers Accuracy: 83.4% symbolic reasoning
AdaBoost and Training data: first week of F-score: 0.82 can recognize and
Bagging recordings quantify high-level
algorithms), Features (if present): sound, lifestyle activities of
symbolic rules location, acceleration, heart- diabetic patients and
rate, respiration-rate help them make
more informed
activity choices.
Wang L. 201546 Smartphone-based Wound Image boundary N/A 30 simulated wound images, Visual evaluation for simulated The proposed
Assessment System for detection: mean- 34 actual patient wound images smartphone camera
Patients with Diabetes shift images Matthews Correlation Coefficient: system enables
segmentation 0.736 diabetic patients and

238
algorithm their caregivers to
Color take a more active
segmentation: K- role in daily wound
means clustering care.
Rigla M. 201847 Gestational Diabetes Mobile NA 20 patients diagnosed with Metabolic and perinatal outcomes Artificial-intelligence-
Management (GDM) telemedicine GDM were similar except for BP, augmented
Using Smart Mobile system (Parallel observational which was lower in patients telemedicine has
Telemedicine prospectively captured using the telemedicine system been proposed as a
clinical data for historical helpful tool to
control) facilitate an efficient
widespread medical
assistance to GDM.

AI, artificial intelligence; ANN, artificial neural network; AP, artificial pancreas; AUC, area under the curve; BP, blood pressure; CBR, case-based reasoning; CKD, chronic kidney disease; CI,
confidence interval; CNN, convolutional neural network; DM, diabetes mellitus; DR, diabetic retinopathy; ECG, electrocardiogram; EMR, electronic medical record; FLC, Fuzzy Logic Controller;
GDM, gestational diabetes management; ICD-9, International Classification of Diseases, Ninth Revision; k-NN, k nearest neighbors algorithm; LASSO, least absolute shrinkage and selection
operator; LR, logistic regression; MLP, multilayer perceptron; MPC, model predictive control; N/A, not applicable; NA; NAFLD, nonalcoholic fatty liver disease; NB, naı̈ve bayes; NLP, natural
language processing; NPV, negative predictive value; PDB, protein databank; PID, proportional integral derivative; PPV, positive predictive value; R2R, run-to-run; RBF, radial basis function; RF,
random forest; ROC, receiver operating characteristic; SNP, single nucleotide polymorphism; SVM, support vector machine.
TRANSFORMING DIABETES CARE THROUGH AI 239

A review of the published articles documented the sub- decision-support tools for clinicians; and patient self-
stantial advances in AI technology over the last 10 years and management aids. Key examples from the published litera-
how it is helping PWDs and their clinicians make more ture are summarized in Table 3 and follow.
informed choices. Examples of the most common AI- DR is the most serious cause of secondary blindness,
powered diabetes care devices and systems identified in the exacting an enormous burden on individuals, families, and
published literature are summarized in Table 3. Table 3 the health care system. The annual diabetic retinal exam
provides examples of the research questions receiving the serves to screen and proactively detect diabetes patients with
most attention among diabetes AI researchers and product early treatable retinopathy. It is estimated that 98% of vision
developers over the past decade. loss from DR and macular edema is avoidable through im-
proved prediction, early detection, and treatment strategies,
and its cost-effectiveness is well established.17 Yet, major
Discussion
barriers to implementing more widespread screenings in-
The published studies suggest that a broad spectrum of clude the limited number of eye care practitioners who are
market-ready AI approaches are being developed, tested, and trained in interpreting retinal images, along with access to
deployed today in the prevention, detection, and treatment of care barriers.
diabetes. The total number of published technical articles re- Today, research documents that deep learning-based AI-
porting advances in the field of diabetes and AI increased grading of DR from retinal photographs is associated with
exponentially in the past decade, from 2600 in 2008, to 5500 in sensitivity and specificity over 90%.18–21 Recently, the US
2013, to more than 10,000 in 2017.13 Millions of patient health Food and Drug Administration (FDA) approved marketing the
records and newly published research exist that need to be first medical device to use AI to screen diabetes patients for
further processed, analyzed, and learned from to create a cur- retinopathy.6 The device, called IDx-DR (IDx LLC, Coral-
rent diabetes knowledge base for patients, researchers, doctors, ville, IA), is a software program that uses an AI algorithm to
and clinicians. Because of AI’s ability to rapidly interpret and analyze images of the eye taken with a retinal camera called
process enormous amounts of data into simple actionable the Topcon NW400 (Topcon Medical Systems, Inc., Oakland,
guidance, these published studies suggest that AI has signifi- NJ). Digital images of the patient’s retinas are uploaded to a
cant potential to improve screening, diagnosis, and manage- cloud server on which IDx-DR software is installed. If the
ment of patients with diabetes.14 images are of sufficient quality, the software provides the
Researchers are employing various AI approaches to inter- doctor with one of 2 results: (1) ‘‘more than mild diabetic
pret the vast amount of relevant data that need to be analyzed retinopathy detected: refer to an eye care professional’’ or (2)
and assessed.15 Table 2 describes the more common AI ap- ‘‘negative for more than mild diabetic retinopathy; rescreen
proaches described in the research and lists their clinical ap- in 12 months.’’ IDx-DR is the first device authorized for
plications in diabetes care. AI involves a wide spectrum of marketing that provides a screening decision without the
increasingly complex algorithms encompassed within the ter- need for a clinician to also interpret the image or results.22
minology of machine learning, deep learning, and cognitive These automatic systems enable non–eye health profes-
computing. In machine learning, experts typically ‘‘train’’ AI sionals in primary care physician offices to perform on-site
systems with large amounts of data and algorithms, which retinal screening and provide on-the-spot normal results or
enable the machine to examine relationships and learn from immediate referrals to the eye specialist without the need for
them. In deep learning, AI systems identify relevant insights for eye specialists, with significantly higher patient satisfaction
diagnostic support, while automatically conducting certain with the simplified process.23,24
complex and time-consuming tasks. Cognitive AI systems go Today, AI-driven predictive modeling proactively identifies
even further by understanding, reasoning, interacting, and diabetes populations with the highest risks of avoidable com-
learning. These systems understand by processing and deeply plications resulting in unnecessary emergency department visits,
interpreting the available data, both structured and unstructured, admissions, and readmissions.25 Larger physician groups, health
at enormous speed and volume. They reason by understanding care systems, and health plans utilize AI to ‘‘mine’’ large sets of
entities and relationships, making connections, proposing hy- digital and unstructured patient data to proactively identify and
potheses, and evaluating evidence. In contrast to the electronic characterize diabetes populations,26,27 find patients at risk for
health record, they provide a more natural interaction between diabetic comorbidities,28–30 identify patients for special diabetes
human and computer, facilitating dialogue, visualization, and disease management programs,31 and discover relevant pro-
collaboration.16 Cognitive AI Systems learn by collecting and teins32 and genes33 associated with and predictive of diabetes.
evaluating feedback at all levels of the system. The result is Today, AI provides practice decision-support tools for
practical knowledge, aids, and devices for diabetes patients and physicians and other health professionals caring for PWDs.
their clinicians that save time, improve efficiency, enhance Machine learning approaches help physicians customize
clinical decision making, empower patients, and have the po- diabetes medications to optimize adherence and health
tential to improve health outcomes and patient and clinician outcomes.34 AI-powered devices help physicians diagnose
satisfaction. Although a more complete explanation of AI is diabetes noninvasively,35 and more accurately measure
beyond the focus of this research article, this study indicates and monitor the severity of diabetic neuropathy36 and di-
that AI is a growing presence in diabetes care with the potential abetic wounds.37
to transform millions of people’s lives. Today, research suggests that diabetes management for
Today, research suggests that AI approaches are rapidly both PWDs and their clinicians is being simplified and im-
transforming care in 4 vital areas: improved screening and proved by new sensors, pumps, smartphone applications,
detection of diabetic retinopathy (DR) and macular edema; and other breakthroughs in AI to achieve better blood glu-
individualized predictive risk stratification and treatment; cose control,38 reduce hypoglycemic episodes,39 and
240 DANKWA-MULLAN ET AL.

improve patient satisfaction and reported outcomes.40 A every 5 minutes and sent to a personal app, the Guardian
2017 meta-review of published clinical trials of the latest, Connect system was accurate and was able to alert patients
automated, personal or real-time continuous glucose moni- of about 98.5% of hypoglycemic events so that they could
toring devices (RT-GCM) using computerized AI algo- proactively take action to normalize blood sugar.48
rithms concluded that a wide range of AI-powered RT-CMG This information also can be shared and monitored with
devices are entering the market to enable PWDs and their caretakers and family members in real time or via text
clinicians to assess and improve glycemic control, reduce message. In addition, Guardian Connect CGM is connected
hypoglycemic episodes, especially at night, and to improve to the [Link] smart diabetes assistant. Utilizing AI tech-
A1c levels. nology from IBM Watson Health, the [Link] assistant
Published research documents the extensive testing under continually analyzes how a patient’s blood glucose levels
way with the ‘‘artificial pancreas,’’ known also as a Closed respond to factors such as food intake, insulin dosages, and
Loop System, which combines continuous glucose mea- daily routines. Relative to baseline metrics, [Link] con-
surement with algorithm-driven insulin pumps to reduce ferred in 256 Guardian Connect users tested a 36-minute/
hypoglycemia and improve diabetes self-care.41 The latest- day improvement in blood glucose time-in-range or 9 full
generation sensors, which are more accurate and sensitive days a year, a 30-minute/day decrease in time >180 mg/dl,
for hypoglycemia, and the development of algorithms that and a 6-minute/day decrease in time <70 mg/dl, all statisti-
allow insulin infusion to be suspended during hypoglycemia cally significant. During the course of the 31+ patient-years
and glucagon to be administered, provide a safe and effec- of use, [Link] generated 655 insights for PWDs related to
tive system for persons at high risk of hypoglycemia. A hypoglycemia and 699 related to hyperglycemia. In addi-
meta-review of 12 published transition and home studies of tion, 134 [Link] users were randomly given Fitbits ex-
10 to 58 patients comparing clinical performance and pa- ercise monitoring applications during the course of the
tient acceptance of Artificial Pancreas Devices (APDs) with study. Results showed that glucose responses to meals and
traditional monitoring concluded that research, testing, and activity vary greatly, demonstrating the importance of per-
validation has moved from the laboratory to free-living, sonalization in diabetes self-management. Notably, 231 of
unsupervised home settings in the past decade, with accu- the 256 (90%) users recorded at least 2 weeks of data,
racy and reliability of the latest APD devices compatible demonstrating a solid pattern of engagement with the AI-
with safe operation and high patient satisfaction.42 A recent powered diabetes self-management application.49
review identified 18 closed-loop APDs being tested – 6 Many challenges remain before diabetes AI apps, devices,
APDs in the home setting, 5 in outpatient settings, and 7 in and systems become ubiquitous in the health care market-
inpatient settings – with planned commercial availability in place. One major challenge is technical interoperability
2018 and 2019.43 between systems: the ability of 2 or more systems to ex-
Today, research findings document the promise of dia- change and use the information.50 In addition, expensive up-
betes apps to assist users in tracking and analyzing their data front and ongoing costs, physician cooperation, and the
in a hassle-free way and to deliver personalized data-driven complexity of meeting Meaningful Use criteria stifle adop-
insights that PWDs may apply in their daily life. Today, tion and innovation.51
best-in-class apps provide comprehensive nutrition data- Another major challenge is the limitation in reproducing
bases that tell a user the nutritional content after scanning AI results from published studies. A most basic problem is
the barcode, allow them to search for restaurant menu items that researchers often do not share their source codes,
or popular meals by their names, or recognize food items on sometimes for competitive reasons. A survey of 400 algo-
a plate.44 Smartphone sensors using machine learning and rithms presented in papers at 2 top AI conferences in the
symbolic reasoning can recognize and quantify high-level past few years revealed that only 6% of the presenters
lifestyle activities of patients with diabetes and help them shared the algorithm’s code. Only one third shared the data
make more informed activity choices.45 An AI smartphone used to test their algorithms, and just half shared the
camera system enables patients with diabetes and their ‘‘pseudocode’’–a limited summary of a source code algo-
caregivers to take a more active role in daily wound care, rithm.52 In addition, assuming one can obtain and run the
and may potentially accelerate wound healing, save travel original pseudocode, it still might not do what is expected.
cost, and reduce health care expenses.46 AI-augmented tel- In the area of AI called machine learning, in which com-
emedicine has been studied to facilitate medical assistance puters derive expertise from experience, the training data for
in the homes of pregnant women with gestational diabetes, an algorithm – for example, the key information to train
with a high degree of patient acceptance. A research study speech-recognition learning systems – also can influence its
incorporated computer-interpretable clinical practice performance.
guidelines, and access to data from the electronic health Despite these challenges, this review of recently-
record as well as from glucose, blood pressure, and activity published, high-impact, and clinically-relevant studies sug-
sensors.47 gests that diabetes is attracting top health care technology
On March 14, 2018, the FDA approved Medtronic’s companies as well as start-ups that are using innovative AI
Guardian Connect, the first AI-powered continuous glucose technologies and approaches to tackle daily challenges faced
monitoring (CGM) system, for use in PWDs between the by PWDs. Many of the applications have received regula-
ages of 14 and 75 years. Guardian Connect utilizes a pre- tory approval in the past few years and are on the market
dictive algorithm that alerts patients of significant swings in today. Many more are on the way with the aim to disrupt
blood glucose levels up to 60 minutes prior to the event. and transform diabetes care by improving accuracy, effi-
When combined with the Guardian Sensor 3, which is ciency, ease of use, simplicity, and enjoyment on behalf of
placed on the abdomen to monitor blood glucose levels PWDs and their providers, caregivers, and family. The
TRANSFORMING DIABETES CARE THROUGH AI 241

published literature suggests that the combination of con- assistance to patient safety. N Engl J Med 2018;378:1271–
tinuous monitoring and real-time feedback to PWDs may be 1273.
able to identify meaningful patterns and lead to personalized 12. Young K. Newly approved software uses AI to improve di-
insights that increase patient and clinician engagement, abetic retinopathy. [Link]
confidence, and success in maintaining blood glucose levels 04/12/newly-approved-software-uses-ai-detect-diabetic Ac-
under better control. cessed July 18, 2018.
13. Pfaffl MW. A new mathematical model for relative quan-
tification in real-time RT-PCR. Nucleic Acids Res 2001;29:
Acknowledgment
45e.
The authors would like to acknowledge the important 14. López B, Martin C, Viñas PH. Special section on artificial
contributions of Lisa Latts, MD, MSPH, MBA, to this intelligence for diabetes. Artif Intell Med 2018;85:27–28.
manuscript and the field of artificial intelligence and health 15. Raghupathi W, Raghupathi V. Big data analytics in
care transformation. healthcare: promise and potential. Health Inf Sci Syst
2014;2:3.
16. Hoyt RE, Snider D, Thompson C, Mantravadi S. IBM
Author Disclosure Statement
Watson analytics: automating visualization, descriptive,
Drs. Dankwa-Mullan, Park, Snowdon, and Rhee declare and predictive statistics. JMIR Public Heal Surveill 2016;2:
that there are no conflicts of interest. Drs. Rivo and Se- e157.
pulveda received consulting fees from IBM Watson Health 17. Vijan S, Hofer TP, Hayward RA. Cost-utility analysis of
during the conduct of the study. This research did not re- screening intervals for diabetic retinopathy in patients with
ceive any specific grant from funding agencies in the public, type 2 diabetes mellitus. JAMA 2000;283:889–896.
commercial, or not-for-profit sectors. The views expressed 18. Gulshan V, Peng L, Coram M, et al. Development and
in this article are the authors’ own and not an official po- validation of a deep learning algorithm for detection of di-
sition of IBM. abetic retinopathy in retinal fundus photographs. JAMA
2016;316:2402–2410.
19. Tariq A, Akram MU, Shaukat A, Khan SA. Automated
References
detection and grading of diabetic maculopathy in digital
1. International Diabetes Federation (IDF). IDF diabetes atlas, retinal images. J Digit Imaging 2013;26:803–812.
7th edition. Brussels, Belgium: International Diabetes Fed- 20. Rahim SS, Palade V, Shuttleworth J, Jayne C. Automatic
eration, 2015. screening and classification of diabetic retinopathy fundus
2. Liu J, Wang R, Ganz ML, Paprocki Y, Schneider D, images. In: Mladenov V, Jayne C, Iliadis L, eds. Engineering
Weatherall J. The burden of severe hypoglycemia in type 2 applications of neural networks. Communications in com-
diabetes. Curr Med Res Opin 2018;34:179–186. puter and information science. Vol 459. Basel, Switzerland:
3. Centers for Disease Control and Prevention. Estimates of Springer Nature Switzerland AG, 2014:113–122.
diabetes and its burden in the United States. Washington, 21. Lam C, Yu C, Huang L, Rubin D. Retinal lesion detection
DC: US Department of Health and Human Services, 2017: with deep learning using image patches. Invest Ophthalmol
1–20. Vis Sci 2018;59:590–596.
4. The Diabetes Control and Complication Trial Research 22. US Food and Drug Administration. FDA permits mar-
Group. The effect of intensive treatment of diabetes on the keting of artificial intelligence-based device to detect
development and progression of long-term complications in certain diabetes-related eye problems. [Link]
insulin-dependent diabetes mellitus. N Engl J Med 1993; NewsEvents/Newsroom/PressAnnouncements/ucm604357
329:977–986. .htm Accessed July 18, 2018.
5. Latts L. ADA/IBM Watson Health Study (N>300,000) finds 23. Keel S, Lee PY, Scheetz J, et al. Feasibility and patient
that nearly 60% of people with T2D discontinue therapy acceptability of a novel artificial intelligence-based screen-
after one year. Presented at: American Diabetes Association ing model for diabetic retinopathy at endocrinology outpa-
78th Scientific Session, June 22–26, 2018, Orlando, FL. tient services: a pilot study. Sci Rep 2018;8:4330.
6. National Center for Biotechnology Information; US National 24. Matimba A, Woodward R, Tambo E, Ramsay M, Gwan-
Library of Medicine. PubMed. [Link] zura L, Guramatunhu S. Tele-ophthalmology: opportunities
pubmed/ Accessed July 18, 2018. for improving diabetes eye care in resource- and specialist-
7. Lewis K. Humans and systems: creating natural interfaces to limited Sub-Saharan African countries. J Telemed Telecare
augment human ability. [Link] 2016;22:311–316.
of-things/creating-natural-interfaces/ Accessed July 18, 2018. 25. Han L, Luo S, Yu J, Pan L, Chen S. Rule extraction from
8. Aggarwal M, Madhukar M. IBM’s Watson analytics for support vector machines using ensemble learning approach:
health care: a miracle made true. Hershey, PA: IGI Global, an application for diagnosis of diabetes. IEEE J Biomed
2016:117–134. Health Inform 2015;19:728–734.
9. Turing Archive for the History of Computing. www 26. Shankaracharya, Odedra D, Samanta S, Vidyarthi AS.
.[Link]/turing_archive/archive/index/archiveindex Computational intelligence-based diagnosis tool for the
.html Accessed July 18, 2018. detection of prediabetes and type 2 diabetes in India. Rev
10. Research 2 Guidance. Top 3 therapy fields with the best Diabet Stud 2012;9:55–62.
market potential for digital health apps. [Link] 27. Wei W-Q, Tao C, Jiang G, Chute CG. A high throughput
[Link]/top-3-therapy-fields-with-the-best-market- semantic concept frequency based approach for patient
potential-for-digital-health-apps/ Accessed July 18, 2018. identification: a case study using type 2 diabetes mellitus
11. Yeung S, Downing NL, Fei-Fei L, Milstein A. Bedside clinical notes. AMIA Annu Symp Proc 2010;2010:857–
computer vision—moving artificial intelligence from driver 861.
242 DANKWA-MULLAN ET AL.

28. Corey KE, Kartoun U, Zheng H, Shaw SY. Development 41. DeJournett L, DeJournett J. In silico testing of an artificial-
and validation of an algorithm to identify nonalcoholic fatty intelligence-based artificial pancreas designed for use in the
liver disease in the electronic medical record. Dig Dis Sci intensive care unit setting. J Diabetes Sci Technol 2016;10:
2016;61:913–919. 1360–1371.
29. Neves J, Martins MR, Vilhena J, et al. A soft computing 42. Thabit H, Tauschmann M, Allen JM, et al. Home use of an
approach to kidney diseases evaluation. J Med Syst 2015; artificial beta cell in type 1 diabetes. N Engl J Med 2015;
39:131. 373:2129–2140.
30. Rau H-H, Hsu C-Y, Lin Y-A, et al. Development of a web- 43. Trevitt S, Simpson S, Wood A. Artificial pancreas device
based liver cancer prediction model for type II diabetes pa- systems for the closed-loop control of type 1 diabetes: what
tients by using an artificial neural network. Comput Methods systems are in development? J Diabetes Sci Technol 2016;
Programs Biomed 2016;125:58–65. 10:714–723.
31. Weber C, Neeser K. Using individualized predictive dis- 44. Zhang W, Yu Q, Siddiquie B, Divakaran A, Sawhney H.
ease modeling to identify patients with the potential to ‘‘Snap-n-Eat’’: food recognition and nutrition estimation on
benefit from a disease management program for diabetes a smartphone. J Diabetes Sci Technol 2015;9:525–533.
mellitus. Dis Manag 2006;9:242–256. 45. Cvetković B, Janko V, Romero AE, Kafalı Ö, Stathis K,
32. Vyas R, Bapat S, Jain E, Karthikeyan M, Tambe S, Kulkarni Luštrek M. Activity recognition for diabetic patients using
BD. Building and analysis of protein-protein interactions a smartphone. J Med Syst 2016;40:256.
related to diabetes mellitus using support vector machine, 46. Wang L, Pedersen PC, Strong DM, Tulu B, Agu E, Ignotz
biomedical text mining and network analysis. Comput Biol R. Smartphone-based wound assessment system for pa-
Chem 2016;65:37–44. tients with diabetes. IEEE Trans Biomed Eng 2015;62:477–
33. López B, Torrent-Fontbona F, Viñas R, Fernández-Real 488.
JM. Single nucleotide polymorphism relevance learning 47. Rigla M, Martı́nez-Sarriegui I, Garcı́a-Sáez G, Pons B,
with random forests for type 2 diabetes risk prediction. Hernando ME. Gestational diabetes management using
Artif Intell Med 2018;85:43–49. smart mobile telemedicine. J Diabetes Sci Technol 2018;
34. Lo-Ciganic WH, Donohue JM, Thorpe JM, et al. Using 12:260–264.
machine learning to examine medication adherence thresh- 48. [Link]. Adult accuracy study of the Elite 3
olds and risk of hospitalization. Med Care 2015;53:720–728. Glucose Censor (E3). [Link]
35. Shu T, Zhang B, Yan Tang Y. An extensive analysis of show/study/NCT02246582?term=NCT02246582&rank=1&
various texture feature extractors to detect diabetes mellitus sect=X0123456 Accessed July 18, 2018.
using facial specific regions. Comput Biol Med 2017;83: 49. Neemuchwala H. [Link] improves time-in-range by
69–83. 36 mins/day, time >180 by 30 mins/day, time <70 by 6 mins/
36. Katigari MR, Ayatollahi H, Malek M, Haghighi MK. Fuzzy day. Presented at: American Diabetes Association 78th Sci-
expert system for diagnosing diabetic neuropathy. World J entific Session, June 22–26, 2018, Orlando, FL.
Diabetes 2017;8:80–88. 50. Benson T. Principles of health interoperability HL7 and
37. Wang L, Pedersen PC, Agu E, Strong DiM, Tulu B. Area SNOMED, 2nd edition. London: Springer Verlag, 2012.
determination of diabetic foot ulcer images using a cas- 51. Adler-Milstein J, DesRoches CM, Kralovec P, et al. Elec-
caded two-wtage SVM-based classification. IEEE Trans tronic health record adoption in us hospitals: Progress
Biomed Eng 2017;64:2098–2109. continues, but challenges persist. Health Aff (Millwood)
38. Mauseth R, Lord SM, Hirsch IB, Kircher RC, Matheson 2015;34:2174–2180.
DP, Greenbaum CJ. Stress testing of an artificial pancreas 52. Hutson TM. Missing data hinder replication of artificial
system with pizza and exercise leads to improvements in intelligence studies. [Link]/news/2018/02/
the system’s fuzzy logic controller. J Diabetes Sci Technol missing-data-hinder-replication-artificial-intelligence-studies?
2015;9:1253–1259. Accessed July 18, 2018.
39. Ling SH, Nguyen HT. Natural occurrence of nocturnal
hypoglycemia detection using hybrid particle swarm opti- Address correspondence to:
mized fuzzy reasoning model. Artif Intell Med 2012;55: Marc Rivo, MD, MPH
177–184. Population Health Innovations, Inc.
40. Herrero P, Pesl P, Reddy M, Oliver N, Georgiou P, Tou- 4566 Prairie Avenue
mazou C. Advanced insulin bolus advisor based on run-to- Miami Beach, FL 33140
run control and case-based reasoning. IEEE J Biomed
Health Inform 2015;19:1087–1096. E-mail: marcleerivo@[Link]