Baaa 010

Database, 2020, 1–35
doi: 10.1093/database/baaa010
Review
Downloaded from https://academic.oup.com/database/article/doi/10.1093/database/baaa010/5809229 by Pazmany Peter Catholic University user on 27 September 2023
Review
Artificial intelligence with multi-functional

machine learning platform development for
better healthcare and precision medicine
Zeeshan Ahmed1,2,3,4,*, Khalid Mohamed3 , Saman Zeeshan5 and
XinQi Dong1,2
1 Institute for Health, Health Care Policy and Aging Research, Rutgers, The State University of New Jersey,
112 Paterson Street, New Brunswick, NJ, USA, 2 Department of Medicine, Rutgers Robert Wood Johnson
Medical School, Rutgers Biomedical and Health Sciences, 125 Paterson Street, New Brunswick, NJ,
USA, 3 Department of Genetics and Genome Sciences, School of Medicine, University of Connecticut
Health Center, 263 Farmington Ave., Farmington, CT, USA, 4 Institute for Systems Genomics, University
of Connecticut, 67 North Eagleville Road, Storrs, CT, USA and 5 The Jackson Laboratory for Genomic
Medicine, 10 Discovery Drive, Farmington, CT, USA
*Corresponding author: Email: zahmed@ifh.rutgers.edu
Citation details: Ahmed,Z., Mohamed,K., Zeeshan,S. and Dong, X. Artificial intelligence with multi-functional machine
learning platform development for better healthcare and precision medicine. Database (2020) Vol. 2020: article ID
baaa010; doi:10.1093/database/baaa010
Received 2 November 2019; Revised 5 January 2020; Accepted 21 January 2020
Abstract
Precision medicine is one of the recent and powerful developments in medical care,
which has the potential to improve the traditional symptom-driven practice of medicine,
allowing earlier interventions using advanced diagnostics and tailoring better and
economically personalized treatments. Identifying the best pathway to personalized and
population medicine involves the ability to analyze comprehensive patient information
together with broader aspects to monitor and distinguish between sick and relatively
healthy people, which will lead to a better understanding of biological indicators that
can signal shifts in health. While the complexities of disease at the individual level have
made it difficult to utilize healthcare information in clinical decision-making, some of
the existing constraints have been greatly minimized by technological advancements.
To implement effective precision medicine with enhanced ability to positively impact
patient outcomes and provide real-time decision support, it is important to harness the
power of electronic health records by integrating disparate data sources and discovering
patient-specific patterns of disease progression. Useful analytic tools, technologies,
databases, and approaches are required to augment networking and interoperability of
clinical, laboratory and public health systems, as well as addressing ethical and social
issues related to the privacy and protection of healthcare data with effective balance.
Developing multifunctional machine learning platforms for clinical data extraction,
aggregation, management and analysis can support clinicians by efficiently stratifying
© The Author(s) 2020. Published by Oxford University Press. Page 1 of 35

This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits
unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.
(page number not for citation purposes)
Page 2 of 35 Database, Vol. 2020, Article ID baaa010
subjects to understand specific scenarios and optimize decision-making. Implementation

of artificial intelligence in healthcare is a compelling vision that has the potential in
leading to the significant improvements for achieving the goals of providing real-time,
better personalized and population medicine at lower costs. In this study, we focused on
analyzing and discussing various published artificial intelligence and machine learning
solutions, approaches and perspectives, aiming to advance academic solutions in paving
the way for a new data-centric era of discovery in healthcare.
Introduction precision medicine has evolved as a central innovation pillar

for leading research in transforming health and holds great
Over the centuries, quests for answers have led us to take
promise in patient treatment (3,4). Precision medicine has
giant leaps. It was only in the last century that the discovery
the potential to improve the traditional symptom-driven
of antibiotics freed us from many of the dreaded diseases of
practice of medicine by intelligently integrating multi-omics
the past. Still, in the context of recent published literature
profiles with clinical, imaging, epidemiological and demo-
(e.g. accessible through PubMed), over 138 000 studies
graphic details to allow a wide range of earlier interventions
discuss medication errors, and over 450 000 include delayed
for advanced diagnostics and tailoring better and econom-
treatment (date accessed 10 October 2019, using search
ical personalized treatment (Figure 1). This requires a pro-
query keywords including ‘medication error’ and ‘delayed
gressive healthcare environment that can enable clinicians
treatment’). Still, the problem of people dying from medical
and researchers to gain a complete picture of the patient to
care gone wrong has been vastly underappreciated and
deepen their understanding, using additional basic details
not well recognized. Today, we stand on the threshold of
from healthcare data e.g. phenotypic information, life style
the new medical revolution, just as big and far-reaching.
factors and social determinants that can impact treatment
Despite all of our scientific knowledge, much of medicine
is still based on the treatment of symptoms and performing decisions. It is primarily based on ‘4Ps’—Predictive, Pre-
learned trials based on treatments, which works for most ventive, Personalized and Participatory—treatment of each
patients to bring symptom relief, reduce the risk of com- individual patient and aims to enable clinicians to efficiently
plications and improve survival chances, but not for all. understand how personalized clinical data variations can
To get new insights into disease taxonomy, etiology and contribute to health and accurately diagnose and predict the
pathogenesis, it is important to understand how diseases most appropriate course of action for a patient (5). While
are related to each other. Breakthroughs in prescription the complexities of diseases at the individual level have
medication, surgical treatment and mental health inter- made it difficult to utilize healthcare information in clinical
ventions are among the reasons we live longer. However, decision-making, some of the existing constraints have been
providing the correct in-time treatment plan for patients minimized by technological advancements (6). To imple-
with knowledge about their current medications and drug ment effective personalized and population health with
allergies is currently a tedious and error-prone task [1]. The enhanced ability to positively impact patient outcomes,
widespread growth of prescribing and consuming medica- it is important to harness the power of electronic health
tions has increased the need for applications that support records (EHR) by integrating disparate data sources and
medication reconciliation. Furthermore, living healthcare discovering patient-specific patterns of disease progression
issues include misdiagnosis, overtreatment, decreased pro- to provide real-time decision support. The significance of
ductivity, under-utilized clinical data handling, significant healthcare data mining cannot be denied, but the challenges
cost and spending (Figure 1). These miscalculations can be of big data management loom large (7).
reduced to a great extent with the use of advancements in Over the years, biotechnology has evolved immensely.
information technology at every level of care. Computers are becoming faster in speed and micro in size,
Medical error is the third leading cause of death after heterogeneity is increasing in datasets and their volume is
heart failure and cancer (2). According to recent studies, growing robustly. These expansions are fueling the engine
approximately 180 000 to 251 000 people are dying every of artificial intelligence (AI) for discovering many technical
year in the USA due to medical errors (2). This number refinements to solve complex problems in almost every field
has been rising due to increasing complexity and reduced of life, including science and medicine. AI is the branch
quality of our current medical system, which includes com- of computer science with the capability of a machine to
munication breakdown, misdiagnosis, poorly coordinated imitate and even enhance intelligent human behavior. One
care and growing cost. In recent years, the concept of of the expected roles in life and medical sciences is to
Database, Vol. 2020, Article ID baaa010 Page 3 of 35
Figure 1. Role of artificial intelligence in traditional healthcare data analytics, and in precision medicine. Addressing key issues in healthcare (e.g.
misdiagnoses, overtreatment, one-size-fits-all approaches, repetitive, decreased productivity, under-utilized data, significant cost & spending), and
finding key biomarkers to provide economic and personalized treatment by intelligently analyzing heterogeneous data.
deal with extensive research studies aimed at support- ysis algorithms for clinical data interpretation, exploration
ing real-time decision-making and producing solutions to and drawing inference; unavailability of effective open-
complex problems through knowledge and data intensive source tools that combine various approaches to model
computational and simulated analysis (8). Healthcare data biological interactions; integration of clinical and analytic
includes information about a patient’s lifestyle, medical systems; interdisciplinary field barriers; high cost (21);
history, encountered visits with practices, laboratory and implementation of secure frameworks for data collection,
imaging tests, diagnoses, prescribed medications, performed simplification, conversion from raw form to knowledge,
surgical procedures and consulted providers (9). Adequate, management and distribution (22,23); automatic cleansing
analytical and intelligent access to healthcare data has the of faulty and error-prone EHRs; correctly identifying
potential to revolutionize the field of medicine by improving prescription medication; and implementing predictive
the quality and transition of care, improving outcomes by diagnostics (10). It is not possible to easily track and
reducing cost, detecting diseases at earlier stages (10,11) prospectively follow the clinical progress and outcomes
and developing a better understanding of biological mecha- in patients over time (e.g. having a critical predictor of
nisms by modeling complex biological interactions through future clinical events where a patient may show the disease
a holistic integration and analysis of knowledge (12). The months or years down the road). In past decades, various
ability to stratify patients, understand scenarios and opti- systems have been developed in both commercial and
mize decision-making would consistently improve based on academic sectors (11,24,25,26,27,28,29,30) for this pur-
the myriad data obtained during the care-delivery process. pose. Academic systems put significant values on analytics,
Innovative and robust big data platforms are necessary while commercial systems focus on supporting clinical
to improve the quality and transition of healthcare by operations. However, independently utilizing traditional
analyzing heterogeneous healthcare, which can be of huge approaches, both sectors are unable to identify problems
volume, velocity, variety and veracity. by their effects and significantly help in clinical decision-
To effectively implement healthcare data analytic pro- making. However, major concerns include handling and
cesses, various big data management challenges (1,9,13,14, evaluation of electronic medical records (EMR), and
15,16,17,18,19) have to be overcome, which include repetitive tasks; patient medication adherence; inefficient
inadequacy of analyzable clinical data (20); existence of therapeutic treatment for cancer and other critical diseases;
multiple data standards, structures, types and formats; cost ineffectiveness; and addressing ethical issues related to
rapid growth in heterogeneous data; understanding of anal- AI and ML implementation in healthcare.
AI & ML in health intelligence, precision similar groups of subjects, associations between subject
medicine and resource management features and outcomes of interest; and apply natural lan-
Intelligent big data platforms are necessary to improve guage processing (NLP) methods to extract information
the quality and transition of healthcare by expediting from unstructured clinical data e.g. narrative text, such as
investigation of active hidden factors in clinical data with physical examination, clinical laboratory reports, operative
machine learning algorithms to obtain actionable gap-based notes and discharge summaries.
information about patients for early detection and preven-
tion of constitutional disorders like cancer, and streamlining AI in healthcare for better prevention, detection,
data sharing by developing efficient communication across diagnosis and treatment of disease (31)
healthcare units and scientific laboratories. Its application
The authors emphasized the idea of embracing changes
in healthcare could be another great leap in medicine
with the advancement in technology with the potential
and a transformational force for guiding personalized and
integration of AI into the field of healthcare in a way that
population medicine with several computational benefits.
is beneficial to each healthcare worker. They focused in
In the recent past, multiple AI and ML-based efforts have
utilizing AI to obviate repetitive tasks to enhance patient–
been made for deciphering diseases to facilitate predictive
physician relationship and increase practice of empathy
diagnosis and thereby guide treatment factors, e.g. drawing
and emotional intelligence. Authors focused on deep learn-
disease relationships using clinical manifestations, EHR
ing algorithm implementation with increased data flow
and data generated using wearable technology. To get
that allows machines to self-develop a complex function
a detailed overview of available academic solutions, we
with improved predictability, as long as a large amount of
reviewed contributions and compared various AI and ML
data is fed as input. They developed a deep convoluted
claimed solutions, approaches and discussions (Table 1),
neural network for skin cancer detection, image analysis
and real-time examples (Table 2) published within the last
for diabetic retinopathy evaluation, smartphone-based AI
5 years [31–63]. Our focus here was to discuss valued
platform to measure adherence in patients on direct oral
contributions of all mentioned AI and ML algorithms
anticoagulants (64) and patient’s visit length reduction (65).
(section Theoretical background of AI, ML and examples
The strengths of this study include the potential to augment
in healthcare and approaches. We provide detailed and
healthcare access in areas where specialists might not be
individualized overview of presented approaches. Overall,
physically available, and medication can still be prescribed
the review study and the contributions of AI and ML are
utilizing the combined efforts of AI and a primary care
divided into three categories: Health Intelligence, Precision
physician, especially in developing countries. However, one
Medicine, and Healthcare Resource Management and
missing aspect is to address the loss of privacy, as well
Ethical Challenges.
as possibilities of patient data exploitation. The potential
for Health Insurance Portability and Accountability Act
Health intelligence approaches (HIPAA) protection is likely feasible (1,9).
Health intelligence can play a vital role at various levels
of clinical research and analytics that can lead to sig-
nificant improvements in achieving goals for the provi- ML in medicine with better patient–provider
sion of better personalized and population healthcare. In interactions (32)
the past decade, various operational and research-based This approach is focused on examining the essential struc-
healthcare data management and analytic systems have tural changes in the healthcare systems that are necessary to
been developed in both academia and commercial sectors. unleash the full potential of ML in medicine. It emphasizes
Our interest includes comprehensive solutions that imple- developing the concepts of ML in medicine, which may be
ment healthcare data analytics process; provide features to centralized around the idea of personalized diagnosis and
manage, analyze, visualize and share EHR in de-identified treatment on the basis of all known information about the
form; help in automatically capturing information about patient and collective experience. Giving rationale, authors
patient demographics, scheduled appointments, pre-exam highlighted proof-of-concept models that have been tested
questionnaire results, consulted providers, conducted lab so far, e.g. difficulties in finding a relationship between
tests, diagnoses, treatment plans, objective test results, med- current ML models and traditional statistical models, need
ications, surgical procedures and claims; support the clini- for a tremendous amount of data to train ML classifiers for
cal decision-making process with AI techniques to create establishing general and complex associations and training
classifiers, which can be trained on structured clinical data clinicians in AI for accurate data interpretation. Further-
generated from different clinical activities and can learn more, they effectively debated about consumption of physi-
Table 1. Feature and variability analysis of reviewed approaches, and real time implementation of ML algorithms
Approach Objectives Approach AI & ML
AI power digital Reduce repetitive tasks and Increased task automation Deep convoluted neural network for
medicine (31) burdens of electronic medical with improved image skin cancer detection and reducing
records through the utilization processing. Monitor visit length. Deep neural network to
of AI and ML. medication adherence and evaluate images for diabetic
detect any changes. retinopathy. Smartphone-based AI
platform to measure adherence in
patients on direct oral
anticoagulants.
ML in medicine (32) Examining the essential Accumulation of large data Applied deep learning on the current
structural changes in the set and implement ML to EHRs data to generate associations
healthcare system that are anticipate events, develop and meaningful data for
necessary to unleash the full search engine, and monitor personalized diagnosis and
potential of machine learning data flow. treatment.
in medicine.
Precision medicine with Applying ML to the EHRs to Analyze patterns within the Supervised learning (support vector
electronic medical generate personalized medicine subset of population who machine, discriminant analysis, naïve
records (33) by converting EHR into present similar clinical Bayes, nearest neighbor and neural
reliable risk predictors, and phenotypes of complex network), unsupervised learning
incorporating patient’s disease. encompass (linear & logistical
variabilities for treatment and regression, decision tree, cluster
prevention of disease. analysis, and neural network).
AI, ML and the evolution of Examining AI integration in AI methods for the SVM model development for
healthcare (34) healthcare. extraction of big data and physiological data segmentation and
aid clinicians in care delivery analysis, disease progression
prediction, and diagnosis.
Solving healthcare problems Tailoring medical treatment Use of information ML for the implementation of
with precision medicine (35) with respect to the technology for precision medicine, which includes
individualized characteristics multidisciplinary data storage and analysis for
of patients. collaboration establishment determining the association between
between clinicians and disease outcome (e.g. disease risk,
researchers. prognosis, or treatment),
identification of patient
characteristics and optimal
treatment.
Role of AI in precision Examining role of AI in Analyzing large scaled Combining DL with human
medicine (36) precision medicine clinical dataset. pathologist to improve success rate
implementation. of diagnosis.
AI towards health in Utilization of AI in poor Implementing NLP over Pattern identification and tracking
resource-poor settings (37) settings, and improving health EHR for surveillance and disease transmission through ML.
outcome in those areas. out breaking predictions.
Integrated precision medicine Early diagnosis of chronic Feature extraction from Predictive, proactive intervention in
and role of EHR in conditions through proper clinical data, and utilization healthcare through AI, and clinical
personalized treatment (38) extraction of clinical insights. of silico dataset. decision support tool development.
AI in healthcare (39) Analyzing AI applications in Precise analysis at the ML algorithms to extract and cluster
healthcare, and their potential extracted useful information data, and perform principal
outcome in future. from a large patient component analysis, SVM to
population. determine model parameters, and
identify imaging biomarkers, NLP
for text processing and classification,
and DL for diagnostic imaging and
electro diagnosis.
(Continued)
Table 1. Continued
ML Knowledge Base with Examine three main pillars Developing knowledgebase ML approaches for pattern
ontology for pattern integrating personalized of existing phenotypes, recognition and development of
recognition in personalized medicine into everyday clinical patient enrollments, and statistical models (sample size and
medicine (40) practice, which are phenotype data expansion. effect size). Knowledge base of all
categories, population size and existing phenotype categories and
statistical analysis. disease. Organized clinical dataset of
population size. Software platform
for statistical analysis of
high-dimensional healthcare and
multi-omics data.
Data science, AI, and ML for Predictive modeling for better Data Science (DS) and AI to ML for healthcare data analysis and
laboratory medicine (41) collaboration between mimics the human processes optimization, and reducing cost,
hospitals without sharing data and improve the process of improving efficiency of staff and
and complying privacy decision-making. resources.
regulation.
AI to solve the human Solve the human resource crisis Implementation of AI Artificial narrow intelligence for
resource crisis in in healthcare with AI. techniques (43). performing a single task. Artificial
healthcare (42) general intelligence for
understanding and reasoning
environment like humans. Artificial
superintelligence for scientific
creativity. Deep learning for image
recognition, natural language
processing and translation.
Data analytics and ML for Analyzing EHR for the Converting electronic ML algorithm for structured and
disease identification in identification of wide range of healthcare record into unstructured big data analysis for
EHR (44) medical conditions and reliable risk predictors. the identification of wide range of
diagnosis. medical conditions and diagnosis.
AI, Big Data and Cancer (45) Application of AI and large Application of cognitive Cognitive computer systems for
scaled database for cancer computer systems for providing rapid access to accurate
diagnosis and treatment, approaching cancer information and treatment
worldwide. diagnosis and treatment procedures, and assisting in
(read, remember, decision-making.
recommend, and remind).
Use of EHR in comparative Reporting caveats in existing Literature review and Implementing ML for overcoming
effectiveness research (46) healthcare systems. reporting caveats. existing big data limitations in
healthcare systems.
Deep learning health care Reporting unintended Creating more precise ML for prognosis modeling in
system (47) consequences due to the analytics platform for oncology, and pattern recognition in
application of ML in existing prognosis modeling and radiology and pathology.
healthcare systems. pattern recognition.
DL to transform Transform healthcare by using Outperforming clinical Implementation of DL for the digital
healthcare (48) ML. systems and modeling image analysis.
complex relationships
among active hidden factors
of data
High-performance medicine Exploring importance and Literature review and field Deep neural networks for pattern
with AI (49) pitfalls of AI in medicine. analysis. recognition and analysis medical
images. NLP in drug discovery by
analyzing biomedical literature.
(Continued)
Table 1. Continued
Intelligent digital Improving diagnostic accuracy Implemented, examined and DL for analyzing whole-slide
pathology (50) and efficiency with the use of compared the performance pathology images.
ML. of DL at test dataset.
ML for prediction in Implementing ML for better Evaluating positives and ML algorithms for addressing
EHR (51) understanding heterogeneous negatives of ML algorithms. different clinical questions by
treatment effects to implement analyzing and finding nonlinear
precision medicine. relationships in the EHR.
Unintended consequences of Safe, effective, efficient and Implementing DL in DL for digital imaging, curating data
ML (52) humanistic care. healthcare analytic systems sets, integrative heterogeneous data
development and modeling analysis, identifying novel
tools. associations, and remote monitoring
and digital consultations.
Finding the missing link for Biomedical data integration Identify and discussed AI and ML tools development to
big biomedical data (53) and analysis located at challenges in biomedical analyze biomedical data for better
heterogeneous sources. data linking. clinical decision-making.
ML classifies cancer (54) Identification of novel tumor Application of ML for the ML for analyzing histological data.
classes. identification of tumor by Supervised ML for analyzing CNS
analyzing histology and tumor type genome-wide
genomics data. methylation data to identify
methylation patterns. Unsupervised
ML to search patterns in the data
sets to develop classification
categories.
Analyzing and visualizing Finding future strands of Quantitative review of the DL, new ML algorithms and
knowledge structures of research, including new health informatics field, advanced big data analytics for
health informatics (55) algorithms, tracking tools and employing text mining and better-personalized treatment.
Internet of Things-based bibliometric research
decision support systems. methods.
Big data and ML algorithms AI tools development based on Identification of clinical HCI-based AI and ML applications
for healthcare delivery (56) incremental learning to refine problems, annotation of for different clinical developments in
the predictive accuracies. extracted healthcare data, oncological.
application of appropriate
ML algorithms and its effect
on decision-making,
addressing legal and ethical
implications, assessment of
ML effect in trail, designing
freeze and submission of
dossier for medical devices,
training clinicians of ML
tool, and monitoring for
adverse outcomes.
Intelligent health data AI for advanced health data Health data analytics Implementation of AI and ML based
analytics (57) analytics. process involving a analysis with the inclusion of health
methodical order of data data preprocessing, selecting
processing, modeling, and algorithm based on expected
analysis steps. outcome, developing analytical
models, and interpreting results.
(Continued)
Table 1. Continued
Ethical challenges of Challenges of implementing Literature review and field Addressing current challenges in
implementing ML in ML in healthcare analysis. healthcare systems due to the
healthcare (58) implementation of ML.
Data science, AI, and ML for Implementing Data science, AI, Framework including ML for finding patterns, discovering
laboratory medicine (59) and ML for laboratory defining tasks, metrics, inefficiencies, predicting outcomes
medicine. models and datasets. and taking factual decisions.
Causal inference and Examined the implications of RWE framework. ML for data classification and
ML (60) progress in observational prediction in RWE to support
research design and healthcare clinical and regulatory
databases. decision-making.
Big data analytics in Application of big data Conceptual architecture of ML for data mining and analysis.
healthcare (61) analytics in healthcare. big data analytics, which
includes developing multi
source data input,
transformation, structure,
management and analysis
using traditional SQL,
OLAP and mining.
ML and genomics in Substantial improvements to Combining the latest ML models to address the challenges
precision medicine (62) address clinical and genomic computational data of gene variations and similarities
data security problems. protection principles with among patients.
legal and ethical perspectives
to construct a secure
framework for data sharing.
ML in cancer prognosis and ML to detect key features by Data preprocessing with ML (ANNs, BNs, SVMs,
prediction (63) predictive modeling of focus on data modification graph-based SSL and DT) to model
complex and heterogeneous via dimensionality reduction the progression and treatment of
datasets for progression and and feature detection. cancerous conditions thru examining
treatment of cancerous complex datasets and revealing their
conditions, risks and outcomes. relevance.
cians’ valuable time due to increased features (e.g. check by authors. They discussed use of support vector machine
boxes) in EHRs for administrative and billing purposes, (SVM)- and deep learning-based model development for
which might prevent them from providing the best quality physiological data segmentation and analysis, disease pro-
care for their patients. At the same time, they claimed that gression prediction and diagnosis in radiology. Authors
the integration of ML in EHRs may lead to potential fear discussed processes, which include designing effective mod-
of overreliance as well as decreased vigilance of errors and els to aid in the diagnosis based on the information that
automation bias. Authors claimed a training ML classifier resemble certain diseases, image analysis and interpretation
at EHRs for pattern detection to allow physicians to antic- to improve the decision-making performance of clinicians.
ipate future events in high-risk patients, to obtain accurate Authors also raised ethical concerns in utilizing ML, primar-
and comprehensive diagnosis and provide with a quick ily in governance and management of big data and future
search engine in locating the pertinent information within of employment.
a patient’s chart, less clicking, voice dictation and better
predictive typing.
Medical AI for increasing availability of
healthcare data and rapid development of big
AI, ML and the evolution of healthcare (34) data analytic methods (39)
Overview and implementation of AI and ML in healthcare A study analyzed different AI applications in healthcare and
for the extraction of big data, and aiding clinicians in discussed their potential future outcomes. Authors argued
providing better care delivery were recently highlighted that currently AI research in healthcare is mainly focused on
Table 2. Real time examples of AI and ML algorithms (support vector machine, deep learning, logistic regression, discriminant
analysis, decision tree, Random forest, linear regression, naïve Bayes, K-nearest neighbor, hidden Markov, genetic algorithm)
in healthcare
ML algorithms Examples in healthcare
Support vector machine Symptoms classification and analysis to improve diagnostic accuracy.
Identifying imaging biomarkers of neurological and psychiatric disease.
Diagnosing mental illness.
SVM with leave-one-out cross-validation for multiple myeloma by analyzing SNPs.
SVM with 20-fold cross-validation for breast cancer by analyzing SNPs.
SVM with hold-out for breast cancer by analyzing clinical, pathologic and epidemiologic data.
SVM with hold-out for cervical cancer by analyzing clinical and pathologic data.
SVM with 10-fold cross-validation for breast cancer by analyzing clinical and population data.
SVM with cross-validation for oral cancer by analyzing clinical and genomic data.
SVM with leave-one-out cross-validation for breast cancer by analyzing genomic data.
SVM with cross-validation for oral cancer by analyzing clinical, molecular data.
Deep learning Evaluating images for diabetic retinopathy.
Identification of type 2 diabetes (T2D) subgroups.
Measure medication adherence via camera interface.
Detection and segmentation of lung and liver tumors by analyzing CT scans.
Diagnosing eye diseases (diabetic retinopathy) by analyzing retinal images.
Diagnosing cardiac anomalies by analyzing images of MRI of heart ventricles.
Detecting malignant lung nodules by analyzing radiographs.
Producing glioma survival predictions by analyzing histological imaging and genomic marker data.
Histological diagnoses prediction in women with cytological abnormalities.
Oncology diagnosis:
• Thoracic (lung cancer)
• Abdominal and pelvic (tomography and magnetic resonance imaging)
• Colonoscopy (colonic polyps)
• Mammography (microcalcifications)
• Brain (brain tumors)
• Radiation oncology (segmenting tumors for radiation, and quantifying specific
radiographic characteristics by analyzing 3D shape of a tumor from)
• Dermatology (skin cancer)
• Pathology (digital whole-slide of biopsy samples)
• Prostate (cancer tumors by analyzing ultrasound of biopsy cores)
• DNA and RNA sequencing (RNA-binding and DNA-binding proteins)
Logistic regression Risk assessment of complex diseases (e.g. tuberculosis, breast cancer, coronary heart disease).
Predicting patient survival rate.
Diagnosing coronary heart disease (CHD).
Non-Hodgkin’s lymphoma diagnosis with multivariable logistic regression modeling.
Identification of pulmonary thromboembolism by analyzing prognostic factors.
Discriminant analysis Identify surgical and operative factors to classify patients for surgical procedure.
Predict the clinical diagnosis of primary immunodeficiencies.
Patient data satisfaction.
Prediction of depression elements in cancer patients.
Classification of BOLD fMRI response to naturalistic movie stimuli.
Identify protein-coding regions of rice genes.
Parkinson’s disease symptoms recognition.
Risk assessment of for chronic illnesses.
Diagnosis of hypercalcemia.
Predicting patient care visits by identifying discriminatory characteristics.
(Continued)
Table 2. Continued
Decision tree Real-time healthcare monitoring medical decision support system, sensors anomaly detections, and data
mining model for pollution prediction.
Supporting clinical decisions.
Strategies alternate therapies for oncology patients.
Collaborative clinical decision-making in mental health care.
Identify predictors of health outcomes.
Find factors related to hypertension.
Discover factors associated with pressure ulcers (PUs) among elderly people.
Identify the potential recipients of telehealth services.
Patient data stratification for interpretable decision-making for precision medicine.
Content analysis for patient aids decision.
Diabetic foot amputation risk analysis.
Support understanding of antenatal lifestyle interventions.
Random Forest Diagnosing mental illness.
Detecting knee osteoarthritis.
Monitoring medical wireless sensors.
Diagnosing Alzheimer disease.
Predicting metabolic pathways.
Predicting outcomes of a patients encounter with behavioral health providers.
Healthcare cost prediction.
Mortality prediction for intensive care unit (ICU) patients.
Classification of Alzheimer’s disease.
Identifying social and economic factors to study social determinants of health.
Predicting disease risks from imbalanced data.
Identify associates of diabetic peripheral neuropathy diagnosis.
Predicting the risk of emergency admission.
Detecting patients ready to discharge from intensive care.
Nonparametric estimation of heterogeneous treatment effects.
Diagnose sleep disorders.
Predicting the depression in patients suffering with Alzheimer’s disease.
Predicting myopia by analyzing EHR.
Linear regression Identification of prognostically relevant risk factors.
Predict hand surgery.
Monitor prescribing patterns and ensure treatment appropriateness.
Mean on decision-making in health care.
Reducing high costs of the health system.
Analysis of skewed healthcare cost data.
Understand HIV/AIDS prevalence patterns.
Naïve Bayes Predictive modeling for different diseases (brain, asthma, prostate and breast cancer etc.).
Risk prediction using censored and time-to-event data.
Mucopolysaccharidosis type II detection.
Predicting Alzheimer’s disease from genome-wide data.
Measuring quality healthcare services.
Finding audit targets in performance-based financing in health.
Modeling medical diagnosis for decision support.
EHR classification.
Classifier and genetic score for risk prediction.
Decision support system for heart disease.
(Continued)
Table 2. Continued
K-nearest neighbor Diagnostic performance of the model.
Preserving privacy of medical diagnosis in e-Health cloud.
Medical dataset classification.
Classification of lymph node metastasis in gastric cancer.
Pattern classification for breast cancer diagnosis.
Pattern classification for health monitoring applications.
Pancreatic cancer prediction combining published literature and EHR data.
Disease diagnosis and detection of Parkinson.
Hand-gesture recognition for elderly individuals.
ECG features extraction for mobile healthcare applications.
Hidden Markov Analyzing sequence data (predicting exons and introns, identifying ORFs, insertions, deletions,
substitutions, functional motifs in proteins, aligning two sequences, and switching from exon to intron in a
DNA sequence)
Modeling ‘Healthy’ and ‘Unhealthy’ unobserved health states.
Analyze time-series personal health checkup data.
Improving length of stay prediction and reducing health care costs.
Mining adverse drug reactions from online healthcare forums.
Monitor, model and cluster medical inpatient journeys.
Analyzing healthcare service utilization after transport-related injuries.
Monitoring circadian in telemetric activity data.
Predicting patients entering states with a high number of asynchronies.
Analyzing subject-specific seizure, automatic segmentation of infant cry signals.
Genetic algorithm Detecting microcalcifications in mammograms leading breast cancer.
Developing non-invasive technique for cervical cancer detection.
Analyzing microarray data from cancer cell lines.
Investigating relationships between soil trace elements and cervical cancer mortality.
Parameter estimation for determining tissue elasticity.
Predicting risk of a major adverse cardiac event (MACE).
Detect QRS complexes.
Detecting hypoglycemia EEG signals.
Predicting time to reach full cervical dilation.
Selecting optimal features of cardiotocogram recordings.
Identifying autism by analyzing gene expression microarray data.
Predicting outcome of patients with non-small cell lung cancer (NSCLC).
Diagnosing patients by classifying lung sounds into normal, wheeze, and crackle.
Choosing appropriate highly active antiretroviral therapy (HAART) to control HIV.
Improving the selection of gantry angles to optimize stereotactic radiotherapy.
Training robot for physiotherapy of the lower limb.
Estimating Cobb angle from torso asymmetry in scoliosis.
Analyzing mutations in Parkinson’s disease.
Predicting of tacrolimus blood levels, scheduling patient admission in ophthalmic hospital.
only a few health conditions, e.g. cancer, neural disease and and classification; and deep learning for diagnostic imaging
cardiovascular disease. They criticized that existing health- and electronic diagnosis.
care systems do not provide incentives for data sharing and
have no structure for the implementation of AI. Further
classified applications of AI consist of ML algorithms to Data analytics and ML for disease identification
extract and cluster useful information from a large patient in EHR (44)
population to assist in making real-time inferences for Managing large volumes of data is a huge challenge for
health risk alerts and health outcome predictions, perform doing big data analytics. A group of researchers used an
principal component analysis and reduce diagnostic and ML algorithm for structured and unstructured big data
therapeutic errors; SVM to determine model parameters analysis to identify a wide range of medical conditions
and identify imaging biomarkers; NLP for text processing and diagnosis from the large-scale EHR database, inclu-
ding information about test results, historical information, EHRs; incorrect diagnosis and medication codes (e.g. ICD,
management plans, and billing codes, etc. Authors initially National Drug Code (NDC)), and their conversion into
applied their algorithm for the identification of potential research descriptions and vice versa; data extraction and
predictors and keyword associations in the EHR to find integration from clinical notes in to EHR; multiple data
presence of pseudoexfoliation syndrome (PXF) by using sources in EHRs affecting data provenance; mismatches of
NLP, least absolute shrinkage and selection operator. The data granularity with comparative research; and differences
reported outcomes were based on positive and negative pre- in research protocols and clinical care.
dictive values of the algorithm, which were also validated by
the glaucoma specialists. Authors highlighted the potential Deep learning in the healthcare system (47)
of their algorithm with its application for data mining and
Deep learning is a dominant ML approach, which has been
predictive analytics in many other applications as well.
greatly augmented in healthcare, analytic systems develop-
ment and modeling tools. This review presented potential
AI, big data and cancer (45) healthcare applications based on multiple driving factors,
including learning digital imaging (e.g. radiology, radiother-
The application of cognitive computer systems has been
apy, pathology, ophthalmology and dermatology) with deep
very successful for approaching cancer diagnosis and treat-
learning to facilitate effective decision-making and therapy,
ment (read, remember, recommend and remind). Accord-
digitization of EHR and applying ML for curating and
ing to a team of researchers, cognitive computer systems
analyzing data sets, integrative heterogeneous data analysis
can support physicians by providing rapid access to accu-
using deep learning, applying deep learning for hypothesis
rate information and treatment procedures, assisting in
generation by identifying novel associations to establish
decision-making. Such systems have the potential to dis-
causation and causal pathways, appropriate deployment of
tribute cancer knowledge into clinical practice and remote
AI and ML-based platforms for remote monitoring and
areas worldwide, by optimizing clinical research and tri-
digital consultations and improving performance of deep
als with reduced bureaucracy and cost. They presented
learning with exposure to larger datasets. Authors predicted
the possibilities of AI in future cancer care and research,
safe, effective, efficient and humanistic care in the future
which includes developing international cancer networks,
with the successful application of ML.
identifying beneficial therapies for rare and highly aggres-
sive cancers, observing different therapeutic outcomes by
different parameters, analyzing associations of cancer with Deep learning to transform healthcare (48)
other disease-specific attributes, discovering new cancer eti- Deep learning has the potential to transform healthcare by
ologies, incorporating pertinent patient and cancer charac- outperforming clinical systems and modeling complex rela-
teristics into clinic-based uses, conducting economic broad- tionships among active hidden factors of data. This study
based cancer trials, uncovering genomic and molecular highlights the background, workflow and challenges of
events sensitive to existing or new treatments and analyz- deep learning with its successful implementation of digital
ing and developing new treatment pathways. Furthermore, image analysis, and some examples like analyzing fundus
authors predicted that large-scaled AI databases may bene- images of the retina to predict cardiovascular disease and
fit cancer programs and help cancer treatment and research. identifying images for melanoma, basal and squamous cell
carcinoma by matching sensitivity criteria. They suggested
developing and training classifiers based on the deep neu-
Use of operational EHR in comparative
ral network to make predictions on big clinical datasets
effectiveness research (46)
by using the nonlinear features for modeling regularities.
EHR has a great role in improving the quality and cost of Authors claimed that in the coming years, with the con-
healthcare, advancing biomedical science and facilitating tinuous increase in the volume of datasets and without
clinical research. Supporting this claim, this overview any enhancements to the basic learning techniques, more
presented some examples of running and accomplished promising results can be achieved by ML.
research projects, e.g. Electronic Medical Records and
Genomics (eMERGE) Network, Strategic Health IT
Advanced Research Projects (SHARP) Program and Health Intelligent digital pathology with deep learning
Maintenance Organization Research Network’s Virtual (50)
Data Warehouse Project. The authors also discussed multi- The data efficiency of deep learning can be used to augment
ple caveats in existing healthcare systems e.g. inaccurate information by improving diagnostic accuracy and effi-
data entry in EHRs; incomplete patient information in ciency by analyzing whole-slide pathology images, which
cannot be easily observed by the human eye. Performance informatics field should involve DL, new ML algorithms
of deep learning at test sets of 129 whole-slide images and advanced big data analytics.
was examined and compared for identifying metastases
in hematoxylin and eosin–stained tissue sections of lymph Intelligent health data analytics to improve health
nodes from women with breast cancer. Authors of the study system management, health outcomes,
optimized deep learning algorithms for obtaining impres- knowledge discovery and healthcare
sive results, even better than the participating panel of 11 innovation (57)
pathologists. They predicted a bright future of deep learning
Various healthcare systems are generating heterogeneous
applications in digital pathology and discussed the need of
data of significant volumes which demands exploitation
new intelligent tool development for handling diagnostic
of healthcare data for resource optimization, patient satis-
sensitivity and specificity.
faction, improved care quality and health outcomes. Rec-
ommended use of AI and ML for advanced health data
ML for predictions and finding nonlinear analytics with the arrangement of active partners in the
relationships in the EHR (51) healthcare processes requires ubiquitous services, thera-
ML real-time applications have been used to address dif- peutic decision support, ethnographic health surveillance,
ferent clinical questions by analyzing and finding nonlinear integration of health-related data sources, personalized and
relationships in the EHR. A research study highlighted one predictive medicine by learning non-linear associations and
important aspect that could be one of the ML drawbacks drawing causal relationships among inherent data elements.
and that is the utilization of most of the ML algorithms to A team of scientists designed a health data analytics process
solve clinical problems when they were originally proposed involving a methodical order of data processing, modeling,
for other matrices. The authors of the study provided and analysis steps categorized as data- and knowledge-
some examples to justify their claim and discussed that the driven methods for Decision, Predictive, Descriptive, Opti-
primary purpose of operational EHR systems development mization, Comparative, Prescriptive and Semantic analy-
is to not support any kind of ML implementation for sis. They advocated implementation of AI and ML-based
predictions. Due to the lack of robust and structured clinical analyses with the inclusion of health data pre-processing,
data, authors showed uncertainty in achieving high-quality selecting algorithm based on expected outcome, developing
results with the use of ML. They stressed the need for new analytical models and interpreting results.
tools based on the ML algorithms especially designed with
prior thresholds for improvements. They also stated that ML in cancer prognosis and prediction (63)
the present discovery phase needs to implement ML for
ML can be applied to detect key features by predictive
better understanding of heterogeneous treatment effects to
modeling of complex and heterogeneous datasets for
implement precision medicine.
progression and treatment of cancerous conditions, risks
and outcomes. Many research groups have highlighted
Analyzing and visualizing knowledge structures the implementation of different ML algorithms in cancer
of health informatics to uncover the explicit and research for estimating unknown dependencies to predict
hidden patterns (55) new outputs of the system, including artificial neural
A quantitative review of the health informatics field uncov- network, Bayesian network, SVMs, graph-based semi-
ers the scientific growth in this field. As discussed, despite supervised learning (SSL) and decision tree. There are
a long period of progress, there is still no proper char- numerous ML real-world applications. The artificial neural
acterization of the knowledge and no common language, network has been established for breast cancer by analyzing
which necessitates illumination of the knowledge struc- mammographic and demographic data with k-fold cross-
tures, text mining methods, scientometric analysis, tracking validation rate and lung cancer by analyzing clinical and
tools, Internet of Things-based decision support systems gene expression data; SVM has been used for analyzing
and social network visualization to identify hidden pat- single-nucleotide polymorphisms (SNPs) with leave-one-
terns. The study conducted produced six clusters and nine out cross-validation for multiple myeloma, with k-fold
research themes to support health information technol- cross-validation for breast cancer and with k-fold cross-
ogy by improving patient safety by reducing medication validation for breast cancer by analyzing clinical and
errors, and associating decision support, knowledge repre- population data; the Bayesian network can efficiently
sentation, telehealth innovations and professional behav- analyze clinical and pathologic data with cross-validation
ioral changes in medicine. For the provision of better- for colon carcinomatosis and with k-fold cross-validation
personalized treatment, authors suggested that the health for oral cancer by analyzing clinical and imaging tissue
genomic; blood genomic data; SVM hold-out for breast medicine, which can support large-scaled clinical data inte-
cancer by analyzing clinical, pathologic and epidemiologic gration, and communication between different EMRs acces-
data and for cervical cancer by analyzing clinical and sible to patients through different health centers. Authors
pathologic data; graph-based SSL algorithm with k- proposed a universal EMR platform development by inte-
fold cross-validation for colon cancer; breast cancer by grating population perspectives for the establishment of
analyzing protein–protein interactions (PPIs) and gene protocols to identify subgroups that fit distinct clinical
expression data and k-fold cross-validation for breast phenotypes of complex disease and provides avenue for dif-
cancer by analyzing Surveillance, Epidemiology, and ferent treatment methods. Authors suggested an additional
End Results (SEER); SVM with cross-validation for oral fifth ‘P’—Population-wise Perspective—to the existing 4P
cancer by analyzing clinical and genomic data; SVM precision medicine concept. They justified this claim by
with leave-one-out cross-validation for breast cancer by examining a study about identification of a type 2 dia-
analyzing genomic data; Bayesian network with hold-out betes subgroup through a topological analysis of a popu-
for breast cancer by analyzing clinical, microarray data; lation’s EMR and determined three major distinct clusters
decision tree with cross-validation for breast cancer by (66). Authors also suggested that it might be difficult to
analyzing SEER; SSL co-training algorithm with k-fold implement ML divisions, which include supervised learning
cross-validation for Breast cancer by analyzing SEER. (SVM, discriminant analysis, Naïve Bayes, nearest neigh-
Authors acknowledged the active contributions of ML bor and neural network), unsupervised learning (linear
in accurate cancer susceptibility, recurrence and survival & logistical regression, decision tree, cluster analysis and
predictions. However, some concerns exist due to lack of neural network) and deep learning. Authors recommended
external validation regarding the predictive performance multi-cluster environment implementation for analyzing
of models utilizing integrated clinical and genomic data. patterns within the subset of populations, which might
They underlined that application of ML methods in present similar clinical phenotypes of complex diseases with
cancer requires significant validation in order to utilize the assumption that the treatment based on one cluster
them in everyday clinical practice. Authors suggested might not be as effective for another subgroup.
data pre-processing with focus on data modification via
dimensionality reduction and feature detection.
Solving healthcare problems with precision
medicine (35)
Precision medicine approaches Precision medicine has been groundbreaking in tailoring
The underlying assumption here is that precision medicine individualized and effective medical treatments based on
will provide tailored healthcare to patients and will yield the characteristics of each patient, and different suscepti-
lower rates of associated adverse outcomes. A classic exam- bilities to a particular disease, e.g. trastuzumab for HER2-
ple of precision medicine is the customization of disease positive breast cancer. Many authors are convinced by
treatment for a single individual which in the old paradigm the importance of information technology and ML for
was a one-size-fits-all medicine; an effective treatment is the the implementation of precision medicine, which includes
treatment known to benefit most of the target population. data storage and analysis for determining the association
However, a certain treatment may actually yield benefit between disease outcome (e.g. disease risk, prognosis or
to only a few individuals. The rest of the population will treatment), identification of patient characteristics and opti-
not benefit from the treatment and may even incur adverse mal treatment. Furthermore, authors highlight the require-
effects. This exemplifies the need for AI and ML-based sys- ment for multidisciplinary collaborations between clini-
tems bridging multiple domains in a secure environment for cians and researchers.
heterogeneous healthcare data analysis and visualization.
Role of AI in patients the point-of-care, advanced

Precision medicine with EMR analysis for analytics and foundation of precision medicine
prevention and treatment of diseases (33) (36)
Different healthcare institutions do not necessarily utilize The scientific community criticizes the current healthcare
the same EMR system with the possibilities of effectively structure, to be based on one-size fits all, and not transi-
communicating with each other, which makes it difficult for tioning from trial and error to evidence-based medicine.
physicians to track patients’ overall medical history. A team The authors recognize the importance of innovative
of authors brought about the importance of population technologies (e.g. genome sequencing, health sensors,
perspective and need of foundation platforms for precision advanced biotech) and essential roles of AI in precision
medicine implementation. Authors proposed three-factor- are not even encoded. Challenges in Radiology are of
based approaches, which include point of care, large-scaled metadata mostly not formatted according to Digital
clinical datasets for training classifier and analytics and Imaging and Communications in Medicine (DICOM), and
finally building foundations for precision medicine. The reports are unstructured. Challenges in Pathology hold
author’s combined deep learning prediction with human generation of unstructured reports, and frequent utilization
pathologist’s diagnosis resulted in a success rate of 99.5% of unstandardized nomenclature. Challenges in Clinical
and reduced human error by 85% at an early stage. Authors Evidence & Outcomes consist of insufficient data types, and
also presented guidelines for implementing AI in precision unavailability of universally adopted data models. Chal-
medicine, which included creating ethical standards for AI lenges in Procedures mainly cover the process for approving
applications, gradual and incremental development of AI new procedural codes. Authors have discussed the evolution
and training medical professionals about AI. of EHR to effectively delivering personalized treatment.
With discussion of two case studies, they emphasized the
implementation of precision medicine for the personalized
Integrated precision medicine towards delivery of care based on patient-specific patterns of disease
targeted and personalized treatment for a given progression and determination of precise therapies.
patient (38)
There has been a lot of debate over the current state
ML knowledgebase with ontology for pattern
of clinical decision support and how it can be improved
recognition in personalized medicine (40)
in providing precision medicine. Many scholars have dis-
cussed involving ML in healthcare settings for the diag- Personalized medicine is a broad and rapidly advancing
nosis and better treatment of chronic disease in clinical, field in healthcare, primarily based on disease-related
translational and public health. Furthermore, they have clinical, genomic, metabolomics and environmental infor-
highlighted the significance of big data management, pri- mation. However, failure to correctly identify disease
vacy, de-identification and data sharing. Focus on early is one of the major reasons for misleading diagnosis,
diagnosis of chronic conditions through proper extraction treatment and prognosis for the patient. A research
of clinical insights and utilization of in silico datasets could group presented essential components (pattern recognition,
allow the replacement of animal and human models when knowledge base, ontology and patient profile) for accurate
conducting clinical trials by generating virtual patients with disease examination needed for successfully integrating
specific characteristics that enhance the outcome of each personalized medicine into everyday clinical practice.
study. According to them, 25% of drug discovery occurs by Utilizing ML approaches for pattern recognition and
chance, which can be highly accidental. They signify using development of statistical models (sample size and effect
predictive, proactive intervention in healthcare through AI size), creating a knowledgebase of all existing phenotype
and clinical decision support system development for lever- categories and disease, organization of clinical datasets of
aging big data analysis to make better predictions on the population size and open software platform development
potential outcome of patients, ultimately supporting better for statistical analysis of high-dimensional healthcare and
decision-making by the physicians. multi-omics data are crucial for practical realization of
Authors have reviewed and reported analysis of the precision medicine.
different clinical decision support states, improvements in
patient outcome and various limitations and challenges in
Laboratory, Medication, Diagnosis, Radiology, Pathology, ML classifies cancer by visual assessment of
Clinical Evidence & Outcomes, and Procedures, based on tumor cells (54)
the availability of different data structures and standards in One of the key technological advancements in the diagnosis
healthcare systems. Challenges in the laboratory mainly of brain tumors was microscope-based analysis. Overcom-
involve incorporation of genetic results into EMR in a ing the visual limitation of such techniques for leading dif-
searchable way; tests conducted at external labs cannot be ferent classifications of a given sample by different individ-
incorporated due to lack of standardization, not encoding uals, a group of authors discussed ML as a precise solution
lab tests with Logical Observation Identifiers Names and for accurate diagnosis by analyzing molecular data. A pre-
Codes (LOINC), missing genetic information in EHR. vious study trained an ML classifier at a maximum number
Challenges in Medication include over time and ineffective of images of tumors that were classified by physicians, as
drug combination. Challenges in Diagnosis surround it is not possible to get precise conclusions, especially when
handling of International Classification of Diseases (ICD) tumor is histologically indistinguishable. Authors discussed
codes, as not all codes are not billable and some diagnoses the application of supervised ML for analyzing central
nervous system (CNS) tumor-type genome-wide methyla- AI towards health in resource-poor settings (37)
tion data to identify methylation patterns. They mentioned Importance and utilization of AI to improve health
application of unsupervised ML to search patterns in the outcomes in low-income settings and regions has always
data sets to develop classification categories. Authors sug- been a point of discussion. One research focus has been on
gested application of ML for molecular analysis and visual clearly identifying problems and issues related to the inte-
inspection, as the disease can be the manifestation of both gration of global healthcare, application and deployment
molecular and cellular changes. of AI in real-time environments, high-quality healthcare
data collection in under-developed areas, automatic data
ML and genomics in precision medicine (62) extraction from hand-written notes in local languages,
availability of high-performance computing and cloud-
Authors studied major concerns regarding data security
based environments for data management and analysis, and
like breaches in patient privacy and also presented the
construction of knowledge bases and expert systems (67).
latest developments in the field of computational data
Furthermore, concerns about data security, consent, and
protection. They discussed advancements in the era of big
ownership, and ethical, patient safety and privacy-related
data for computer-aided diagnosis, and a recent concept
issues that accompanies the utilization of AI are also
of precision medicine for providing care for clinical, envi-
present. Many solutions have been presented to address
ronmental and genetic characteristics. Authors emphasized
the problems of implementation of NLP in EHR for
developing new ML algorithms for computational analysis
surveillance and predictions, AI implementation at pre-
of genomic data to systematize the process of finding
existing systems, weather and land pattern identification
genetic similarities among patients with similar prognosis,
and tracking disease transmission through ML algorithms,
or response to a treatment. They also stressed the need of
implementing advanced expert systems, utilizing NLP to
implementing intelligent procedures to ensure the privacy
translate hand-written notes according to WHO standard-
of data, as they believe that data anonymization is not
ized medical terminologies and developing local dictio-
enough to guarantee unidentifiability due to auxiliary
naries, establishment of environments capable of working
information. They discussed some existing solutions,
offline and synchronized with the remote databases,
which includes de-identification by data suppression,
adopting cloud computing for the implementation of public
k-anonymization, learning from noisy data, homomor-
health without established IT structure in Low and Middle
phic encryption, multi-party computation, cryptographic
Income Countries (LMICs), and using ‘blockchain’ for
hardware and protecting genomic databases. Authors
cryptocurrency and addressing issues related to the privacy
acknowledged the current contributions of ML but with
and transparency. The scientific fraternity has predicted
the expectations for substantial improvements to address
tremendous cost-saving and improved care delivery in
clinical data security problems and development of new
coming years with the implementation of such along with
significant ML models to address the challenges of gene
related tasks.
variations and similarities among patients. They suggested
combining the latest computational data protection
principles with legal and ethical perspectives to construct a Data science, AI and ML for laboratory medicine
secure framework for data sharing. (41)
Data science has demonstrated success in laboratory
medicine and reinforcing its value in transforming the
Healthcare resource management and healthcare system. This is a review on the perspective
ethical challenges applications of data science current problems in healthcare,
Resource management is very important in any field of which includes the need of significant computational power
life, especially in healthcare. Aligning people and technol- to contribute to solving data optimization problems, over-
ogy with organizational goals can positively impact with fitting in experimental designs, lack of data standardization,
efficient implementation of planned workflow in achieving large-scaled datasets for training ML classifiers and other
on time high-quality results. However, inefficient resource concerns in clinical laboratories, e.g. protected health
management may lead to over-exaggeration of organiza- information challenges, financial limitations and ethical
tional resources, which includes, time, cost, manpower and concerns. Authors discussed the concept that data science
computational, bench and infrastructure resources. Further- and AI mimics human processes and improves the process
more, it is important to address the ethical and data privacy of decision-making. They reasoned the use of predictive
challenges, when implementing traditional state-of-the-art modeling for better collaboration between hospitals
and intelligent healthcare data analytics. without sharing data and complying privacy regulation
through ML (supervised and unsupervised learning) for predictive and proactive interventions in clinical decision
healthcare data (clinical data, imaging, laboratory tests) support systems. They reviewed AI approaches, dividing
mining, analysis and optimization, identification of large them into three categories (43), artificial narrow intelligence
patient clusters with certain disease characteristics, reducing (performing a single task), artificial general intelligence
medical errors and cost and improving efficiency of staff (agent-based system, known as human-level AI) and arti-
and resources. Data science has its own applicability and ficial superintelligence (agent-based system, but smarter
ethical challenges especially related to AI integration. AI than better human). Authors debated the potential role
encompassing data science when adopted can even improve of AI in filling the human resource gaps by integrating it
human thought processes in efficient decision-making with with physicians to improve diagnostics and help in better
different ML algorithms, data mining, knowledge discovery, decision-making. At the same time, they also highlighted
improving complex analytical tasks and calculating the ethical implication of using AI technology as integrating
clinical pathways. Authors of a study acknowledged part of the healthcare system.
success stories of ML in healthcare (radiology, pathology,
dermatology and genomics), weather forecasting, structural
recognition, NLP, games, analyses of financial transactions High-performance medicine with AI for
and improvements in industrial processes by finding improving workflow and reducing medical errors
patterns, discovering inefficiencies, predicting outcomes (49)
and taking factual decisions. They provided examples There are multiple challenges of implementing AI in
of ML application in healthcare for the detection of medicine. A research group underlined the obstacles and
cervical cancer (neural networks), prediction of histological caveats, especially when applied in radiology, pathology,
diagnoses (artificial neural network-based decision-support dermatology, ophthalmology, cardiology, gastroenterology
scoring systems) and radical hysterectomies with gene and mental health. They predicted adaptation of AI by
expression analysis (neuronal network), DNA (epigenetic almost every type of clinician, which mainly includes deep
modifications) and RNA (messenger, long non-coding or neural networks for pattern recognition and analysis of
double-stranded) data analysis. ML is contributing to better medical images (e.g. medical scans, pathology slides, skin
personalized treatment by monitoring patient activities lesions, retinal images, electrocardiograms, endoscopy,
(analyzing data received by sensors for continuous mea- faces and vital signs); applying deep learning to EHR for
surement, e.g. glucose); analyzing molecular biomarkers; estimating the risk of a patient’s hospital readmission;
predicting drug efficacies, treatment responses and disease supporting doctors in decision-making for resuscitation;
pathways; and identifying molecular factors and genetic determining patients at risk of developing sepsis and
variants. The authors presented an ML framework, which other diseases; and predicting biological age, and critical
includes defining tasks, matrices, models and datasets. diseases leading to death. Authors highlighted current
They showed concerns regarding implementation of human challenges to the field of life sciences that can be solved
rights by the Universal Declaration of Human Rights by ML algorithms, like identification and isolation of rare
at the 1948 United Nations General Assembly for data cells, multi-omics data analysis, classification of somatic
privacy, protection, de-identification and encryption for and germline mutations and gene–gene interactions and
data handling, collection and sharing. Authors emphasized prediction of protein structure and PPIs, the microbiome
the consideration of key ethical concerns, which include and single cells. They also discussed the use of NLP in
consenting patient, AI human warranty and regulation drug discovery by analyzing biomedical literature; mining
of healthcare data according to principles of bioethical molecular structures; predicting off-target effects, toxicity
law. and right dose for experimental drugs; developing cellular
assays; and using AI cryptography for determining uniden-
tified drug interactions. Together with these advantages,
AI to solve the human resource crisis in authors also discussed formidable obstacles and drawbacks
healthcare (42) due to the field of AI, which includes data privacy and
The healthcare workforce crisis is widening across the security, data hacking and breaches, uncertainty in use of
globe. With the increase in the number of chronic and black boxes of algorithms to resolve output, state of AI hype
complex diseases, the demand towards efficient healthcare in validation and readiness for implementing in patient care.
system is consistently growing. However, the lack of access Authors anticipated the bright future of AI applications
to care, differing quality and doctor shortages are increasing by predicting useful clinical outcomes in health systems,
worldwide. Some authors have presented AI as the solution algorithmic interpretation of images and data, reducing
in healthcare and discussed its progress in the direction of errors, inefficiencies and cost.
Big data and ML algorithms for better healthcare deep learning models for feature extraction, development of
delivery (56) graphical processing units, predicting in-hospital mortality,
ML is considered a branch of AI encompassing algorithmic unplanned readmissions, prolonged stays, and discharge
methods to solve scientific, healthcare and various other diagnoses. However, they showed lack of confidence in ML
problems without traditional computer programming; for shielding against the normal challenges in observational
this involves interplay between large population-specific data analysis, as screening some concerns require evidence.
datasets and model building. Suggested AI tool develop-
ment should be based on incremental learning (ability to Promise and potential of big data analytics in
continuously improve with the inclusion of new data) to healthcare (61)
refine predictive accuracies. One study proposed analytics The effective use of big data by digitizing, combining large
exercise starting with the identification of clinical problems, hospital networks and implementing efficient analytical
annotation of extracted healthcare data, application of approaches has been widely successful. According to a
appropriate ML algorithms and its effect on decision- team of scientists, there are tremendous benefits of such
making, addressing legal and ethical implications, assess- approach in detecting diseases at earlier stages; predicting
ment of ML effect in trial, designing freeze and submission risk for medical complications; managing individual
of dossier for medical devices, training clinicians on use of and population health, avoiding frauds; and addressing
ML tools and monitoring for adverse outcomes. Authors numerous healthcare questions. Furthermore, it can reduce
of the study presented ML algorithms suitable for different waste and inefficiency by determining clinically relevant
clinical applications and reported real-time clinical devel- and cost-effective ways to diagnose and treat patients,
opments in oncology with the implementation of different applying ML algorithms to predict models and analyzing
AI and ML applications and algorithms. Authors discussed EHR and disease patterns to discover adverse treatment
the importance of human–computer interaction in AI and effects. Authors also discussed the current challenges,
ML platform development to effectively support decision- covering unavailability of user-friendly and transparent
making processes in healthcare. They presented different real-time big data platforms; lags between data acquisition,
real-time examples, which include HealthSuite, GE Health- collection, cleansing, processing and standardization;
care, Lumada, DISCOVERY and CURATE.AI. They recom- missing ability to manipulate data at different levels for
mended model linking EHR in future ML platform devel- granularity, privacy and security enablement and quality
opment for addressing healthcare data analytics concerns. assurance; difficulties in the management of large, diverse
and complex data with traditional approaches; changing
Causal inference with ML (60) healthcare reimbursement models; lack of professional
tools, infrastructure and techniques to leverage big data
Another study examined the implications of progress of AI
effectively; and absence of dynamic analytics algorithms
in observational research design and healthcare databases,
for efficient data modeling. Authors presented a conceptual
and implementation of ML for data classification and pre-
architecture of big data analytics, from developing multi-
diction in Real-World Evidence (RWE) to support clinical
source data input, transformation, structure, management
and regulatory decision-making. They discussed American
and analysis using traditional SQL, OLAP, to data mining.
Recovery and Reinvestment Act (ARRA) explicitly pro-
hibiting cost-effectiveness and focusing on broad interven-
tions for better diagnoses, treatments, disease management Unintended consequences of ML (52)
programs and healthcare organization models by mainly Authors reviewed some of the factors driving wide adoption
analyzing EHR and establishing patient-centered Outcomes of deep learning and other forms of ML in the health ecosys-
Research. Authors discussed a wide range of increased size tem. They stated that ML has the potential of promoting
and nature of biomedical operational and research data, changes in specialty that requires accurate prognosis mod-
which requires efficient utilization of ML approaches to els (e.g. oncology) and pattern recognition (e.g. radiology,
restructure, dimensionality reduction, clustering, modeling, pathology). However, authors anticipated some unintended
linking, classification, analysis and predictions. Meeting consequences due to the application of ML-based deci-
RWE objectives, authors acknowledged ML as one of the sion support systems in healthcare, which includes overre-
powerful tools today for bringing significant improvement liance on automation, more potential for decision errors
to care service. They exemplified ML with the implemen- by physicians, misleading diagnosis, misinterpretation of
tation of SVMs to predict hospitalization; regression-based data, intrinsic uncertainty in medicine, inevitable intrinsic
methods for reducing the risk of overfitting; k-fold cross- uncertainties and rationale and inscrutable outcome of ML
validation for splitting one’s sample into two models; and algorithms. Authors presented unintended consequences in
clinical and operational research support in reducing the and researchers from policy enactment, programming, task-
odds and better implementing ML in medicine. Authors forces (59), etc.
appealed for developing more precise analytics platforms
for pathology images, next-generation radiology tools, con-
verting EHR into reliable risk predictors and monitoring Theoretical background of AI, ML and
patients’ health through wearables and personal devices. examples in healthcare
AI aimed to improve the intellectual capabilities and perfor-
Finding the missing link for big biomedical data mances of machines to solve complex and big data-oriented
(53) problems by classifying interaction patterns among vari-
ables, learning from experiences, strategizing and predicting
The problem of integration and analysis of large-scale
better orientations. AI has been in business for almost over
biomedical data located in heterogeneous data sources can
50 years, and its applications are heavily demanded in
be solved by intelligently conforming biomedical data to
numerous fields of life, science, technology and medicine
support physicians and researchers in conducting new stud-
(68,69). AI developments are based on supervised, unsuper-
ies and drawing new hypothesis, leading to novel inter-
vised and reinforced learning principles, which include com-
ventions. This study lists challenges in biomedical data
putational command line, desktop, web-based, robotics and
linking, which includes identification of potential sources,
smartphone applications with different analytics capabili-
determination of linking values, lack of a national unique
ties, e.g. machine translation, speech recognition, NLP, data
patient identifier and data privacy and security concerns.
mining, risk modeling, image recognition, machine vision,
The authors also presented some potential solutions to these
knowledge bases, expert systems and agent-based systems.
challenges and support personalized treatment, urging the
AI is categorized in four main types for decision-making,
identification of potential sources of health information,
which are reactive machines (based on the current situation
considering non-traditional data (e.g. social media, pur-
without learning from experiences), limited memory (based
chase history, census records, etc.) to assemble a universal
on short memory, and learning from experiences), theory
view of a patient and probabilistic linkage to the unique
of mind (based on humanlike capabilities and abilities
patient identifier problem. They recommended future (AI
to attribute mental states) and self-awareness (based on
and ML) tools development to analyze biomedical data for
human-level consciousness).
better clinical decision-making.
In the last few years, AI has become more popular and
seriously considered for analyzing diverse clinical data
Ethical challenges of implementing ML in (EHR, images, etc.) for accurate diagnosis and effective
healthcare (58) treatment in different practices (70), e.g. radiology (71)
The current trend towards AI and ML algorithm devel- (early diagnosis, enhance visualization of pathologies and
opment and application comes with its set of issues and predicting emergency situations (64,73,71)), oncology
challenges. Despite the helpful contributions of AI and ML (diagnosis of breast (75), skin (76), lung cancer (77)), cardi-
in healthcare, they have some concerns, especially related ology (interpreting electrocardiogram readings, echocardio-
to the ethical challenges, which include unexpected risks graphy with 3D cardiac imaging, cardiac CT angiography
due to the utilization of algorithms in medicine originally for calcification of the coronary vessels, cardiac MRI for
proposed for some other development; likelihood of using measuring perfusion and blood flow and longitudinal
ML as the source of the communal medical mind; mirroring evaluation to find predictors of heart failure (78,79,80)),
human prejudices in decision-making by learning from dif- gastroenterology (analyzing endoscopic images for screen-
ferent unforeseen biases; in extreme premature conditions ing regimens to abnormal findings (81)), ophthalmology
relying on ML results can lead to fatal conclusions; mostly (detection of diabetic retinopathy in retinal fundus
applied ML algorithms are designed to perform in uneth- photographs (64)), pediatric (augmenting diagnostic
ical conditions; ethical strain of gaining industrial profits evaluations (82)) and surgery (robotic-assisted surgery
with clinical decision-support systems without informing (83,84)), but not limited to these (85). One of the
its users; increase in ethically problematic outcomes due most recent trends is the utilization of AI in Precision
to constructed ML black boxes; and reimagining of con- Medicine (86,87,88) with the application of ML algorithms
fidentiality due to ML. A review justifies these limitations for analyzing heterogeneous patient data, e.g. clinical,
in AI and ML with some examples, e.g. Uber’s software tool genomics, metabolomics, imaging, claims, labs, nutrients
Greyball and Volkswagen’s algorithm. Authors are hopeful and life-style.
that these issues will be addressed in future ML-based ML is a branch of AI that utilizes and proposes different
healthcare systems, with the involvement or professionals algorithms for learning from numerous data variables and
Figure 2. Data classification, clustering and regression for healthcare data analytics. ML application process includes creating and labeling of raw
data, training classifier for data modeling using appropriate algorithm and analyzing and reporting results.
revealing multifaceted relationships among data features from death certificates and autopsy reports; identification
to predict accuracies in different contexts and support of heart failure, cancer and other chronic disease causing
decision-making processes. The overall ML process starts symptoms; risk predictions, interventions, paneling and
with the manifestation of either or both data and labels, resourcing (97,98) and clinical decision-making (99). Most
training classifier to learn and model using an algorithm and commonly used ML algorithms in medicine includes SVM,
then performing data evaluation and analysis to estimate deep learning, logistic regression, DA, decision tree, random
final results. ML is mainly categorized in three learning forest, linear regression, Naïve Bayes, K-nearest neighbor
approaches: classification, cluster, and regression (Figure 2). (KNN) and hidden Markov model (HMM) (Figure 3).
Classification and regression are based on supervised learn- SVM (100) is one of the most widely used ML algorithms
ing, while clustering involves unsupervised learning. Clas- in bioinformatics and healthcare (101). It is considered a
sification predicts discrete, categorical response values by great supervised learning method (102,103) for accurately
using labels and parameters, e.g. determining if a biopsy working in general practice. SVM was proposed in 1963
sample is cancerous or not (positive/negative). Cluster is to (100) to model diverse and high-dimensional data (101)
partition data into sub-groups, e.g. what is the prevalence with kernel methods to generate nonlinear decision bound-
of disease recurrence (positive/negative) in a certain pop- aries and train classifier (104,105,106). It assists in the
ulation due to pollution or chemical spill (common rela- field of medicine with vast variety of contributions, e.g.
tionship between dataset). Regression predicts continuous- symptoms classification and analysis to improve diagnostic
response numeric values to identify distribution trends, e.g. accuracy, identifying imaging biomarkers of neurological
how long before a patient is readmitted to the hospital and psychiatric disease, validation for multiple myeloma
following his/her discharge (positive/negative). and breast cancer by analyzing SNPs, hold-out for breast
ML is becoming the transformational force in health- and cervical cancer by analyzing clinical, pathologic and
care (89) for guiding individual and population health epidemiologic data, validation for oral cancer by analyzing
with several computational benefits, which includes real- clinical, molecular and genomic data, and diagnosing men-
time patient monitoring; disease patterns analysis (90); tal illness (106,107,108). Its ML benefits include modeling
disease diagnosis and prescription of medicines; patient- nonlinear class boundaries, unlikely overfitting, reduced
centric care provision with enhanced treatment (91); clinical computational complexity to quadratic optimization prob-
errors reduction (92); prognostic scoring (93); therapeutic lems and controllable complexity of decision rule and fre-
decision-making (94); identification of sepsis and high risk quency of error (109). However, the complex structure
for medical emergencies (95); identification of phenotypes; of this algorithm is its limitations leading to slow data
screening claims data (96); extraction of clinical codes processing speed, and difficulties in determining optimal
Figure 3. Applying machine learning algorithms for clinical, genomics, metabolomics, imaging, claims, labs, nutrients and life style data fusion,
integration and analysis. Machine learning algorithms include, support vector machine, deep learning, logistic regression, discriminant analysis,
decision tree, Random forest, linear regression, naïve Bayes, K-nearest neighbor, hidden Markov model and genetic algorithm.
parameters especially when training dataset is not linearly widely used algorithms in medicine for analyzing different
classified (108). types of images from various healthcare disciplines but
Deep learning is a dominant approach based on artificial especially oncology, e.g. thoracic (lung cancer); abdominal
neural network. It was inspired by the ability of the brain and pelvic (computed tomography (CT) and magnetic
to learn complicated patterns, and one of the most popular resonance imaging (MRI)); colonoscopy (colonic polyps);
ML algorithms in healthcare today, aiming to advance mammography (microcalcifications); brain (brain tumors);
clinical medicine and care delivery (47). The artificial neural radiation oncology (segmenting tumors for radiation,
network base gives its several nested layers of neurons and quantifying specific radiographic characteristics by
(110) to learn complex relationships between features and analyzing 3D shape of a tumor (113,114)); dermatology
labels from heterogeneous clinical data (32). Deep learning (skin cancer (115,76)); pathology (digital whole-slide
has four categories: deep belief network, deep neural net- of biopsy samples) (115); prostate (cancer tumors by
work, convolutional neural network and recurrent neural analyzing ultrasound of biopsy cores); malignant lung
network (39) for implementing pattern recognition and nodules by analyzing radiographs; glioma by analyzing
predictive modeling at high-dimensional big data sets. It is histological imaging and genomic marker data; and
fundamentally a uniquely different paradigm in ML DNA and RNA sequencing (RNA-binding and DNA-
(111,112) and its capabilities include multitasking, auto- binding proteins). Furthermore, deep learning has been
matic construction of complex features, digitization of applied for the diagnosis of several other diseases, e.g.
EHR and image-based data, integrating heterogeneous nodular BCC, dermal nevus and seborrheic keratosis in
data sets assembled from diverse sources, combining dermatopathology (116); diabetic retinopathy (64); type 2
with wearables for remote monitoring (47). Its ML diabetes subgroups (66); diabetic retinopathy by analyzing
benefits include application in classification or regression retinal images; histological prediction in women with
with ability to represent Boolean functions (AND, OR, cytological abnormalities; measure medication adherence
NOT), handling noisy inputs and classifying instances for via camera interface (65); and cardiac anomalies and
more than one output. However, its limitations include congestive heart failure by analyzing images of MRI of
difficulties in understanding structures of algorithms, heart ventricles (117).
possibility of too many overfitting attributes and its Logistic regression is a statistical method to assess the
optimal network structure which can only be created by relationships between various predictor categorical and
experimentation (108). Deep learning is one of the most continuous variables and dichotomous binary outcome
(118). It has been applied in the fields of medicine (clinical order of training algorithm instances with no effect,
practice, surgery), epidemiology and biochemistry to and no overfitting problem while pruning, allowing
perform predictive and explanatory modeling by obtaining predictive model implementation with high precision,
odds ratios and risk factors explaining variations in permanence and ease of clarification, well-mapped non-
specific outcomes (119,120,121,122). Its examples of real- linear relationships and suitability for both classification
time implementation in the field of medicine include risk or regression problems. However, its limitations include
assessment of complex diseases (123) (e.g. tuberculosis mutually exclusive classes, dependency on order of attribute
(124), breast cancer (125)), predicting patient survival selection, being error-prone when training on excessively
rate and diagnosing coronary heart disease (CHD) (123), complex decision tree, time-consumption and branching
non-Hodgkin’s lymphoma diagnosis with multivariable of missing values for an attribute (108). It has been well
logistic regression modeling (126), and identification applied in the field of medicine for real-time healthcare
of pulmonary thromboembolism (PTE) by analyzing monitoring, medical decision support system, anomaly
prognostic factors (127). Its ML benefits include concurrent detecting and sensor and a data mining model for pollution
multiple explanatory variables analysis with reduced effect prediction. A few real-time examples include supporting
of confounding factors and modeling categorical dependent clinical decisions (147), strategies for alternating therapies
variables (128,129). However, its limitation lies in handling in oncology patients (148), collaborative clinical decision-
continuous explanatory variables with more than two levels making in mental health care (149), identifying predictors
as it is based on variables with a constant range of values of health outcomes (150), finding factors related to hyper-
(122), difficulties in understanding odds and probabilities tension (151), discovering factors associated with pressure
(119), right predictor variable selection (118,122), reference ulcers (PUs) among elderly people (152), identifying the
group setup and managing relationships between input and potential recipients of telehealth services (153), patient
output variables. data stratification for interpretable decision-making for
Discriminant analysis is a widely applied technique in precision medicine (154), content analysis for patient aids
medical studies for pattern recognition (130). Discriminant decision (155), diabetic foot amputation risk analysis
analysis is used to predict and classify group members by (156) and support understanding of antenatal lifestyle
building one or multiple functions (131,132,133) based interventions (157).
on normal distribution and equal variance–covariance of Random forest is also known as the random decision
independent variables (134,135). Its real-time applications forest, a combination of algorithms to build predictive
include identifying surgical and operative factors to accu- models for classification and regression problems (158).
rately classify patients for surgical procedure (136), predict Its classifier generates a set of decision trees based on
the clinical diagnosis of primary immunodeficiencies training of randomly selected subsets to aggregate the elects
(137), patients’ symptom-relief satisfaction data (138), from different decision trees to get final object. Its ML
prediction of depression elements in cancer patients (135), benefits include overcoming the problem of overfitting,
classification of BOLD fMRI response to naturalistic movie less variance, not requiring input data preparation, flexi-
stimuli (139), identify protein coding regions of rice genes bility and high accuracy even with missing large portions
(140), Parkinson’s disease symptoms recognition (141), of the data. However, its limitations include its complex-
risk assessment for chronic illnesses (142), diagnosis of ity, difficulty in implementing, requirements for additional
hypercalcemia (143) and predicting patient care visits by computational resources, less intuitiveness and more time
identifying discriminatory characteristics (144). Its ML consumption than most other algorithms. Nevertheless, it
benefits include robustness, reduced dimensionality, easy has been applied in the field medicine for data mining,
implementation as it requires fewer parameters to be real-time patient monitoring, disease classification, imple-
estimated (137). However, its limitations include over- mentation in wearables and personal devices and mod-
fitting to dataset (131), limit performance at novel datasets eling big data based on engine recommendations. Some
(145) and lack of cross-validation (137). of its reported contributions include diagnosing mental
The decision tree utilizes a tree structure modeling illness (106), detecting knee osteoarthritis (159), monitor-
approach with conditional control statements for establishing medical wireless sensors (160), diagnosing Alzheimer’s
ing an efficient decision-making process (146). The decision disease (161), predicting metabolic pathways (162), pre-
tree is based on the concept of classification rules, following dicting outcomes of a patient’s encounter with behavioral
paths from root to leaf. Its internal nodes represent ‘test’ health providers (163), healthcare cost prediction (164),
on an attribute, branch represents the outcome of the mortality prediction for intensive care unit (ICU) patients
test and leaf represents decision taken after computing all (165), classification of Alzheimer’s disease (166), identify-
attributes. Its ML benefits include ease of understanding, ing social and economic factors to study social determinants
of health (167), predicting disease risks from imbalanced updating description concept, expecting similar classifica-
data (168), identifying associates of diabetic peripheral neu- tions and relevancy from instances with similar attributes
ropathy diagnosis (169), predicting the risk of emergency and increase in computational complexity with the number
admission (170), detecting patients ready to discharge from of attributes (108). KNN has been used in many scientific
intensive care (171), nonparametric estimation of hetero- fields but with limited applications in the field of medicine
geneous treatment effects (172), diagnosing sleep disorders (195). Some of real-time examples include modeling diag-
(173) and predicting depression in patients suffering from nostic performance (195), preserving privacy of medical
Alzheimer’s disease (174). diagnosis in e-Health cloud (196), medical dataset classi-
Linear regression is an ML approach to model relation- fication (197), classification of lymph node metastasis in
ships between dependent and independent variables using gastric cancer (198), pattern classification for breast cancer
linear predictor functions to identify errors of prediction diagnosis (199), pattern classification for health monitor-
in a scatter plot and characterize relationships among ing applications (200) and pancreatic cancer prediction
multiple factors (175). Linear regression has been applied combining published literature and EHR data (201).
in the field of medicine for many computational analyses HMM was originally proposed to solve speech problems
and predictions, from identification of prognostically by making complex and instinctive probabilistic models for
relevant risk factors (175), predicting hand surgery (176), finding and processing hidden states and paths (202,203,
monitoring treatment prescribing patterns and ensuring 204). Since the late 1980s, it has been effectively applied in
its appropriateness (177), averaging decision-making in the field of life sciences, especially in biology for analyzing
healthcare (178), reducing high costs of the health system sequence data (e.g. predicting exons and introns, identifying
(179), analyzing skewed healthcare cost data (180) and ORFs, insertions, deletions, substitutions, functional motifs
understanding human immunodeficiency virus (HIV) in proteins, aligning two sequences and switching from
prevalence patterns (181). exon to intron in a DNA sequence (205)) by capturing
Naïve Bayes is a supervised ML technique based on hidden information from observable sequential symbols.
Bayes’ theorem for data mining, classification, and predic- Later, it was well adapted in the field of medicine and
tive modeling (182) to find a maximum probability value its real-time contributions include modeling ‘Healthy’ and
from a conditional probability chain (183). Naïve Bayes ‘Unhealthy’ unobserved health states (206); analyzing time-
is well applied in the field of health performing predictive series data on personal health check-up (206); improving
modelling for different diseases (brain, asthma, prostate, length of hospital stay prediction and reducing health care
and breast cancer etc.) (184), predicting risk using censored costs (207); mining adverse drug reactions from online
and time-to-event data (185), detecting Mucopolysaccha- healthcare forums (208); monitoring, modeling and clus-
ridosis type II (186), predicting Alzheimer’s disease from tering medical inpatient journeys (209); analyzing health-
genome-wide data (187), measuring quality healthcare ser- care service utilization after transport-related injuries (210);
vices (188), finding audit targets in performance-based monitoring circadian in telemetric activity data (211); pre-
financing in health (189), modeling medical diagnosis for dicting patients entering states with a high number of
decision support (190), classifying EHR (191), classifying asynchronies (212); and analyzing subject-specific seizure
and genetic scoring for risk prediction (192), and designing and automatic segmentation of infant cry signals (213).
a decision support system for heart disease (193). Its ML However, its modeling limitations include computing prob-
benefits include modelling based on statistical foundation, ability of sequence observation, choosing an accurate corre-
ease to understand and train algorithm, and usefulness sponding state, adjusting parameters (214) and dealing with
across multiple domains. However, its limitations include correlations between residues due to the underlying decency
difficulties in handling redundant attributes, distribution assumption problem (215).
of statistically independent attributes, and management of The genetic algorithm was inspired by Charles Darwin’s
class frequencies affecting accuracy (108). theory of natural evolution to solve constrained and
KNN is also a supervised learning-based algorithm used unconstrained data optimization and standardization prob-
for classification and regression in pattern recognition, data lems by repeatedly modifying a population of individual
mining and intrusion detection by classifying points to given solutions (215). Primarily, it is based on heuristic search
categories from a training dataset (194). Its ML benefits with three active rules: selection of data elements (parents),
include implementation of instances for fast non-linear data crossover rules for two parents from children and mutation
classifications, robustness in addressing irrelevant or novel for random change (216). Its ML benefits include easier
attributes, well-handled instances with noise and missing implementation than other algorithms; application for fea-
attribute values and applicability for both regression and ture classification, selection and optimization; and relative
classification. However, its limitations include languidity in success. However, its computational limitations include
development of non-trivial scoring function, complications of clinical operations, research and analytics to achieve
in training classifiers of given data, and its being not the significant improvements in providing better individualized
best method to find optima. It has been vigorously involved and population healthcare at lower costs, and together
in the fields of life and medical sciences for benefitting with better work-life style for clinicians and staff. The
analytics, development in radiology, oncology, cardiology, major barriers in successfully implementing AI in healthcare
endocrinology, pediatrics, surgery, pulmonology, infectious include gaps in AI building blocks and infrastructures;
diseases, radiotherapy, rehabilitation medicine, orthopedics, low quality data availability; business model sustainabil-
neurology, pharmacotherapy, health care management, ity; regulations and policies for data collection, usage and
obstetrics and gynecology (217). It has been successful in sharing; disruption in the physician–patient relationships;
detecting microcalcifications in mammograms of breast integration to operational health systems; reduced evidence
cancer (218,219), developing a non-invasive technique for and reproducibility; selection of most appropriate ML algo-
cervical cancer detection (220), analyzing microarray data rithm; lack of understanding of AI and ML processes to
from cancer cell lines (221), investigating relationships predict; hazard of dehumanization of healthcare data and
between soil trace elements and cervical cancer mortality job insecurity threat; conflicts of interest and impartial
(222), parameter estimation for determining tissue elasticity access; accountability exploitation of AI; ‘sanity’ check to
(223), predicting risk of a major adverse cardiac event minimize any bias; handling of misleading and erroneous
(MACE) (224), detecting QRS complexes (225), detecting results; and data privacy, ethics, consent and ownership
hypoglycemia EEG signals (226), predicting time to reach (108,238,239,240).
full cervical dilation (227), selecting optimal features Determining which AI approach to use for which task
for cardiotocogram recordings (228), identifying autism is a challenge in itself. Classifying tasks based on available
by analyzing gene expression microarray data (229), predictor variables is a key step to correctly addressing
predicting outcomes of patients with non-small cell lung the problem. Traditional AI models can be used for simple
cancer (NSCLC) (230), diagnosing patients by classifying prediction tasks while complex tasks require more complex
lung sounds into normal, wheeze and crackle (231), models. The next step is to consider how the model will
choosing appropriate highly active antiretroviral therapy be used in practice. Setting rules to identify if a task needs
(HAART) to control HIV (232), improving the selection of to pre-process data before execution or if a complex task
gantry angles to optimize stereotactic radiotherapy (233), can be broken down into simple tasks is amenable to a
training robots for physiotherapy of the lower limb (234), traditional model. Creating and updating these rules is very
estimating Cobb angle from torso asymmetry in scoliosis time-consuming but is very useful. Availability of training
(235), analyzing mutations in Parkinson’s disease (236), data is another key factor in intelligent automation. Simple
predicting tacrolimus blood levels and scheduling patient models can work with little data and few variables, but
admission in an ophthalmic hospital (237). complex models require huge amount of data with mul-
tiple examples and scenarios as it is expected to remove
noisy data and learn to identify complex statistical patterns
Discussion (32). Numerous approaches have been proposed recently
AI and its application in healthcare could be another great to successfully target current challenges of implementing
leap in medicine and a transformational force for guiding AI and ML. We studied and reviewed contributions and
personalized and population medicine with several compu- variability analysis of various approaches in healthcare.
tational benefits. The extent of its popularity in healthcare We defined 15 different features to assess the potential
can be easily determined by the number of AI related pub- of discussed approaches (Table 3). These features are (i)
lished work in medicine. At the time of the study (09 April intelligent interface development; (ii) next-gen radiology
2019), a total of 16 166 AI and ML papers were available and imaging tools development; (iii) global expansion of
through PubMed. The growth of scientific literature in AI medical resources; (iv) automated ETL, linkage and data
increased in the last 10 years with 14 469 papers (2009– mining in HER; (v) risk prediction and containment of
2019), and more than 70% work was published in the last antibiotics resistance; (vi) pathology image analysis; (vii) AI
five years. While the growing importance and relevance of in machines and medical devices; (viii) smart solutions and
AI in healthcare is indisputable, to improve public sector methods for cancer treatment; (vix) EMR analysis for accu-
clinical practice, there is a critical need for development of rate risk predictors; (x) wearable devices for monitoring
intelligent frameworks to connect operational and analyti- patients health; (xi) smartphone applications as diagnostic
cal healthcare systems in a way that experts from multiple tools; (xii) AI-based clinician decision-making; (xiii) search
domains can perform measurement and predictive analysis. engine for healthcare data flow; (xiv) data privacy and
AI has the potential to play a vital role at various levels security; and (xiv) personalized treatments.
Table 3. Variability analysis of reviewed approaches.
Database, Vol. 2020, Article ID baaa010
It is based on 15 different features, which includes 1: intelligent interface development; 2: next-gen radiology and imaging tools development; 3: global expansion of medical resources; 4: automated ETL, linkage and data mining in HER; 5: risk prediction
and containment of antibiotics resistance; 6: pathology images analysis; 7: AI in machines and medical devices; 8: smart solutions and methods for cancer treatment; 9: EMR analysis for accurate risk predictors; 10: wearable devices for monitoring patients
health; 11: smartphone applications as diagnostic tools; 12: AI-based clinician decision-making; 13: search engine for healthcare data flow; 14: data privacy and security; 15: personalized treatments. Color coding, red represents ‘absence’, and blue represents
‘presence’ of respective feature discussion
Page 25 of 35
Although many AI and ML-based approaches have computer-aided design (CAD) (34); ML models to create
embraced the data gold rush in biomedicine, many knowledge base systems of phenotypes (40); ML algorithms
concerns remain, which include integration of biomedical to predict outbreak patterns and surveillance for new
data located in heterogeneous data sources; handling emergence (37); integrative approach to help decision-
extensively available irrelevant, error-prone and missing making processes between physicians and AI (89); ML
data; unnecessary follow-up diagnosis and treatment as a algorithms to enhance and optimize cancer treatment and
result of overloaded health information; loss of data privacy development of new drug treatments (45); ML algorithms
and reimagining of confidentiality; ethical strain of gaining to perform longitudinal population studies for analyzing
industrial profits with clinical decision-support systems the effects of treatments (56); ML-based hybrid model
without informing the users; possibilities of patient data classifier to enhance overall healthcare predictability (72);
exploitation, employment insecurity due to governance and ML algorithms to transform clinical research into
and management of big data; regulation of healthcare a much higher capacity and lower cost information
data analysis according to principles of bioethical law; processing care service (74).
evidence-based observational data analysis and screening, We also considered most adaptable traditional (statis-
slow data processing speed and difficulties in determining tical) AI and ML algorithms (SVM, deep learning, logis-
optimal parameters; difficulties in understanding structures tic regression, discriminant analysis, decision tree, random
of algorithms; possibility of too many overfitting attributes forest, linear regression, naïve Bayes, KNN, HMM and
and its optimal network structure; handling continuous genetic algorithm) in healthcare and justified their contri-
explanatory variables with more than two levels; difficulties butions with living examples for different clinical appli-
in understanding odds and probabilities (119); correct cations, e.g. deep neural network for diagnosing eye dis-
predictor variable selection, reference group setup, and eases (diabetic retinopathy) by analyzing retinal images;
managing relationships between input and output variables diagnosing cardiac anomalies by analyzing MRI images
(118,122); error-prone branching of missing values for of heart ventricles; detecting malignant lung nodules by
an attribute (108); lack of data availability on social analyzing radiographs; grading prostate cancer tumors by
determinants of health; difficulties in handling redun- analyzing ultrasound of biopsy cores; and producing glioma
dant attributes, distribution of statistically independent survival predictions by analyzing histological imaging and
attributes and management of class frequencies affecting genomic marker data. Some other non-traditional algo-
accuracy (108); increased computational complexity with rithms were used, e.g., ‘Watson for Oncology’ (241,242)
increasing number of attributes; computing probability of and ‘CURATE.AI’ (243,244), for analyzing EHR of can-
sequence observation; choosing accurate corresponding cer patients and recommend treatments and drug doses;
states and adjusting parameters (214); dealing with TREWScore (245,246,247,248) for analyzing EHR for pre-
correlations between residues due to underlying decency dicting septic shock risk and analyze CT brain scans for
assumption problem (215); and prediction modeling three-dimensional convolutional neural network neurolog-
with biases in confounding, causal inference, complexity- ical disorder classification; ‘PowerLook Density Assessment
based model selection, benchmark development and 3.4’ (249) for analyzing mammogram images for breast
pragmatic interoperability including reproducibility and density assessment; ‘OsteoDetect’ (84,250,251) for detect-
generalizability. ing distal radius fractures in the wrist; cloud-based deep
Potential solutions to these concerns are based on design neural network algorithm for diagnosing cardiac anoma-
and implementation of healthcare IT infrastructure (35,49); lies and segmentation of lung and liver tumors by ana-
EMR documentation and automation of repetitive tasks lyzing CT scans (56); and AI system for breast cancer
with ML algorithms (31); deep learning to mine (41), train prediction (252). Furthermore, major clinical oncological
(39) and learn complex relationships between features and developments with the implementation of different AI and
labels (32); ML algorithms to effectively link data between ML applications and algorithms include Chatbot; breast
different platforms (33); development of effective clinical MRI interpretation; breast lesion classification; colorec-
decision support systems (CDS) through the utilization of tal polyp classification; identification of colorectal can-
ML algorithms (38); translation of local languages into cer biomarkers in cell-free blood assays; gastric mucosal
EHRs and cloud-based data sharing (37); deep learning disease classification; detection of lung nodules in low-
for integrative EMR analysis from diverse sources (47); dose lung CT screening; prediction and evaluation of risk
computer vision algorithms to identify accurate indigestion and malignancy and classification of dysplastic nevi, spitz
of medications for patient (31); deep learning for speech nevi and basal and squamous cell carcinoma; detection of
recognition, image interpretation and language translation esophageal cancer; and identification of pancreatic cancer
(48); AI algorithms for diagnostic disease modeling and biomarkers (56).
The goal of this study is to highlight recent contributions of automatically linking and communicating with similar
and effectiveness of AI and ML in the development of com- systems.
putational systems towards better healthcare and precision Unmet clinical research and operational data analytics
medicine. We have reviewed and discussed multiple AI and needs development of intelligent and secure systems to
ML-based approaches (Tables 1, 2 and 3) and algorithms support practice transformation for implementing precision
with variable colors. Deep learning (253,254) has proven to medicine at a global level. Overarching goals include new
be one of the most trending algorithms today, but this does multi-functional platforms founded on the clinical, AI and
not undermine the importance of other machine learning scientific premise that integrates and analyzes heteroge-
algorithms. We believe that a right approach and algorithm neous clinical data received from multiple platforms. We
should be chosen for the development of the most effective need to implement and train AI classifiers at available
solutions to the targeted problems. In spite of various structured clinical dataset of over a million subjects to help
traditional and AI-based solutions, current limitations and in early detection and diagnosis of common, frequently
challenges by the healthcare community include uneven dis- occurring and rare diseases and predicting the performance
tribution of resources towards the future of digital health- of provided treatments. We need to further explore AI
care; business model unsustainability; disruption in the methodologies to design models segregating disorders, iden-
physician–patient relationships; hazards of dehumanization tify causative medical conditions and determine the best
of healthcare data and job insecurity threat; prevention drug therapies, especially when adjusted for age, race and
of early adopters for dragging down to the lowest com- gender. The goal should be to accelerate clinical care and
mon technology denominator; inconvenient adoption of discovery by satisfying research aims, improving quality
digital processes for healthcare interoperability; shortages and transition of care, obtaining actionable care gap-based
in development of customized IT infrastructure for data information about patients and developing communication
science; global unavailability of de-identified healthcare and coordination across hospitals, specialists, community-
data for research; understanding of how healthcare data of based providers, sub-acute care, nurses, quality inspectors,
different people relates to one another; need to harmonize management, researchers and analysts. We need to strive
big data with the definitions of the clinical phenotypes and to address possible challenges that continue to slow the
diagnosis; limited EHR systems for assimilating operational advancements of this breakthrough treatment approach.
and analytics interests; EHR offering limited patient par- Along with AI- and ML-based methodological devel-
ticipation; unavailability of high-quality open-source EHR opments, it is important to address the issues related to
systems inviting third-party extensions and assertions; and healthcare data privacy and security. In most of the cases,
inflexible EHR database schemas not geared for precision academic and applied research environments do not have
medicine. access to the healthcare data. The major reason is secure
To effectively meet the goals of healthcare data analytics, handling of protected health information (PHI) and lack of
while dealing with the aforementioned community, and trust by the healthcare institutions in providing access to
traditional and AI-based challenges, significant efforts are their medical records. We need to develop secure research-
required from experts in multidisciplinary sciences. To based HIPAA-compliant frameworks for efficient PHI stor-
facilitate and improve public sector clinical research and age, pre-processing, de-identification and integration to
practice, there is a critical need for academic frameworks serve a large community of users, support organizational
that can connect operational and analytical systems in policies and provide efficient access and connectivity. It
a way that experts from multiple domains can perform is mandatory to implement HIPAA rules for system user
measurement and descriptive analysis, even without strong data security, which includes application and data crit-
computational background. There is a need to develop icality, risk management and analysis, information sys-
standalone, user-friendly, standardized, open-source, and tem activity review, contingency plan, device and media
comprehensive solutions, which implement healthcare controls, disaster recovery plan, data backup plan, emer-
data analysis by connecting all kinds of patient data gency mode operations plan, device and media controls,
generated from any of the existing commercial EHR access controls, security incident procedures, vulnerability
systems at any level, which includes patient’s demographic assessment, penetration testing, physical security, business
information, personal life style, medical history, recent visits associate agreements, polices and procedures, evaluation,
to the practices, providers attended, diagnosis performed, audit, assignment of responsibility sanctions, workstation
lab tests conducted, longitudinal images, medications use and security. Furthermore, it is important to provide
and procedures, samples taken for wet and dry lab an external layer to the overall healthcare data analytics
experimentations for research and treatment of disease with systems for placing clinical data in a distributed centralized
no cure. Furthermore, the ideal system should be capable structure; administering security by encrypting data as well
as offering a multi-user-based graphical interface with con- References

trolled access; and supporting data backup with redundant 1. Ahmed,Z., Kim,J. and Liang,B.T. (2019) MAV-clic: framework
data management plans. towards management. Analysis and visualization of clinical big
data. J. Am. Med. Inf. Assoc. Open, 2, 23–28.
2. Makary,M.A. and Daniel,M. (2016) Medical error—the third
Conclusion leading cause of death in the US. BMJ., 353, i2139.
3. Ritchie, M. D., de Andrade, M. & Kuivaniemi, H. The foun-
Precision medicine is progressing but with many challenges
dation of precision medicine: integration of electronic health
lying ahead (255), which require addition of useful analytic
records with genomics through basic, clinical, and translational
tools, technologies, databases and approaches (4,6) to effi- research. Front. Genet., 6, 104 (2015).
ciently augment networking and interoperability of clinical, 4. Sboner,A. and Elemento,O. (2016) A primer on precision
laboratory and public health systems, as well as address eth- medicine informatics. Brief. Bioinform., 17, 145–153.
ical and social issues related to the privacy and protection of 5. Zeeshan,S., Ruoyun,X., Liang,B.T. and Ahmed,Z. (2019) 100
healthcare and omics data with effective balance. This will years of evolving gene-disease complexities and scientific debu-
also require more efficient management of massive amounts tants. Brief. Bioinform., bbz038.
6. Karczewski,K.J. and Snyder,M.P. (2018) Integrative omics for
of generated data, as well as earlier mined consensus and
health and disease. Nat. Rev. Genet., 108, 1111.
actionable data. Most efforts involved currently are manual 7. Marx,V. (2013) Biology: the big challenges of big data. Nature,
and time-consuming, whether it is extraction of healthcare 498, 255–260.
data from operational clinical systems, identification of 8. Combi,C. (2017) Editorial from the new editor-in-chief: arti-
common and rare functional variants, metabolite pene- ficial intelligence in medicine and the forthcoming challenges.
trance using listed features and abnormalities, examining Artif. Intell. Med., 76, 37–39.
relations between genomic variations and metabolite lev- 9. Ahmed,Z. and Liang,B.T. (2019) Systematically dealing prac-
tical issues associated to healthcare data analytics. In: Lecture
els, analyzing biochemical pathways in metabolites with
Notes in Networks and Systems 69 Springer Nature.
patterns of multimodal distributions for candidate genes
10. Beam,A.L. and Kohane,I.S. (2018) Big data and machine
and management and assimilation of healthcare, along with learning in health care. JAMA, 319, 1317–1318.
epidemiological and omics data generated at each step of 11. Raghupathi,W. and Raghupathi,V. (2014) Big data analytics in
entry, production and analysis. Cutting-edge, new AI and healthcare: promise and potential. Health Inform. Sci. Syst.,
ML-based big data platform development has the potential 2, 3.
to revolutionize the field of medicine and improve the 12. Alyass,A., Turcotte,M. and Meyre,D. (2015) From big data
quality and transition of healthcare by intelligently ana- analysis to personalized medicine for all: challenges and oppor-
tunities. BMC Med. Genom., 8, 33.
lyzing structured clinical data available in great count and
13. McShane,L.M., Cavenagh,M.M., Lively,T.G. et al. (2013)
volume, posing unprecedented challenges in data storage,
Criteria for the use of omics-based predictors in clinical trials.
processing, exchange and curation, and developing a better Nature, 502, 317–320.
understanding of biology. 14. Berger,B., Peng,J. and Singh,M. (2013) Computational solu-
tions for omics data. Nat. Rev. Genet., 14, 333–346.
15. Kim,M.O., Coiera,E. and Magrabi,F. (2017) Problems with
Supplementary data health information technology and their effects on care delivery
Supplementary data is available at Database online. and patient outcomes: a systematic review. J Am. Med. Inform.
Assoc., 7, 246–260.
16. Sligo,J., Gauld,R., Roberts,V. and Villa,L. (2017) A literature
Author contributions review for large-scale health information system project plan-
ZA proposed and lead the study. ZA, KM, and SZ conducted review ning, implementation and evaluation. Int. J. Med. Inf., 97,
research, and evaluated reported findings. XD guided the study. ZA
86–97.
drafted the manuscript, and all authors participated in writing and
17. Lu,Z. and Su,J. (2010) Clinical data management: current
review.
status, challenges, and future directions from industry perspec-
tives. Open Access J. Clin. Trials, 2, 93–105.
18. Haux,R., Knaup,P. and Leiner,F. (2007) On educating about
Acknowledgements
medical data management the other side of the electronic
We are grateful to the Ahmed lab, and appreciate all colleagues health record. Methods Inf. Med., 46, 74–79.
and institutions, who provided direct and indirect insight and 19. Rumsfeld,J.S., Joynt,K.E. and Maddox,T.M. (2016) Big data
expertise that greatly assisted this research. We would like to give
analytics to im-prove cardiovascular care: promise and chal-
special thanks to Dr. Christopher Bonin for providing editorial
lenges. Nat. Rev. Cardiol., 13, 350–359.
support.
20. Van Panhuis,W.G., Paul,P., Emerson,C. et al. (2014) A system-
Conflict of interest. The authors have declared that no conflict of atic review of barriers to data sharing in public health. BMC
interests exists. Public Health, 14, 1144.
21. Fegan,G.W. and Lang,T.A. (2008) Could an open-source clin- 39. Jiang,F., Jiang,Y., Zhi,H. et al. (2017) Artificial intelligence in
ical trial data-management system be what we have all been healthcare: past, present and future. Stroke Vasc. Neurol., 2,
looking for? PLoS Med., 5, e6. 230–243.
22. Wang,X., Williams,C., Liu,Z.H. and Croghan,J. (2017) Big data 40. Emmert-Streib,F. and Dehmer,M. (2019) A machine learning
management challenges in health research—a literature review. perspective on personalized medicine: An automized compre-
Brief Bioinform, bbx086. hensive knowledge base with ontology for pattern recognition.
23. Duffy,D.J. (2016) Problems, challenges and promises: per- Mach. Learn Knowl Extr., 1, 149–156.
spectives on precision medicine. Brief. Bioinform., 17, 41. Gruson,D., Helleputte,T., Rousseau,P. and Gruson,D. (2019)
494–504. Data science, artificial intelligence, and machine learning:
24. Xu,R., Li,L. and Wang,Q. (2014) dRiskKB: a large-scale opportunities for laboratory medicine and the value of positive
disease-disease risk relationship knowledge base constructed regulation. Clin. Biochem., 69, 1–7.
from biomedical text. BMC Bioinform., 15, 105. 42. Meskó,B., Hetényi,G. and Győrffy,Z. (2018) Will artificial
25. Xu,R. and Wang,Q. (2015) PhenoPredict: a disease phenome- intelligence solve the human resource crisis in healthcare? BMC
wide drug repositioning approach towards schizophrenia drug Health Serv. Res., 18, 545.
discovery. J. Biomed. Inform., 56, 348–355. 43. Bostrom,N. (2014) Superintelligence: Paths, Dangers, Strate-
26. Brown,A.S. and Patel,C.J.M.S.H.D.D. (2017) Literature-based gies. Oxford University Press.
drug-drug similarity for drug repositioning. JAMIA, 24, 44. Stein,J.D., Rahman,M., Andrews,C. et al. (2019) Evaluation
614–618. of an algorithm for identifying ocular conditions in electronic
27. Luo,Y., Szolovits,P., Dighe,A.S. and Baron,J.M. (2017) 3D- health record data. JAMA Ophthalmol., 137, 491–497.
MICE: integration of cross-sectional and longitudinal impu- 45. Kantarjian,H. and Yu,P.P. (2015) Artificial intelligence, big
tation for multi-analyte longitudinal clinical data. J. Am. Med. data, and cancer. JAMA Oncol., 1, 573–574.
Inform. Assoc, ocx133. 46. Hersh,W.R., Weiner,M.G., Embi,P.J. et al. (2013) Caveats for
28. Walker,J.G., Bickerstaffe,A., Hewabandu,N. et al. (2017) The the use of operational electronic health record data in compar-
CRISP colorectal cancer risk prediction tool: an exploratory ative effectiveness research. Med. Care, 51, S30–S37.
study using simulated consultations in Australian primary care. 47. Naylor,C.D. (2018) On the prospects for a (deep) learning
BMC Med. Inform. Decis. Mak., 17, 13. health care system. JAMA, 320, 1099–1100.
29. Liu,L., Liu,L., Fu,X. et al. (2017) A cloud-based framework 48. Hinton,G. (2018) Deep learning-a technology with the poten-
for large-scale traditional Chinese medical record retrieval. tial to transform health care. JAMA, 320, 1101–1102.
J. Biomed. Inform., 77, 21–33. 49. Topol,E.J. (2019) High-performance medicine: the conver-
30. Ahmed,Z., Zeeshan,S. and Dandekar,T. (2016) Mining biomed- gence of human and artificial intelligence. Nat. Med., 25,
ical images towards valuable information retrieval in biomed- 44–56.
ical and life sciences, Database baw118. 50. Acs,B. and Rimm,D.L. (2018) Not just digital pathology, intel-
31. Fogel,A.L. and Kvedar,J.C. (2018) Artificial intelligence pow- ligent digital pathology. JAMA Oncol., 4, 403–404.
ers digital medicine. NPJ Digit. Med., 1, 5. 51. Rose,S. (2018) Machine learning for prediction in electronic
32. Rajkomar,A., Dean,J. and Kohane,I. (2019) Machine learning health data. JAMA Netw Open., 1, e181404.
in medicine. N. Engl. J. Med., 380, 1347–1358. 52. Cabitza,F., Rasoini,R. and Gensini,G.F. (2017) Unintended
33. Nayak, L., Ray, I. & De, R. K. Precision medicine with elec- consequences of machine learning in medicine. JAMA, 318,
tronic medical records: from the patients and for the patients. 517–518.
Ann. Transl. Med. 4(Suppl 1), S61 (2016). 53. Weber,G.M., Mandl,K.D. and Kohane,I.S. (2014) Finding the
34. Jones,L.D., Golan,D., Hanna,S.A. and Ramachandran,M. missing link for big biomedical data. JAMA, 311, 2479–2480.
(2018) Artificial intelligence, machine learning and the evolu- 54. Wong,D. and Yip,S. (2018) Machine learning classifies cancer.
tion of healthcare: a bright future or cause for concern? Bone Nature, 555, 446–447.
Joint Res., 7, 223–225. 55. Saheb,T. and Saheb,M. (2019) Analyzing and visualizing
35. Redekop,W.K., Bakker,L.J. and Aarts,J. (2018) Healthcare knowledge structures of health informatics from 1974 to 2018:
problems cannot be solved using health technologies alone: a bibliometric and social network analysis. Healthc. Inform.
the example of precision medicine. Health Policy Technol., 7, Res., 25, 61–72.
3–4. 56. Ngiam,K.Y. and Khor,I.W. (2019) Big data and machine learn-
36. Mesko,B. (2017) The role of artificial intelligence in precision ing algorithms for health-care delivery. Lancet Oncol., 20,
medicine. Expert Rev. Precis. Med. Drug Dev., 2, 239–241. e262–e273.
37. Wahl,B., Cossy-Gantner,A., Germann,S. and Schwalbe,N.R. 57. Abidi,S.S.R. and Abidi,S.R. (2019) Intelligent health data ana-
(2018) Artificial intelligence (AI) and global health: how can lytics: a convergence of artificial intelligence and big data.
AI contribute to health in resource-poor settings? BMJ Glob. Healthc. Manage. Forum., 32, 178–182.
Health, 3, e000798. 58. Char,D.S., Shah,N.H. and Magnus,D. (2018) Implementing
38. Sitapati, A., Kim, H., Berkovich, B., Marmor, R., Singh, S., machine learning in health care - addressing ethical challenges.
El-Kareh, R., Clay, B. & Ohno-Machado, L. Integrated preci- N. Engl. J. Med., 378, 981–983.
sion medicine: the role of electronic health records in delivering 59. He,J., Baxter,S.L., Xu,J. et al. (2019) The practical implemen-
personalized treatment. Wiley Interdiscip. Rev. Syst. Biol. Med. tation of artificial intelligence technologies in medicine. Nat
9, 3 (2017). Med., 25, 30–36.
60. Crown,W.H. (2019) Real-world evidence, causal inference, and 80. Seetharam,K., Shrestha,S. and Sengupta,P.P. (2019) Artificial
machine learning. Value Health, 22, 587–592. intelligence in cardiovascular medicine. Curr. Treat Options
61. Raghupathi,W. and Raghupathi,V. (2014) Big data analytics in Cardio. Med., 21, 25.
healthcare: promise and potential. Health Inform. Sci. Syst., 81. Karkanis,S., Magoulas,G. and Theofanous,N. (2000) Image
2, 3. recognition and neuronal networks: intelligent systems for the
62. Azencott,C.A. (2018) Machine learning and genomics: preci- improvement of imaging information. Min. Invas. Ther. Allied
sion medicine versus patient privacy. Philos. Trans. A Math. Technol., 9, 225–230.
Phys. Eng. Sci., 13, 376. 82. Liang,H. et al. (2019) Evaluation and accurate diagnoses of
63. Kourou,K., Exarchos,T.P., Exarchos,K.P. et al. (2014) Machine pediatric diseases using artificial intelligence. Nat. Med., 25,
learning applications in cancer prognosis and prediction. Com- 433–438.
put. Struct. Biotechnol. J., 13, 8–17. 83. Herron,D.M., Marohn,M. and SAGES-MIRA Robotic Surgery
64. Gulshan,V. et al. (2016) Development and validation of a deep Consensus Group (2008) A consensus document on robotic
learning algorithm for detection of diabetic retinopathy in surgery. Surg. Endosc., 22, 313–315.
retinal fundus photographs. JAMA, 316, 2402–2410. 84. Craft,J.A. (2018) Artificial intelligence and the softer side of
65. Labovitz,D.L., Shafner,L., Reyes Gil,M. et al. (2017) Using medicine. Missouri Med., 115, 406–409.
artificial intelligence to reduce the risk of nonadherence in 85. Galimova,R.M., Buzaev,I.V., Ramilevich,K.A. et al. (2019)
patients on anticoagulation therapy. Stroke, 48, 1416–1419. Artificial intelligence-developments in medicine in the last two
66. Li,L. et al. (2015) Identification of type 2 diabetes subgroups years. Chronic Dis. Transl. Med., 5, 64–68.
through topological analysis of patient similarity. Sci. Transl. 86. Mesko,B. (2017) The role of artificial intelligence in pre-
Med., 7, 311ra174. cision medicine. Expert Rev. Precis. Med. Drug Dev., 2,
67. Sheikhtaheri,A., Sadoughi,F. and Hashemi Dehaghi,Z. (2014) 239–241.
Developing and using expert systems and neural networks in 87. van Hartskamp,M., Consoli,S., Verhaegh,W. et al. (2019)
medicine: a review on benefits and challenges. J. Med. Syst., 38, Artificial intelligence in clinical health care applications. Inter-
110. act. J. Med. Res., e12100, 8.
68. Buch,V.H., Ahmed,I. and Maruthappu,M. (2018) Artificial 88. Schork,N. and Artificial,J. (2019) Intelligence and personalized
intelligence in medicine: current trends and future possibilities. medicine. Cancer Treat. Res., 178, 265–283.
Br. J. Gen. Pract., 68, 143–144. 89. Mesko,B. (2019) Artificial intelligence is the stethoscope of
69. Bali,J., Garg,R. and Bali,R.T. (2019) Artificial intelligence the 21st century. The Medical Futurist. https://medicalfuturi
(AI) in healthcare and biomedical research: why a strong st.com/ibm-watson-is-the-stethoscope-of-the-21st-century.
computational/AI bioethics framework is required? Indian J. Updated July 18,2017. Accessed 23 August 2019.
Ophthalmol., 67, 3–6. 90. Challen,R., Denny,J., Pitt,M. et al. (2019) Artificial intelligence,
70. Mintz,Y. and Brodie,R. (2019) Introduction to artificial intelli- bias and clinical safety. BMJ Qual. Saf., 28, 231–237.
gence in medicine. Minim. Invasive. Ther. Allied Technol., 28, 91. Kaura,P., Sharmab,M. and Mittalc,M. (2018) Big data and
73–81. machine learning based secure healthcare framework. Procedia
71. Prevedello,L.M. et al. (2017) Automated critical test findings Comput. Sci., 132, 1049–1059.
identification and online notification system using artifical 92. Kaushal,R., Shojania,K.G. and Bates,D.W. (2003) Effects of
intelligence. Radiol., 285, 923–931. computerized physician order entry and clinical decision sup-
72. Neesha Jothia,N.A. (2015) Data Mining in Healthcare - a port systems on medication safety: a systematic review. Arch.
review. Procedia Comput. Sci., 72, 306–313. Intern. Med., 163, 1409–1416.
73. Bejnordi,B.E. et al. (2017) Diagnostic assessment of deep 93. Bouch,D.C. and Thompson,J.P. (2008) Severity scoring systems
learning algorithms for detection of lymph node metastasis in in the critically ill. Cont. Edu. Anaesth. Crit. Care Pain., 8,
women with breast cancer. JAMA, 318, 2199–2210. 181–185.
74. Beam,A.L. and I,K. (2016) Translating Artificial intelligence 94. Gianfrancesco,M.A., Tamang,S., Yazdany,J. and Schmajuk,G.
into clinical care. JAMA, 316, 2368–2369. (2018) Potential biases in machine learning algorithms
75. Somashekhar, S. P. et al. Watson for oncology and breast using electronic health record data. JAMA Int. Med., 178,
cancer treatment recommendations: agreement with an expert 1544–1547.
multidisciplinary tumor board. Ann. Oncol. 29, 1–6 (2018). 95. Sidey-Gibbons,J. and Sidey-Gibbons,C.J. (2019) Machine
76. Esteva,A. et al. (2017) Dermatologist-level classification of skin learning in medicine: a practical introduction. BMC Med. Res.
cancer with deep neural networks. Nature, 542, 115–118. Methodol., 19, 64.
77. Yu,K.H. et al. (2016) Predicting non-small cell lung cancer 96. Panch, T., Szolovits, P. & Atun, R. Artificial intelligence,
prognosis by fully automated microscopic pathology image machine learning and health systems. J. Global Health 8,
features. Nat. Commun., 7, 12474. 020303 (2018).
78. Slomka,P.J. et al. (2017) Cardiac imaging: working towards 97. Hippisley-Cox,J., Coupland,C., Vinogradova,Y. et al. (2008)
fully-automated machine analysis & interpretation. Expert Predicting cardiovascular risk in England and Wales: prospec-
Rev. Med. Dev., 14, 197–212. tive derivation and validation of QRISK2. BMJ Clin. Res. Ed.,
79. Choi,E. et al. (2017) Using recurrent neural network models 336, 1475–1482.
for early detection of heart failure. J. Am. Med. Inform. Assoc., 98. Rajkomar,A., Yim,J.W., Grumbach,K. and Parekh,A. (2016)
24, 351–370. Weighting primary care patient panel size: a novel electronic
health record-derived measure using machine learning. JMIR 120. McHugh,M.L. (2009) The odds ratio: calculation, usage, and
Med. Inform., 4, e29. interpretation. Biochem. Med., 19, 120–126.
99. Sullivan, T. Next up for EHRs: Vendors adding artificial 121. Kalil,A.C., Mattei,J., Florescu,D.F. et al. (2010) Recommenda-
intelligence into the workflow. Healthcare IT News. https:// tions for the assessment and reporting of multivariable logistic
www.healthcareitnews.com/news/next-ehrs-vendors-adding-a regression in transplantation literature. Am. J. Transplant., 10,
rtificial-intelligence-workflow. Updated March 13, 2018. 1686–1694.
Accessed 23 August 2019. (2018). 122. Tolles,J. and Meurer,W.J. (2016) Logistic regression: relating
100. Vapnik,V. (1963) Pattern recognition using generalized portrait patient characteristics to outcomes. JAMA, 316, 533–534.
method. Autom. Remote Control., 24, 774–780. 123. Xu,W., Zhao,Y., Nian,S. et al. (2018) Differential analy-
101. Noble, W. S. What is a support vector machine? Nat. Biotech. sis of disease risk assessment using binary logistic regres-
24, 1564–1567 (2006). sion with different analysis strategies. J. Int. Med. Res., 46,
102. Kaura,P., Sharmab,M. and Mittalc,M. (2018) Big data and 3656–3664.
machine learning based secure healthcare framework. Procedia 124. Mamiya,H. et al. (2015) Towards probabilistic decision sup-
Comput. Sci., 132, 1049–1059. port in public health practice: predicting recent transmission
103. Hearst,M.A., Dumais,S.T., Osuna,E. et al. (1998) Support of tuberculosis from patient attributes. J Biomed. Inform., 53,
vector machines. IEEE Intell. Syst. Appl., 13, 18–28. 237–242.
104. Boser,B.E., Guyon,I.M. and Vapnik,V.N. (1992) A training 125. García-Laencina,P.J. et al. (2015) Missing data imputation
algorithm for optimal margin classifiers. In: 5th Annual ACM on the 5-year survival prediction of breast cancer patients
Workshop on COLT. ACM Press, pp. 144–152. with unknown discrete values. Comput. Biol. Med., 59,
105. Shawe-Taylor,J. and Cristianini,N. (2004) Kernel Methods for 125–133.
Pattern Analysis. Cambridge University Press. 126. Nick,T.G. and Logistic Regression,C.K.M. (2007) Topics in
106. Huang, S., Cai, N., Pacheco, P. P., Narrandes, S., Wang, Y. & Xu, biostatistics. Methods Mol. Biol., 404.
W. Applications of support vector machine (SVM) learning in 127. Yoo,H.H. et al. (2003) Logistic regression analysis of potential
cancer genomics. Cancer Genom. Proteom. 15, 41–51 (2018). prognostic factors for pulmonary thromboembolism. Chest.,
107. Cho,G., Yim,J., Choi,Y. et al. (2019) Review of machine 123, 813–821.
learning algorithms for diagnosing mental illness. Psychiatry 128. Zhang,W.T. and Kuang,C.W. (2011) SPSS statistical analysis-
Investig., 16, 262–269. based tutorial, 2nd edn. Higher Education Press, Beijing.
108. Cruz,J.A. and Wishart,D.S. (2007) Applications of machine 129. Hosmer,D.W., Lemeshow,S. and Sturdivant,R.X. (2013)
learning in cancer prediction and prognosis. Cancer Inform., Applied Logistic Regression, 3rd edn. Hoboken Wiley.
2, 59–77. 130. Solberg,H.E. (1978) Discriminant analysis. CRC Crit. Rev.
109. Vapnik,V. (1982) Estimation of Dependences Based on Empir- Clin. Lab. Sci., 9, 209–242.
ical Data. Springer Verlag. 131. Davis,J.C. (1986) Statistics and Data Analysis in Geology.
110. Chang,H.Y., Jung,C.K., Woo,J.I. et al. (2019) Artificial intelli- Wiley, New York.
gence in pathology. J. Pathol. Transl. Med., 53, 1–12. 132. Afifi,A.A. and Clark,V. (1996) Computer-aided multivariate
111. LeCun,Y., Bengio,Y. and Hinton,G. (2015) Deep learning. analysis, 3rd edn. Chapman & Hall, London.
Nature, 521, 436–444. 133. Brown,C.A. (1998) Applied Multivariate Statistics in Geohy-
112. Kevin Zhou,S., Greenspan,H. and Shen,D. (2017) Deep Learn- drology and Related sciences. Springer, New York.
ing for Medical Image Analysis. Academic Press. 134. Panagopoulos,G.P. et al. (2016) The contribution of cluster and
113. Hosny,A., Parmar,C., Quackenbush,J. et al. (2018) Artificial discriminant analysis to the classification of complex aquifer
intelligence in radiology. Nat. Rev. Cancer, 18, 500–510. systems. Environ. Monit. Assess., 188, 591.
114. Langlotz,C.P. et al. (2019) A roadmap for foundational 135. Shayan,Z., Mohammad Gholi Mezerji,N., Shayan,L. and
research on artificial intelligence in medical imaging: from the Naseri,P. (2015) Prediction of depression in cancer patients
2018 NIH/RSNA/ACR/the Academy Workshop. Radiology, with different classification criteria, linear discriminant anal-
291, 781–791. ysis versus logistic regression. Glob. J. Health Sci., 8, 41–46.
115. Haenssle,H.A. et al. (2018) Man against machine: diagnostic 136. Winkler-Schwartz,A. et al. (2019) Machine learning identifica-
performance of a deep learning convolutional neural network tion of surgical and operative factors associated with surgical
for dermoscopic melanoma recognition in comparison to 58 expertise in virtual reality simulation. JAMA Netw. Open.,
dermatologists. Ann. Oncol., 29, 1836–1842. 20192, e198363.
116. Olsen,T.G. et al. (2018) Diagnostic performance of deep learn- 137. Murata,C. et al. (2019) Discriminant analysis to predict the
ing algorithms applied to three common diagnoses in der- clinical diagnosis of primary immunodeficiencies: a prelimi-
matopathology. J. Pathol. Inform., 9, 32. nary report. Rev. Alerg. Mex., 62, 125–133.
117. Rajkomar,A. et al. (2018) Scalable and accurate deep learning 138. Dansky,K.H. and Brannon,D. (1996) Discriminant analysis:
with electronic health records. Nat. Digital Med., 1, 18. a technique for adding value to patient satisfaction surveys.
118. Ranganathan,P., Pramesh,C.S. and Aggarwal,R. (2017) Com- Hosp. Health Services Admin., 41, 503–513.
mon pitfalls in statistical analysis: Logistic regression. Perspect. 139. Mandelkow, H., de Zwart, J. A. & Duyn, J. H. Linear dis-
Clin. Res., 8, 148–151. criminant analysis achieves high classification accuracy for
119. Sperandei,S. (2014) Understanding logistic regression analysis. the BOLD fMRI response to naturalistic movie stimuli. Front.
Biochem. Med., 24, 12–18. Hum. Neurosci. 10, 128 (2016).
140. Jin,J. and An,J. (2011) Robust discriminant analysis and its 159. Kotti,M., Duffell,L.D., Faisal,A.A. and McGregor,A.H. (2017)
application to identify protein coding regions of rice genes. Detecting knee osteoarthritis and its discriminating parameters
Math Biosci., 232, 96–100. using random forests. Med. Eng. Phys., 43, 19–29.
141. Armañanzas,R., Bielza,C., Chaudhuri,K.R. et al. (2013) Unveil- 160. Pachauri,I. and Sharma,S. (2015) Anomaly detection in med-
ing relevant non-motor Parkinson’s disease severity symptoms ical wireless sensor networks using machine learning algo-
using a machine learning approach. Artif. Intell. Med., 58, rithms. Procedia Comput. Sci., 70, 325–333.
195–202. 161. O’Bryant,S.E. et al. (2010) A serum protein–based algorithm
142. C.-H. Jen, C.-C. Wang, B. C. Jiang, Y.-H. Chu & M.-S. Chen. for the detection of Alzheimer disease. Arch. Neurol., 67,
Application of classification techniques on development an 1077–1081.
early-warning system for chronic illnesses. Expert Syst. Appl. 162. Toubiana,D. et al. (2019) Combined network analysis and
39, 10, 8852–8858 (2012). machine learning allows the prediction of metabolic pathways
143. Johnson,K.R., Mascall,G.C. and Howarth,A.T. (1984) Differ- from tomato metabolomics data. Commun. Biol., 2, 214.
ential laboratory diagnosis of hypercalcemia. Crit. Rev. Clin. 163. Roysden,N. and Wright,A. (2015) Predicting health care
Lab. Sci., 21, 51–97. utilization after behavioral health referral using natural
144. Lee,E.K., Yuan,F., Hirsh,D.A. et al. (2012) A clinical deci- language processing and machine learning. AMIA, 2015,
sion tool for predicting patient care characteristics: patients 2063–2072.
returning within 72 hours in the emergency department. 164. Morid,M.A., Kawamoto,K., Ault,T. et al. (2018) Supervised
AMIA ... Annual Symposium proceedings. AMIA Symposium, learning methods for predicting healthcare costs: systematic lit-
2012, 495–504. erature review and empirical evaluation. AMIA ... Annual sym-
145. Deo,R.C. (2015) Machine learning in medicine. Circulation, posium proceedings. AMIA Symposium, 2017, 1312–1321.
132, 1920–1930. 165. Lee,J. (2017) Patient-specific predictive modeling using ran-
146. Quinlan,J.R. (1986) Induction of decision trees. Machine dom forests: An observational study for the critically ill. JMIR
Learning, 1, 81–106. Med. Inform., 5, e3.
147. Bae,J.M.T. (2014) Clinical decision analysis using decision tree. 166. Sarica,A., Cerasa,A. and Quattrone,A. (2017) Random For-
Epidemiol. Health, 36, e2014025. est algorithm for the classification of neuroimaging data in
148. Montbriand,M.J. (1995) Decision tree model describing alter- Alzheimer’s disease: a systematic review. Front. Aging Neu-
nate health care choices made by oncology patients. Cancer rosci., 9, 329.
Nurs., 18, 104–117. 167. Seligman,B., Tuljapurkar,S. and Rehkopf,D. (2017) Machine
149. Bonner,G. (2001) Decision making for health care profession- learning approaches to the social determinants of health in the
als: use of decision trees within the community mental health health and retirement study. SSM Popul. Health, 4, 95–99.
setting. J. Adv. Nursing, 35, 349–356. 168. Khalilia,M., Chakraborty,S. and Popescu,M. (2011) Predicting
150. Tsai,W.M., Zhang,H., Buta,E. et al. (2016) A modified classi- disease risks from highly imbalanced data using random forest.
fication tree method for personalized medicine decisions. Stat. BMC Med. Inform. Dec. Mak., 11, 51.
Interface, 9, 239–253. 169. DuBrava,S. et al. (2017) Using random Forest models to iden-
151. Tayefi,M. et al. (2017) The application of a decision tree to tify correlates of a diabetic peripheral neuropathy diagnosis
establish the parameters associated with hypertension. Com- from electronic health record data. Pain Med., 18, 107–115.
put. Methods Programs Biomed., 139, 83–91. 170. Rahimian,F. et al. (2018) Predicting the risk of emergency
152. Moon,M. and Lee,S.K. (2017) Applying of decision tree anal- admission with machine learning: development and valida-
ysis to risk factors associated with pressure ulcers in long-term tion using linked electronic health records. PLoS Med., 15,
care facilities. Healthc. Inform. Res., 23, 43–52. e1002695.
153. Chern,C.C., Chen,Y.J. and Siao,B. (2019) Decision tree-based 171. McWilliams,C.J. et al. (2019) Towards a decision support
classifier in providing telehealth service. BMC Med. Inform. tool for intensive care discharge: machine learning algorithm
Decis. Mak., 19, 104. development using electronic healthcare data from MIMIC-III
154. Valdes,G., Luna,J.M., Eaton,E. et al. (2016) MediBoost: a and Bristol, UK. BMJ Open, 9, e025925.
patient stratification tool for interpretable decision making in 172. Wager,S. and Athey,S. (2018) Estimation and inference of
the era of precision medicine. Sci. Rep., 6, 37854. heterogeneous treatment effects using random forests. J. Am.
155. Gheondea-Eladi,A. (2019) Patient decision aids: a content Stat. Assoc., 113, 1228–1242.
analysis based on a decision tree structure. BMC Med. Inform. 173. Yulita,I.N., Fanany,M.I. and Arymurthy,A.M. (2018) Fast con-
Dec. Mak., 19, 137. volutional method for automatic sleep stage classification.
156. Kasbekar,P.U., Goel,P. and Jadhav,S.P. (2017) A decision tree Healthc. Inform. Res., 24, 170–178.
analysis of diabetic foot amputation risk in Indian patients. 174. Byeon,H. (2019) Developing a random forest classifier for pre-
Front. Endocrinol., 8, 25. dicting the depression and managing the health of caregivers
157. Ainscough,K.M., Lindsay,K.L., O’Sullivan,E.J. et al. (2017) supporting patients with Alzheimer’s disease. Technol. Health
Behaviour change in overweight and obese pregnancy: a deci- Care, 1–14.
sion tree to support the development of antenatal lifestyle 175. Schneider,A., Hommel,G. and Blettner,M. (2010) Linear regres-
interventions. Public Health Nutri., 20, 2642–2648. sion analysis: part 14 of a series on evaluation of scien-
158. Ho,T.K. (1998) The random subspace method for constructing tific publications. Deutsches Arzteblatt international, 107,
decision forests. IEEE T. Pattern Anal., 20, 832–844. 776–782.
176. Bebbington,E. and Furniss,D. (2015) Linear regression analysis 193. Srinivas,K. et al. (2010) Applications of data mining techniques
of hospital episode statistics predicts a large increase in demand in healthcare and prediction of heart attacks. Int. J. Comput.
for elective hand surgery in England. JPRAS., 68, 243–251. Sci. Eng., 2, 250–255.
177. Backenroth,D., Chase,H.S., Wei,Y. and Friedman,C. (2017) 194. Altman,N.S. (1992) An introduction to kernel and nearest-
Monitoring prescribing patterns using regression and elec- neighbor nonparametric regression. Am. Stat., 46, 175–185.
tronic health records. BMC Medical Inform. Dec. Mak., 195. Zhang,Z. (2016) Introduction to machine learning: k-nearest
17, 175. neighbors. Ann. Translat. Med., 4, 218.
178. Morton,V. and Torgerson,D.J. (2003) Effect of regression to the 196. Park,J. and Lee,D.H. (2018) Privacy preserving k-nearest neigh-
mean on decision making in health care. BMJ Clin. Res. Ed., bor for medical diagnosis in e-health cloud. J. Healthc. Eng.,
326, 1083–1084. 2018, 4073103.
179. Madadizadeh,F., Ezati Asar,M. and Bahrampour,A. (2016) 197. Hu,L.Y., Huang,M.W., Ke,S.W. and Tsai,C.F. (2016) The dis-
Quantile regression and its key role in promoting medical tance function effect on k-nearest neighbor classification for
research. Iran. J. Public Health, 45, 116–117. medical datasets. SpringerPlus, 5, 1304.
180. Malehi,A.S., Pourmotahari,F. and Angali,K.A. (2015) Statisti- 198. Li,C., Zhang,S., Zhang,H. et al. (2012) Using the K-nearest
cal models for the analysis of skewed healthcare cost data: a neighbor algorithm for the classification of lymph node metas-
simulation study. Health Econ. Rev., 5, 11. tasis in gastric cancer. Comput. Math. Methods Med., 876545,
181. Madigan,E.A., Curet,O.L. and Zrinyi,M. (2008) Workforce 2012.
analysis using data mining and linear regression to understand 199. Sarkar,M. and Leong,T.Y. (2000) Application of K-nearest
HIV/AIDS prevalence patterns. Hum. Res. Health, 6, 2. neighbors algorithm on breast cancer diagnosis problem.
182. Langley,P., Iba,W. and Thompson,K. (1992) An analysis of AMIA Symposium, 2000, 759–763.
Bayesian classifiers. Proceedings of the Tenth National Con- 200. Vitola,J., Pozo,F., Tibaduiza,D.A. and Anaya,M.A. (2017) Sen-
ference on. Artif. Intell., 223–228. sor data fusion system based on k-nearest neighbor pattern
183. Rish,I. (2001) An empirical study of the naive Bayes classifier. classification for structural health monitoring applications.
IJCAI 2001 Work Empir Methods Artif Intell., 3, 41–46. Sensors Basel. Switzerland, 17, 417.
184. Langarizadeh,M. and Moghbeli,F. (2016) Applying naive 201. Zhao,D. and Weng,C. (2011) Combining PubMed knowl-
Bayesian networks to disease prediction: a systematic review. edge and EHR data to develop a weighted bayesian net-
AIM: journal of the Society for Medical Informatics of Bosnia work for pancreatic cancer prediction. J. Biomed. Inform., 44,
& Herzegovina, 24, 364–369. 859–868.
185. Wolfson,J. et al. (2015) A naive Bayes machine learning 202. Baum,L.E. and Petrie,T. (1966) Statistical inference for prob-
approach to risk prediction using censored, time-to-event data. abilistic functions of finite state Markov chains. Ann. Math.
Stat. Med., 34, 2941–2957. Stat., 37, 1554–1563.
186. Ehsani-Moghaddam,B., Queenan,J.A., MacKenzie,J. and 203. Baum,L.E. and Egon,J.A. (1967) An inequality with applica-
Birtwhistle,R.V. (2018) Mucopolysaccharidosis type II tions to statistical estimation for probabilistic functions of
detection by Naïve Bayes classifier: An example of patient a Markov process and to a model for ecology. Bull. Amer.
classification for a rare disease using electronic medical Meteorol. Soc., 73, 360–363.
records from the Canadian Primary Care Sentinel Surveillance 204. Rabiner,L.R. (1989) A tutorial on hidden Markov models and
Network. PloS One, 13, e0209018. selected applications in speech recognition. Proc. IEEE., 77,
187. Wei,W., Visweswaran,S. and Cooper,G.F. (2011) The applica- 257–286.
tion of naive Bayes model averaging to predict Alzheimer’s 205. Eddy,S.R. (2004) What is a hidden Markov model? Nat.
disease from genome-wide data. JAMIA, 18, 370–375. Biotech., 22, 1315–1316.
188. Doing-Harris,K., Mowery,D.L., Daniels,C. et al. (2017) Under- 206. Wall,M.M. and Li,R. (2009) Multiple indicator hidden Markov
standing patient satisfaction with received healthcare services: model with an application to medical utilization data. Stat.
a natural language processing approach. AMIA Symposium, Med., 28, 293–310.
2016, 524–533. 207. Kawamoto,R. et al. (2013) Hidden markov model for ana-
189. Grover,D., Bauhoff,S. and Friedman,J. (2019) Using supervised lyzing time-series health checkup data. Stud. Health Technol.
learning to select audit targets in performance-based financing Inform., 192, 491–495.
in health: An example from Zambia. PloS One, 14, e0211262. 208. Sotoodeh,M. and Ho,J.C. (2019) Improving length of stay
190. Wagholikar,K.B., Vijayraghavan,S. and Deshpande,A.W. prediction using a hidden Markov model. AMIA Jt. Summits
(2009) Fuzzy naive Bayesian model for medical diagnostic Transl. Sci., 2019, 425–434.
decision support. Conf. Proc. IEEE Eng. Med. Biol. Soc., 209. Sampathkumar,H., Chen,X.W. and Luo,B. (2014) Mining
3409–3412 2009. adverse drug reactions from online healthcare forums using
191. Al-Aidaroos,K.M., Bakar,A.A. and Othman,Z. (2012) Medical hidden Markov model. BMC Med. Inform. Dec. Mak., 14, 91.
data classification with naive Bayes approach. Inform. Technol. 210. Huang,Z., Dong,W., Wang,F. and Duan,H. (2015) Medical
J., 11, 1166–1174. inpatient journey modeling and clustering: a Bayesian hidden
192. Sebastiani,P., Solovieff,N. and Sun,J.X. (2012) Naïve Bayesian Markov model based approach. AMIA Symposium, 2015,
classifier and genetic risk score for genetic risk prediction of 649–658.
a categorical trait: not so different after all! Front. Genet., 211. Esmaili,N., Piccardi,M., Kruger,B. and Girosi,F. (2019) Correc-
3, 26. tion: analysis of healthcare service utilization after transport-
related injuries by a mixture of hidden Markov models. PloS 229. Latkowski,T. and Osowski,S. (2015) Computerized system
One, 14, e0214973. for recognition of autism on the basis of gene expression
212. Huang,Q., Cohen,D., Komarzynski,S. et al. (2018) Hidden microarray data. Comput. Biol. Med., 56, 82–88.
Markov models for monitoring circadian rhythmicity in tele- 230. Jefferson,M.F., Pendleton,N., Lucas,S.B. and Horan,M.A.
metric activity data. J. R. Soc. Interface, 15, 20170885. (1997) Comparison of a genetic algorithm neural network
213. Marchuk,Y. et al. (2018) Predicting patient-ventilator asyn- with logistic regression for predicting outcome after surgery
chronies with hidden Markov models. Sci. Rep., 8, 17614. for patients with nonsmall cell lung carcinoma. Cancer, 79,
214. Naithani,G. et al. (2018) Automatic segmentation of infant 1338–1342.
cry signals using hidden Markov models. EURASIP J. Audio 231. Güler,I., Polat,H. and Ergün,U. (2005) Combining neural net-
Speech Music Process, 1, 1. work and genetic algorithm for prediction of lung sounds.
215. Reeves,C. (2003) Genetic algorithms. Handbook of meta- J. Med. Syst., 29, 217–231.
heuristics: Springer, 3, 55–82. 232. Castiglione,F., Poccia,F., D’Offizi,G. and Bernaschi,M. (2013)
216. Hasancebi,O. and Erbatur,F. (2000) Evaluation of crossover Mutation, fitness, viral diversity, and predictive mark-
techniques in genetic algorithm based optimum structural ers of disease progression in a computational model of
design. Comput. Struc., 78, 435–448. HIV type 1 infection. AIDS Res. Hum. Retroviruses, 20,
217. Ghaheri,A., Shoar,S., Naderan,M. and Hoseini,S.S. (2015) The 1314–1323.
applications of genetic algorithms in medicine. Oman Med. J., 233. Yu,Y., Schell,M.C. and Zhang,J.B. (1997) Decision theoretic
30, 406–416. steering and genetic algorithm optimization: application to
218. Karnan,M. and Thangavel,K. (2007) Automatic detection of stereotactic radiosurgery treatment planning. Med. Phys., 24,
the breast border and nipple position on digital mammograms 1742–1750.
using genetic algorithm for asymmetry approach to detection 234. Pei,Y., Kim,Y., Obinata,G. et al. (2011) Trajectory planning of
of microcalcifications. Comput. Methods Programs Biomed., a robot for lower limb rehabilitation. Conf. Proc. IEEE Eng.
87, 12–20. Med. Biol. Soc., 1259–1263 2011.
219. Pereira,D.C., Ramos,R.P. and do Nascimento,M.Z. (2014) Seg- 235. Jaremko,J.L. et al. (2002) Genetic algorithm-neural network
mentation and detection of breast cancer in mammograms estimation of cobb angle from torso asymmetry in scoliosis.
combining wavelet analysis and genetic algorithm. Comput. J. Biomech. Eng., 124, 496–503.
Methods Programs Biomed., 114, 88–101. 236. Smigrodzki,R. et al. (2005) Genetic algorithm for analysis
220. Baum,K.G., Schmidt,E., Rafferty,K. et al. (2011) Evaluation of mutations in Parkinson’s disease. Artif. Intell. Med., 35,
of novel genetic algorithm generated schemes for positron 227–241.
emission tomography (PET)/magnetic resonance imaging 237. Chen,H.Y., Chen,T.C., Min,D.I. et al. (1999) Prediction of
(MRI) image fusion. J. Digit Imaging, 24, 1031–1043. tacrolimus blood levels by using the neural network with
221. Wold,S., Sjostrom,M. and Eriksson,L. (2001) PLS-regression: genetic algorithm in liver transplantation patients. Ther. Drug
a basic tool of chemometrics. Chemom. Intell. Lab. Syst., 58, Monit., 21, 50–56.
109–130. 238. Prabhu,S.P. (2019) Ethical challenges of machine learning and
222. Tan,C., Chen,H., Wu,T. and Xia,C. (2011) Modeling the rela- deep learning algorithms. Lancet Oncol., 20, 621–622.
tionship between cervical cancer mortality and trace elements 239. Park,S.H. et al. (2019) Ethical challenges regarding artifi-
based on genetic algorithm-partial least squares and support cial intelligence in medicine from the perspective of scientific
vector machines. Biol. Trace Elem. Res., 140, 24–34. editing and peer review. Sci. Ed., 6, 91–98.
223. Khalil,A.S., Bouma,B.E. and Kaazempur Mofrad,M.R. (2006) 240. Park,S.H., Do,K.H., Kim,S. et al. (2019) What should medical
A combined FEM/genetic algorithm for vascular soft tissue students know about artificial intelligence in medicine? J. Educ.
elasticity estimation. Cardiovasc. Eng., 6, 93–102. Eval. Health Prof., 16, 18.
224. Zhou,X. et al. (2009) Identification of biomarkers for risk 241. Somashekhar,S.P. et al. (2018) Watson for oncology and
stratification of cardiovascular events using genetic algorithm breast cancer treatment recommendations: agreement with
with recursive local floating search. Proteomics, 9, 2286–2294. an expert multidisciplinary tumor board. Ann. Oncol., 29,
225. Tu,C., Zeng,Y. and Yang,X. (2005) A new approach to detect 418–423.
QRS complexes based on a histogram and genetic algorithm. 242. Liu,C., Liu,X., Wu,F. et al. (2018) Using artificial intelli-
J. Med. Eng. Technol., 29, 176–180. gence (Watson for oncology) for treatment recommendations
226. Nguyen,L.B., Nguyen,A.V., Ling,S.H. and Nguyen,H.T. (2013) amongst Chinese patients with lung cancer: feasibility study.
Combining genetic algorithm and Levenberg-Marquardt algo- J. Med. Internet Res., 20, e11087.
rithm in training neural network for hypoglycemia detection 243. Theodore,K. et al. (2019) Harnessing CURATE.AI as a digital
using EEG signals. Conf. Proc. IEEE Eng. Med. Biol. Soc., therapeutics platform by identifying N-of-1 learning trajectory
5386–5389 2013. profiles. In: Advanced Therapeutics.
227. Hoh,J.K., Cha,K.J., Park,M.I. et al. (2012) Estimating time to 244. Londhe,V.Y. and Bhasin,B. (2019) Artificial intelligence
full uterine cervical dilation using genetic algorithm. Kaohsi- and its potential in oncology. Drug Discov. Today., 24,
ung J. Med. Sci., 28, 423–428. 228–232.
228. Ocak,H. (2013) A medical decision support system based on 245. Henry,K., Hager,D., Pronovost,P. and Saria,S. (2015) A tar-
support vector machines and the genetic algorithm for the geted real-time early warning score (TREWScore) for septic
evaluation of fetal well-being. J. Med. Syst., 37, 9913. shock. Sci. Transl. Med., 7, 299ra122.
246. Amorim,F.F. and Santana,A.N. (2017) Automated early warn- 250. Drug and Device News (2018) P & T : a peer-reviewed journal
ing system for septic shock: the new way to achieve inten- for formulary management, 43, 382–429.
sive care unit quality improvement? Ann. Transl. Med., 251. Badgeley,M.A. et al. (2019) Deep learning predicts hip fracture
5, 17. using confounding patient and healthcare variables. NPJ Digit.
247. Khurana,H.S. et al. (2016) Real-time automated sampling of Med., 2, 31.
electronic medical records predicts hospital mortality. Am. J. 252. McKinney,S.M. et al. (2020) International evaluation of an AI
Med., 129, 688–698.e2. system for breast cancer screening. Nature, 577, 89–94.
248. Saria,S., Butte,A. and Sheikh,A. (2018) Better medicine through 253. Papadakis,G.Z. et al. (2019) Deep learning opens new horizons
machine learning: What’s real, and what’s artificial? PLoS in personalized medicine. Biomed. Rep., 10, 215–217.
Med., 15, e1002721. 254. Parekh, V. S., & Jacobs, M. A. Deep learning and radiomics in
249. Kerlikowske,K. et al. (2018) Automated and clinical breast precision medicine. Expert Rev. Precis. Med. Drug Dev., 4(2),
imaging reporting and data system density measures predict 59–72 (2019).
risk for screen-detected and interval cancers: a case-control 255. Khoury,M.J. et al. (2016) Precision public health for the era of
study. Ann. Int. Med., 168, 757–765. precision medicine. Am. J. Prev. Med., 50, 398–401.

Baaa 010

Uploaded by

Copyright:

Available Formats

Baaa 010

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Baaa 010

Uploaded by

Copyright:

Available Formats

Database, 2020, 1–35

Artiﬁcial intelligence with multi-functional

© The Author(s) 2020. Published by Oxford University Press. Page 1 of 35

subjects to understand speciﬁc scenarios and optimize decision-making. Implementation

Introduction precision medicine has evolved as a central innovation pillar

Approach Objectives Approach AI & ML

Approach Objectives Approach AI & ML

ML algorithms Examples in healthcare

ML algorithms Examples in healthcare

Role of AI in patients the point-of-care, advanced

as offering a multi-user-based graphical interface with con- References

You might also like