Circulation Research

HYPERTENSION COMPENDIUM

Artificial Intelligence in Hypertension

Seeing Through a Glass Darkly
Sandosh Padmanabhan , Tran Quoc Bao Tran, Anna F. Dominiczak

ABSTRACT: Hypertension remains the largest modifiable cause of mortality worldwide despite the availability of effective
medications and sustained research efforts over the past 100 years. Hypertension requires transformative solutions that
can help reduce the global burden of the disease. Artificial intelligence and machine learning, which have made a substantial
impact on our everyday lives over the last decade may be the route to this transformation. However, artificial intelligence
in health care is still in its nascent stages and realizing its potential requires numerous challenges to be overcome. In this
review, we provide a clinician-centric perspective on artificial intelligence and machine learning as applied to medicine and
hypertension. We focus on the main roadblocks impeding implementation of this technology in clinical care and describe
efforts driving potential solutions. At the juncture, there is a critical requirement for clinical and scientific expertise to work in
tandem with algorithmic innovation followed by rigorous validation and scrutiny to realize the promise of artificial intelligence-
enabled health care for hypertension and other chronic diseases.

Key Words: artificial intelligence ◼ blood pressure ◼ clinical trial ◼ hypertension ◼ machine learning

O
ver the last decade, artificial intelligence (AI) has
Downloaded from http://ahajournals.org by on August 11, 2023

transformed all aspects of life (advertising, finance, Hypertension Compendium

legal, and medical). Concurrently, precision medi-
cine has become embedded into medical parlance with We shall consider hypertension as an ideal disease
notable successes in personalized cancer treatment, exemplar to contextualize AI in medicine.3–5 Globally,
giving rise to the expectation that integration of big hypertension is the leading cause of death and disability-
data and omics along with AI will lead to personalized adjusted life years,6 accounting for more cardiovascular
treatment for all diseases. However, both fields have death than any other modifiable risk factor and second
fallen victim to cognitive biases that have led to mis- only to cigarette smoking as a preventable cause of mor-
calibration of expectations from early successes. There tality. Despite the availability of multiple drugs, hyperten-
is now greater recognition that clinical implementa- sion remains poorly controlled with major shortfalls in
tion of AI-based solutions need evidence generation awareness of hypertension, uptake of antihypertensive
similar to clinical trials.1,2 In this review, we shall pro- therapy, and adequacy of blood pressure (BP) control.7,8
vide a clinician-centric perspective of AI and machine Precision medicine for hypertension has been the sub-
learning (ML) as applied to cardiovascular medicine to ject of recent articles7,8 and treatment guidelines9,10 high-
highlight the critical requirement of scientific expertise lighting complexities in the architecture of the disease,
and scrutiny in tandem with technological innovation to challenges in management, and the need for transfor-
avoid negative consequences leading to an AI winter in mation.3 Despite considerable biologic heterogeneity, the
health care. The combination of advances in data sci- diagnosis and management of hypertension are reliant
ence, biotechnology, and digital technology permitting upon the accurate measurement of BP, which is chal-
the collection and integration of large amounts of data lenging in clinical practice. While the use of standardized
is essential for the application of AI. techniques and appropriate observer training improves

Correspondence to: Anna F. Dominiczak, DBE, MD, BHF Glasgow Cardiovascular Research Centre, Institute of Cardiovascular and Medical Sciences, University of
Glasgow, Wolfson Medical School Bldg, University Ave, Glasgow G12 8QQ. Email [email protected]
For Sources of Funding and Disclosures, see page 1116.
© 2021 American Heart Association, Inc.
Circulation Research is available at www.ahajournals.org/journal/res

1100   April 2, 2021 Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106

 Padmanabhan et al Artificial Intelligence in Hypertension

HYPERTENSION COMPENDIUM
through experience). Currently, the field of AI is still in
Nonstandard Abbreviations and Acronyms its nascent stages at an initial level termed artificial nar-
row intelligence with the 2 subsequent theoretical levels
ABPM ambulatory BP monitoring termed artificial general intelligence and artificial superin-
AI artificial intelligence telligence.13 Artificial narrow intelligence represents pat-
AKI acute kidney injury tern recognition, in a precisely defined single task, using
ANN artificial neural network medium to large data sets with applications in text, voice,
AUROC area under the receiver operating and image-based classification and clustering. Artificial
characteristic curve general intelligence, the second level, represents cog-
BP blood pressure nitive prowess that reaches human levels with the abil-
CONSORT-AI Consolidated Standards of Report-
ity to reason, argue, memorize, and solve problems like
ing Trials-AI humans. The third and highest level, artificial superintel-
EHR electronic health record
ligence, is a theoretical concept representing cognitive
capacity equivalent to more than the entire humanity.13
ML machine learning
The terms AI and ML have been used interchangeably,
SPIRIT-AI Standard Protocol Items: Recom-
but it is important to understand the differences between
mendations for Interventional
Trials-AI
them. AI is a technology which enables a machine to
simulate human behavior, while ML is a subset of AI
TRIPOD-ML Transparent Reporting of a Multi-
variable Prediction Model for Indi-
which allows software or algorithm to automatically learn
vidual Prognosis or Diagnosis-ML from past data without programming explicitly. AI has
a very wide scope and encompasses learning, reason-
ing, and self-correction; ML has limited scope covering
accuracy, this reading is often variable, influenced by a only learning and self-correction when introduced with
number of factors and prone to error.11 The treatment new data. As most of AI applications in health care are
strategy has remained essentially unchanged over the based on ML, we shall describe this in more detail. First,
past half-century, and personalizing treatment has not we start with the 2 main schools of AI that informed the
progressed beyond considering African ancestry and development of ML—symbolists and connectionists.14,15
serum renin levels. Furthermore, there is a lack of inte- The symbolists have sought to build intelligent machines
gration of the considerable genomic, molecular, and by coding in logical rules and representations of the
Downloaded from http://ahajournals.org by on August 11, 2023

physiological research results into screening, diagnostic, world. The connectionists have sought to construct arti-
and management plans.9,10 In >50% of patients, the ini- ficial neural networks (ANN), inspired by biology, to learn
tial treatment does not control BP adequately leading to about the world (See Box 1: Symbolists and Connection-
multiple clinic visits at varying intervals to try dose titra- ists). ML is defined as the study of computer algorithms
tion, switching, or adding drugs until a satisfactory out- that improve automatically through experience, and we
come is achieved or intolerable side effects occur, or no provide a brief overview of ML algorithms below.
further progress seems possible. The high prevalence of
hypertension in relation to the number of primary care or
specialist doctors means more and more responsibility BOX 1: SYMBOLISTS AND
for the follow-up and management of hypertension need CONNECTIONISTS
to be divested to the patient or to intelligent systems
powered by AI. However, several hurdles lie in the path Symbolic AI
of implementing AI to transform hypertension care, and Symbolic AI is also known as rule-based AI, classic AI, and
in this review, we provide a clinician-centric perspective good old-fashioned AI and is based on the notion that the
on AI and ML as applied to medicine and hypertension. world can be understood in the terms of structured repre-
sentations. This relies on inserting human knowledge and
behavioral rules into computer codes processing strings
ARTIFICIAL INTELLIGENCE of characters representing real-world entities or concepts
The formal organization of AI as a discipline started in through symbols. Here, a human-created knowledge base
the 1950s when Alan Turing posed the question can or an expert system linked to an inference engine selects
machines think?12 The simplest definition of AI is human rules to apply to given symbols. The classical example of
intelligence simulated by machines. An AI system com- symbolic AI in medicine is the clinical decision support sys-
prises reasoning (making inferences using data), natu- tem.16 Typically, clinical decision support systems are used
ral language processing (ability to read and understand to automatically remind the clinician of a specific action
human languages), planning (ability to act autonomously (alerts, reminders, order sets, and drug-dose calculations) or
and flexibly to construct a sequence of actions to reach a to provide performance feedback on quality indicators (care
final goal), and ML (algorithms that improve automatically summary dashboards). The applications of clinical decision

Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106 April 2, 2021   1101

 Padmanabhan et al Artificial Intelligence in Hypertension
HYPERTENSION COMPENDIUM

support systems are limited as the real world has a tremen- the algorithm cannot be tested on a large number of sce-
dous amount of data and variations, and it is impossible to narios since patient lives are at stake. Deep learning is
anticipate all permutations and build rules for every possibil- an autonomous, self-teaching system in which algorithms
ity. Nevertheless, within its limited field of application, clinical are trained to find patterns then used to make predictions
decision support systems have been shown to be effective about new data. Its functioning is based on ANNs. Deep
at improving health care process measures across diverse learning and reinforcement learning are both systems
settings, although evidence for clinical, economic, workload, that learn autonomously. The difference between them
and efficiency outcomes remains sparse.17 is that while deep learning applies knowledge learned
from a training data set to a new data set, reinforcement
learning uses continuous feedback to dynamically adjust
Connectionist AI its actions to maximize reward. A technical description of
In contrast to symbolic AI, connectionist AI models pro- different algorithms is beyond the scope of this review,
cesses based on how the human brain works with its but the key features of each algorithm are provided in
interconnected neurons, which follows from the notion Table 1 and Figure. Complex ML models are built on big-
that intelligent decision-making is possible through an data available from electronic health records (EHRs) and
interconnected system of small processing nodes of unit other databases, and frequently, these are accompanied
size.15 A system built with connectionist AI gets more by hype and claims of superiority over clinical experts.
intelligent through increased exposure to data by learn- While guidelines19 have now been established for report-
ing the patterns and relationships associated with it. One ing ML studies, it is critical that the medical readership
example of connectionist AI is ANN, where connections of the studies be familiar with the terminology and prop-
are represented numerically, and learning occurs by erties of metrics used to assess the performance of AI
modifying connection strengths based on experience. tools. In Box 2, we provide a list of performance measures
Each network contains hundreds artificial neurons or used in the evaluation of ML algorithms.
processing elements. Although ANN models get more
intelligent with increased exposure, they require a foun-
dation of accurate information to initiate the learning pro- BOX 2: PERFORMANCE MEASURES FOR
cess. ANN is only one of several types of ML algorithms ML ALGORITHMS
which have advanced the field of AI recently.
Classification Metrics
Downloaded from http://ahajournals.org by on August 11, 2023

We shall use a hypothetical example of an ML algo-

Machine Learning rithm built using a multitude of variables from a large
ML is divided into 4 subcategories: supervised learning, EHR to predict hypertension. In this model, the outcome
unsupervised learning, reinforcement learning, and deep is hypertension, and this label for each patient is pro-
learning.18 Supervised learning is used when the algo- vided by hypertension specialists based on adjudication
rithm’s exact task can be precisely defined with the data of longitudinal health records. The ML algorithm claims
at hand, while unsupervised learning is akin to learning to predict hypertension status using a set of variables
without a teacher. In other words, supervised learning is from the data set. The binary outcome in this example is
done using ground truth, meaning each sample is labeled whether hypertension is present or absent (normoten-
with an outcome or diagnosis by experts that are used sion). To show that a model accurately differentiates one
to build a model that best approximates the relationship outcome from another, there are 4 types of outcomes
between input variables and the known outcome in the that could occur.
data. In contrast, unsupervised learning does not have 1. True positives are when the model correctly pre-
labeled outputs, so its goal is to infer the natural structure dicts hypertension status.
present within a set of data points. From different data 2. True negatives are when the model predicts nor-
sets provided to an unsupervised learning algorithm, it motension and the patient is normotensive.
finds associations on its own. Such a model can discover 3. False positives occur when the model predicts
new drug-drug interactions or cluster patients accord- hypertension while the patient is normotensive.
ing to the attributes they display. Reinforcement learning 4. False negatives occur when the model predicts nor-
is an autonomous, self-teaching system that essentially motension while the correct label is hypertension.
learns by trial and error. It performs actions with the aim of These 4 outcomes are often plotted on a confusion
maximizing rewards, or in other words to achieve the best matrix which is a contingency table yielding the follow-
outcomes. Although reinforcement learning shares some ing metrics with the best prediction approaching 100%
features with unsupervised learning, unlike unsupervised or 1.0.
learning, it involves input from a human expert. Reinforce- 1. Positive predictive value or precision: the propor-
ment learning is typically used in robot and game devel- tion of positive cases that are true positive rather
opment, but its applications in health care are limited as than false positives.

1102   April 2, 2021 Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106

 Padmanabhan et al Artificial Intelligence in Hypertension

HYPERTENSION COMPENDIUM
Table 1. Common Machine Learning Algorithms

Algorithm Underlying principles Strengths Weaknesses

Unsupervised learning
 K-Means The algorithm identifies k number of centroids and computes the Relatively simple and compu- The user must identify the number of
clustering sum of squared Euclidean distance between the centroids (cen- tationally fast to implement. clusters, which can be challenging.
ters of the clusters) and the data points on a coordinate plane.
Flexible approaches with Reduced accuracy if the true under-
Each data point is then allocated to the closest centroid to keep
reasonable results for most lying clusters are not globular or of
the clusters as small the possible. The final centroids are com-
problems. varying sizes and density.
puted by taking the averages of all the data points that belong to
each cluster.
DBSCAN DBSCAN groups together points that are close to each other based Does not require user to User must specify how close points
on the predetermined Euclidean distance and a minimum number of know the number of clus- in each cluster should be to each
points to be included within each cluster. ters. other and the minimum number of
points in a dense region.
Work well with arbitrary Poor performance if the true under-
(nonglobular) cluster shape. lying clusters have different density.
 Hierarchical/ Each data point is initially considered as an individual cluster. At Easy to understand and All the approaches to calculate the
agglomerative each subsequent iteration, the 2 closest clusters are merged. This implement. similarity between clusters has its
clustering process is repeated until only one cluster remains. own disadvantages.
There are several approaches to measure the similarities of different
clusters:
1. Maximum linkage clustering
2. Minimum linkage clustering The algorithm output can Cannot work with data sets contain-
be visualized with a den- ing a mixture of different data types.
3. Average linkage clustering
drogram.
4. Centroid linkage clustering
5. Ward’s minimum variance
The final number of clusters can be determined by selecting an arbi- Work well with arbitrary Poor performance with high-dimen-
trary cutoff on the resulted dendrogram. (nonglobular) cluster shape. sional data.
Mean shift Mean shift begins with a circular sliding window centered at a ran- No need to select the num- The selection of window size can be
domly selected data point, with a predefined radius. At every subse- ber of clusters. challenging.
quent iteration, the window is gradually shifted towards the region of
Downloaded from http://ahajournals.org by on August 11, 2023

higher point density. The algorithm continues until there is no longer

a shift that can increase the point density within the window. If there
is overlapping between multiple sliding windows, the window con-
taining the most points is preserved.
 Expectation The algorithm begins by randomly assigning initial parameters: mean Can identify nonglobular The user must identify the number
maximization and SD based on the user-defined number of clusters. The algo- clusters. of clusters.
rithm then refines the parameters using an iteration of 2 steps: An
expectation (E) step and a maximization (M) step.
In the E step, the algorithm calculates the possibility of each data Allow for the extraction of Can be much slower than other
point belonging to a cluster based on the initial parameter values. membership probability for algorithms.
In the M step, a new set of parameters is computed to maximize each data point.
the likelihood of the distribution provided in the E step. At the
end of the process, each data point is assigned to its most prob-
able cluster.
 Affinity Affinity propagation works by identifying exemplars, which are The number of clusters Slow and memory-heavy, making it
propagation members of the data set that are representative of clusters. does not need to be preset. unoptimized for large data set.
The algorithm takes the similarities between pairs of data
Insensitive to outliers. Assumes the true underlying clus-
points as input. By exchanging real-valued messages between
ters to be globular in shape.
all data points, the algorithm gradually identifies a high-quality
set of exemplars, from which corresponding clusters are pro-
duced. Points that share the same exemplar are grouped in the
same cluster.
Supervised learning
 Logistic Logistic regression is a statistical technique used to estimate the Computationally efficient Unable to handle data set with large
regression probability of a binary response using logistic/sigmoid function. and easy to implement. number of categorical inputs.
Logistic regression can handle multiclass classification by using the
Interpretable Vulnerable to overfitting
one-vs-rest scheme, in which each class is treated as a binary clas-
sification problem. Does not require scaling of Colinearity between input variables
input features. may negatively affect the accuracy.
Provides the probability
score for the observations.
(Continued )

Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106 April 2, 2021   1103

 Padmanabhan et al Artificial Intelligence in Hypertension
HYPERTENSION COMPENDIUM

Table 1. Continued

Algorithm Underlying principles Strengths Weaknesses

 K-Nearest KNN assumes that similar data points are in close proximity to each Simple algorithm to imple- User must define the number of
neighbors other. The algorithm classifies a data point based on the class of ment and interpret. nearest neighbors to use.
the nearest neighbors. The number of the neighbors to be used is
Quick computing time Memory intensive
predefined by the user. The majority class of the neighbors is taken
to be the prediction of the query instance. No major assumptions Poor performance with high-dimen-
about data. sional data
Sensitive to outliers.
Naïve Bayes Naïve Bayes is a classification technique based on Bayes’ Theorem. Quick computing time The assumption of independence of
The classifier calculates the probability of a data point belonging to features rarely holds true in real-life
a class given a set of variables. This is under the assumption that settings.
the effect of a variable on a given class is independent to the other
variables.
Mathematically, for each given class:
1. Calculate probabilities for each input variable, conditional on the Excellent performance if the
given class. assumption of the indepen-
dence of features applies.
2. Calculate the joint conditional probability for all the variables.
3. Compute the conditional probabilities for the class using the Works well with categorical
Bayes rule. input variable.
Finally, the data point is assigned the class with the highest prob-
ability.
 Support SVM checks for a hyperplane (or a line in 2-dimensional data) that Works well when there is Low accuracy when samples
vector machine best separate the 2 different classes. The best hyperplane is de- a clear boundary between overlap.
fined as the one that maximizes the distance to the nearest element classes.
of each class.
Memory efficient. Does not work well with large data
set.
Does not provide probability esti-
mates.
Decision tree Decision tree starts with the best feature as the root node and re- Closely mimics the human Prone to overfitting
cursively divides the data into smaller subsets. This process contin- decision-making process.
Downloaded from http://ahajournals.org by on August 11, 2023

ues until all the data points in a subset have the same classification.
Easy interpretation. Affected by noise.
If a subset contains multiple classification and there are no more
features for splitting, the algorithm returns the most frequent class Work well with both numeri-
as the final classification. The decision of making strategic splits is cal and categorical features.
based on whether a split will increase the homogeneity of the resul-
tant subsets, which can be calculated using the Gini index.
Random forest Random forests are an ensemble learning algorithms, which is a Resilient to overfitting. Take long to train.
combination of multiple base decision tree. Each base decision tree
Works well with nonlinear Low interpretability.
is trained on a subset of the initial training set. The final classification
data.
of a data point is based on the simple majority voting scheme.
Robust to outliers.
No scaling required.
Work well with both numeri-
cal and categorical features
 Boosting algo- Like random forest, boosting algorithms is a form of ensemble learn- Easy to interpret. Sensitive to outliers
rithms (XGBoost, ing. Instead of creating several base models independently like
Resilient to overfitting Slow to train since the model is built
Adaboost, Cat- Random Forests, boosting algorithms create the base Decision Tree
consequentially.
boost, LightGBM) sequentially. Each successive tree focuses on the error of the previ- No scaling required
ous trees by assigning higher weights to the misclassified instances
Good performance
while keeping the weights of the correctly classified instances the
same. The final model is based on the combination of all weighted Work well with both numeri-
base models. cal and categorical features.

Deep learning
Neural networks Neural network consists of neurons, which are the basic processing The hierarchical architec- Require large amounts of data
mathematical functions of the network. Each neuron can transform ture allows neural network
Computationally expensive
the input data with a weighted linear summation, followed by a to capture the complex,
nonlinear activation function. The neurons are organized into several nonlinear relationships from Neural networks are black boxes.
layers, where the output of one layer is used as the input of the next the data.
Heavily depend on the quality of the
layer. Each layer can thus identify the important features from the
training data
input of the previous layer and further process them.
Excellent performance with Design and development of a neural
large data sets. network are significantly more chal-
lenging than other algorithms.

DBSCAN indicates Density-based spatial clustering of applications with noise; KNN, K-Nearest neighbors; and SVM, support vector machine.

1104   April 2, 2021 Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106

 Padmanabhan et al Artificial Intelligence in Hypertension

HYPERTENSION COMPENDIUM
Downloaded from http://ahajournals.org by on August 11, 2023

Figure. Data flow and machine learning algorithms in health care research.
Flow of various types of data (eg, clinical, demographic, and genomic) to integration (eg, electronic health records and biobanks) from where it can
be extracted to perform machine learning modeling. CNN indicates convolutional neural network; DBN, deep belief network; DBSCAN, density-
based spatial clustering of applications with noise; GAN, generative adversarial network; KNN, K-nearest neighbors; LSTM, long short-term
memory; RNN, recurrent neural network; and SVM: support vector machine.

2. Negative predictive value: the proportion of nega- 3. Sensitivity or recall: the proportion of true positive
tive cases that are true negatives rather than false cases that are correctly identified.
negatives.

Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106 April 2, 2021   1105

 Padmanabhan et al Artificial Intelligence in Hypertension
HYPERTENSION COMPENDIUM

4. Specificity: the proportion of true negative cases prediction of hypertension from clinical data in EHRs,21–
which are correctly identified. 23
diagnosis of hypertension,24 prediction of systolic BP
5. Accuracy: the percentage of the correct predic- from retinal fundal images,25 and prediction of absolute
tions. This is helpful only when the data sets are risk reduction in cardiovascular events using data from
symmetrical, where values of false positive and randomized clinical trials.26 These form a smaller subset
false negatives are almost same. of ML studies in cardiovascular diseases in general, and
6. F1 score: this measure represents the harmonic mean these are presented in Table 2, which contains details
of precision (or positive predictive value) and recall of individual studies, the performance metrics evaluated
(sensitivity) in which both are maximized to the largest and the source of the data set on which ML was applied.
extent. F1 is usually more useful than accuracy, espe- Below we describe a sample of hypertension studies
cially if the data set has an uneven class distribution. that used ML methods highlighting the predominant
7. Area under precision-recall curve the receiver oper- usage of ML methods as an analytical technique, rather
ating characteristic curve (AUROC): The PR curve is than extending these to robust validation and clinical
a graphical plot of positive predictive value (or pre- applications.
cision) against sensitivity (or recall) and is a better
measure of accuracy for classification tasks involving Diagnosing Hypertension
highly imbalanced data sets (for instance very few
hypertension cases and large number of normoten- Koshimizu et al57 applied deep neural networks to time-
sive patients in our hypothetical example). An area series data from serial home and clinic BP measure-
under precision-recall curve of 0.5 is preferred. The ments to predict BP variability and mean BP values.
ROC curve is the plot between sensitivity and (1− This study underlines one of the major unmet needs in
specificity) or false positive rate versus true positive hypertension management, the need to accurately clas-
sify hypertension status using out-of-office BP mea-
rate. An AUROC of 1.0 represents perfect predic-
surements. Hypertension guidelines advise the use of
tion; an AUROC equal to or above 0.8 is preferred.
out-of-office BP measurements using ambulatory BP
8. Calibration determines how well the model’s pre-
monitoring (ABPM) or home BP monitoring for the
dicted probability approximates the actual event
diagnosis and classification of hypertension.10,58 ABPM
probability. Calibration is evaluated by plotting the
offers the advantage of diagnosing hypertension and
actual observed event frequency against the aver-
also classifying the ABPM status of patients into cat-
Downloaded from http://ahajournals.org by on August 11, 2023

age predicted probability for each decile of a popu-

egories—true hypertension, white-coat hypertension, or
lation and quantitatively and qualitatively assessing
masked hypertension.59,60 Detecting true hypertension
the deviation from a diagonal line having an inter-
and defining these subgroups of patients is important as
cept of 0 and slope of 1.20
relying on clinic BP alone can lead to adverse outcomes
though the initiation of antihypertensive medication in
Regression Metrics those with white-coat hypertension potentially exposing
the patient to adverse drug effects and greater medica-
Evaluation of ML models predicting a continuous out-
tion costs. Conversely, the failure to diagnose and initi-
come is assessed by the following metrics.
ate therapy in those with masked hypertension exposes
1. Explained variance compares the variance within
the patient to greater risk of hypertension-mediated
the expected outcomes and compares that to the
organ damage, cardiovascular events, and mortality.59–61
variance in the error of the test model.
Despite the numerous benefits of ABPM, there are a
2. Mean squared error is simply defined as the aver-
number of disadvantages encountered in clinical prac-
age of squared differences between the predicted tice, notably the access to and the availability of ABPM
output and the true output. A value closer to 0 indi- devices (often only available in secondary care); the sig-
cate better performance. nificant difference in cost of the device compared with
3. R2 coefficient (coefficient of determination) rep- clinic BP; training of practitioners in measurement and
resents the proportion of variance in the outcome interpretation; inability to record during physical activity;
that the model is capable of predicting based on its presence of cardiac dysrhythmia (eg, atrial fibrillation);
features. An R2 of 0 means prediction is impossible sleep disturbance; and patient discomfort.62–65 The abil-
while R2 of 1 means completely accurate predic- ity to predict accurately ABPM status without ABPM
tion. An R2 above 0.6 indicates a useful algorithm. would allow primary care physicians to effectively diag-
nose and manage patients minimizing risk or harm from
antihypertensive therapy, hypertension undertreatment,
ML APPLICATIONS IN HYPERTENSION or the ABPM device itself. This is potentially achievable
ML has been used to explore some of the questions in with ML using large data sets with participants having
hypertension diagnosis and management and include both ABPM and clinic BP measured.

1106   April 2, 2021 Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106

 Padmanabhan et al Artificial Intelligence in Hypertension

HYPERTENSION COMPENDIUM
Table 2. Recent Applications of Machine Learning in Cardiovascular Research

Authors Sample Cross External Refer-

Outcome and year size Input variables Results validation validation Database ence
Support vector machines
 Future occur- Mezza- 1216 Age, sex, smoking, diabetes, Accuracy: 95% 5-fold Accuracy: 75% First data set: Calabrian 27

rence of ischemic testa et al, cardiovascular comorbidi- cross vali- dialysis registry from the
ROC-AUC: 50% ROC-AUC: 74%
heart disease 2019 ties, cholesterol, antihyper- dation Institute of Clinical Physiol-
tensive drugs, creatinine, Precision: 89% Precision: 77% ogy of the National Re-
AST, ALT, dialysis vintage, search Council of Reggio
hemoglobin, calcium, phos- Calabria, Italy
phate, albumin, C-reactive
Recall: 94% Second data set: data
protein, markers of diabetes,
set derived from the
educational and marital sta- F1 score:91% Recall: 75%
HEMO clinical trial from
tus, malignancies.
F1 score:75% the database repository
National Institute of Dia-
betes and Digestive and
Kidney Diseases

 Predict hyperten- Li et al, 1488 Genes’ degree, K-core, 177 genes were 5-fold No The protein interaction 28

sion genes 2017 and betweenness, gene predicted cross vali- network data set: OPHID
ontology, human phenotype dation (The Online Predicted Hu-
ROC-AUC: 83%
ontology. man Interaction Database)

F1: 78% Hypertension genes data

set: T-HOD database (The
Precision: 79% Text-Mined Hypertension,
Obesity, and Diabetes
Recall: 77%
Candidate Gene Data-
base); HPO database
(Human Phenotype Ontol-
ogy database)

Naïve Bayes
 Prediction of Yakovlev et 5000 Age, sex, minimal hemoglo- Accuracy: 90% No No 5000 electronic medical 29

in-hospital hos- al, 2018 bin level, maximal levels of records of acute coronary
pitality glucose, aspartate and ala- syndrome patients, hospi-
nine transaminases, minimal talized from 2010 to 2016
platelet count
Downloaded from http://ahajournals.org by on August 11, 2023

 Prediction of Frizzell et 56 477 Age, sex, race, socioeco- C statistics: 62% No No Get With The Guidelines 30

30-day all-cause al, 2017 nomic status, medical history, Heart Failure registry from
readmissions in characterization of HF, admis- the American Heart As-
patients hospital- sion and discharge medica- sociation
ized for HF tions, vital signs, weights, se-
lected laboratories treatment,
and discharge interventions.
Density-based spatial clustering of applications with noise
 Automatic detec- Ivanov et 500 MRI slices of the heart short Systolic volume: No No Data set provided by the 31

tion and calcula- al, 2019 axis. Each slice has 30 imag- RMSE: 21.64 mL; National Heart, Lung, and
tion of left ven- es that are recorded sequen- MAPE: 23.4% Blood Institute
tricular volume tially in time and comprise
Diastolic volume:
the whole cardiac cycle.
RMSE: 44.92 mL;
MAPE: 20.18%
Ejection fraction:
RMSE: 7.96%;
MAPE: 12.05%
Boosting algorithm
 Predict the risk of Ye et al, 823 627 Demographics, laboratory, ROC-AUC: 91.7% No Prospective co- Patient electronic health 21

developing essen- 2018 and radiographic test re- hort: ROC-AUC: records extracted from the
tial hypertension sults, essential and second- 87% Maine Health Information
within the next 1 ary diagnoses and proce- Exchange network
y. (5 categories: dures, outpatient medication
very low, low, prescriptions, clinical utility
medium, high, or records, and socioeconomic
very high.) variables
 Predict clinical Wu et al, 508 Left atrial diameter, HDL C-statistic: 0.76 10-fold No A cohort of hypertension 32

outcomes in 2020 cholesterol, big endothelin-1, cross vali- patients who were hospi-
Brier score: 0.05
young patients right arm DBP, right leg dation. talized for 2012–2016 at
with hyperten- SBP, left leg SBP, right leg the Hypertension Center
sion. DBP, left arm SBP, mean of Fuwai Hospital
nocturnal arterial oxygen
saturation, past maximum
SBP, and urea
(Continued )

Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106 April 2, 2021   1107

 Padmanabhan et al Artificial Intelligence in Hypertension
HYPERTENSION COMPENDIUM

Table 2. Continued

Authors Sample Cross External Refer-

Outcome and year size Input variables Results validation validation Database ence

 Predict 5-y all- Motwani et 10 030 Segment stenosis score, ROC-AUC: 79% 10-fold No The Coronary CT Angi- 33

cause mortality al, 2016 segment involvement cross vali- ography Evaluation for
Brier score: 0.08
score, modified Duke in- dation Clinical Outcomes: An
dex, number of segments International Multicenter
with noncalcified, mixed registry.
or calcified plaques, age,
sex, gender, standard
cardiovascular risk fac-
tors, and Framingham risk
score.
 Estimate the Al’Aref et 13 054 25 demographic and clini- Accuracy: 81% 10-fold No The Coronary CT Angi- 34

likelihood of ob- al, 2019 cal variables, including age, cross vali- ography Evaluation for
ROC-AUC: 88%
structive CAD on sex, diabetes, hypertension, dation Clinical Outcomes: An
coronary comput- dyslipidemia, baseline choles- Sensitivity: 80% International Multicenter
ed tomography terol level, and the coronary registry
Specificity: 81%
angiography artery calcium score
PPV: 49%
NPV: 95%
 Predict 2-y Wallert et 51 943 39 variables including age, ROC-AUC: 85% 7-fold No Swedish national quality 35

survival after al, 2017 sex, weight, comorbidities, cross vali- register SWEDEHEART/
Sensitivity: 78%
first myocardial medications at admission dation RIKS-HIA
infarction and discharge, presenting Specificity: 75%
symptoms, ECG parameters
NPV: 97%
(rhythm, QRS, STT), pulmo-
nary rates, troponin, heart PPV: 28%
rate, SBP, reperfusion
Accuracy: 76%
 Prediction of AKI Huang et 947 091 Age, prior HF, cardiogenic ROC-AUC: 75% No ROC-AUC: 79% The NCDR CathPCI 36

risk after percu- al, 2018 shock within 24 h, cardiac registry

Brier score: Brier score:
taneous coronary arrest within 24 h, diabetes,
0.0617 0.0610
intervention CAD presentation, HF within
2 wk, preprocedure GFR, Calibration slope:
preprocedure hemoglobin, 1.008
admission source, BMI, PCI
Downloaded from http://ahajournals.org by on August 11, 2023

Predictive range: Calibration

status, pre-PCT left ventricu-
25.3% slope: 1.003
lar ejection fraction
Predictive range:
29.4%
K-nearest neighbors
 Diagnosis of HF Tabassian 100 Velocity, strain, and strain ROC-AUC: 89% ML model No Metabolic Road to 37

with preserved et al, 2018 rate curves were obtained training/ Diastolic Heart Failure
Accuracy: 85%
ejection fraction from myocardial segments at testing (MEDIA) project: the
rest and during submaximal Sensitivity: 86% process University Hospital of
exercise. was Wales and the Wales
Specificity: 82%
repeated Heart Research Institute
100 times (Cardiff, United King-
dom) and the Scuola
di Medicina of Eastern
Piedmont University (No-
vara, Italy)
Neural networks
 Continuous Tomašev 703 782 Age, sex, ethnicity, diabe- Accuracy: 55.8% 200 boot- No Electronic Health Records 38

prediction of AKI et al, 2019 tes, number of admissions, strapped from the US Department
ROC-AUC: 92.1%
within a 48-h diagnoses (ICD-9), pro- samples of Veterans Affairs
window cedures (CPT code), lab AUPRC: 29.7%
results (biochemistry, hema-
tology, cytology, toxicology,
microbiology, and histopa-
thology), medications and
prescriptions, orders, vital
signs, health factors, and
note titles.
The data of each patient was
recorded as a sequence of
6-h windows.

(Continued )

1108   April 2, 2021 Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106

 Padmanabhan et al Artificial Intelligence in Hypertension

HYPERTENSION COMPENDIUM
Table 2. Continued

Authors Sample Cross External Refer-

Outcome and year size Input variables Results validation validation Database ence

 Detect diabetic Ting et al, 494 661 Retinal photographs (optic Referable diabetic Cluster- 10 external vali- Patients from the Singa- 39

retinopathy and 2017 disc and fovea) retinopathy: ROC- bootstrap dation data sets pore National Diabetic
related eye dis- AUC: 93.6%; sen- with each for referable dia- Retinopathy Screening
eases. sitivity: 90.5%; and patient betic retinopathy: Program between 2010
specificity: 91.6% as the ROC-AUC: and 2013
bootstrap 88.9%–98.3%;
Vision-threatening
sampling sensitivity:
diabetic retinopa-
clusters 91.8%–100%;
thy: ROC-AUC:
and specificity:
95.8%; sensitivity:
73.3%–92.2%
100%; and speci-
ficity: 91.1%
Possible glau-
coma: ROC-AUC:
94.2%; sensitivity:
96.4%; and speci-
ficity: 87.2%
Age-related
macular degenera-
tion: ROC-AUC:
93.1%; sensitivity:
93.2%; and speci-
ficity: 88.7%
 Identify deter- Koren et 30 705 Initial SBP and DBP, pulse ROC-AUC: 0.82 10-fold No Electronic medical charts 40

minants that al, 2018 pressure, age, BMI, weight, cross vali- of Maccabi Health Service
Higher initial BP
contribute to the sex, smoking, number of dation
predicts lower suc-
success of hy- antihypertensive treatments,
cess rates.
pertension drug antihypertensive drug class-
treatment es, proton pump inhibitors,
and HMG CO‐A reductase
inhibitors (statins)
 Predict pres- Maxwell et 110 300 62 examination items con- Accuracy: 92.07% Stratified No The examination records 23

ence of diabetes, al, 2017 sisting of the basic physical 10-fold provided by the Collabora-
ROC-AUC: 96%
hypertension, examination items, blood cross vali- tive Innovation Center on
Downloaded from http://ahajournals.org by on August 11, 2023

fatty liver, or a routine examination, liver AUPRC: 61% dation. Internet Healthcare and
combination of function test, and diagnosis Health Service of Henan
Precision: 91.5%
these 3 chronic results Province, Zhengzhou
diseases. Recall: 86.7% University

F-score: 82.3%
 Predict the pres- LaFreniere 379 027 Age, sex, BMI, SBP and DBP, Accuracy: 82.5% No No Electronic Medical Record 24

ence of hyperten- et al, 2016 HDL and LDL, triglycerides, from Canadian Primary
sion. cholesterol, microalbumin, and Care Sentinel Surveillance
urine albumin creatinine ratio Network
 Predict first car- Weng et 378 256 Sex, age, total cholesterol, ROC-AUC: 76% No No Electronic Medical Re- 41

diovascular event al, 2017 HDL cholesterol, SBP, blood cords from the Clinical
Sensitivity: 68%
over 10 y. pressure treatment, smoking, Practice Research Da-
diabetes, BMI, LDL choles- PPV: 18.4 talink
terol, triglycerides, C-reactive
Specificity:71%
protein, serum fibrinogen,
gamma-glutamyl transferase, NPV:96%
serum creatinine, HbA1c,
forced expiratory volume,
AST/ALT ratio, family history
of CHD, ethnicity, deprivation
index, hypertension, rheuma-
toid arthritis, chronic kidney
disease, atrial fibrillation,
chronic obstructive pulmonary
disease, mental illness, anti-
psychotic drug, oral cortico-
steroids, immunosuppressant
 Diagnosis of con- Acharya et 120 000 Two-second segmented, Accuracy: 99% Stratified Accuracy: 96% EEG signals from the 42

gestive HF al, 2018 regularized ECG signal. 10-fold Beth Israel Deaconess
PPV: 91%
cross vali- Medical Centre (BIDMC)
PPV: 97% dation Sensitivity: 97% Congestive Heart Failure
Database, Fantasia Data-
Sensitivity: 99% Specificity: 96%
base, and MIT-BIH Normal
Specificity: 99% Sinus Rhythm Database
(NSRDB)

(Continued )

Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106 April 2, 2021   1109

 Padmanabhan et al Artificial Intelligence in Hypertension
HYPERTENSION COMPENDIUM

Table 2. Continued

Authors Sample Cross External Refer-

Outcome and year size Input variables Results validation validation Database ence

 10-y CVD pre- Zhao et al, 109 490 Demographics, variables ROC-AUC: 79% 5-fold No De-identified copy of Elec- 43

diction 2019 in the ACC/AHA equa- cross vali- tronic Medical Records
AUPRC: 29%
tions (eg, blood pressure dation (Synthetic Derivative) at
measurements), physical Vanderbilt University Medi-
measurements (eg, BMI), cal Center
laboratory tests including
glucose, triglyceride levels,
and creatinine level, disease
phenotypes, medications,
SNPs associated with CVD.
 Prediction of car- Poplin et 284 335 Retinal fundus photographs Age: MAE=3.26 y No Age: MAE=3.42 Images from UK Biobank 25

diovascular risk al, 2018 y and EyePACS

Gender: ROC-
factors: age, sex,
AUC=97% Gender: ROC-
smoking status,
AUC=97%
SBP, and major Smoking status:
adverse cardiac ROC-AUC=71% Hb1Ac:
events. MAE=1.39
Major adverse car-
diac event: ROC-
AUC=70%
SBP: MAE=11.23
mm Hg
 Heartbeat classi- Elhaj et al, 110 109 ECG recordings Accuracy: 98% No No MIT-BIH database of ar- 44

fication of arrhyth- 2017 rhythmia.

mia detection
 Predict athero- Pfohl et al, 253 547 Race, age, sex, diagnoses, ROC-AUC: 79% No No Stanford Medicine Re- 45

sclerotic CVD risk 2019 procedures, medication search Data Repository s

Brier score:
orders, lab tests, clinical en- from Stanford Hospital and
0.0205
counter types, departments. Clinics and Lucile Packard
Children’s Hospital for be-
tween 1990 and 2018
 Identification of Arbabshi- 46 583 Noncontrast head CT slices. ROC-AUC: 85% No Clinical imple- The data set was ex- 46

intracranial hem- rani et al, mentation tracted from the authors’

Sensitivity: 73%
Downloaded from http://ahajournals.org by on August 11, 2023

orrhage 2018 affiliated health care sys-

Accuracy: 84%
tem’s picture archiving and
Specificity: 80% Sensitivity: 70% communication system.

Specificity: 87%
Detect CVD Moradi et 4110 Frontal chest X-ray images Accuracy: 84% No No NIH PLCO (National Insti- 47

al, 2018 tute of Health; prostrate,

lung, colorectal, and ovar-
ian cancer) data set
 Estimate blood Khalid et 8133 Waveform features from pre- Mean and SD 10-fold No University of Queensland 48

pressure al, 2018 processed photoplethysmo- of difference: cross vali- vital signs data set
graph signal segments. SBP: −0.1±6.5 dation
mm Hg and DBP:
−0.6±5.2 mm Hg
 Predicting lipid- Hu et al, 2480 Sex, ethnicity, deprivation, C-index: 81% 10-fold No The PREDICT Cardiovas- 49

lowering medica- 2020 diabetes history, smoking cross vali- cular Disease Cohort in
Accuracy: 79%
tion persistence status, BMI, age on the day dation New Zealand Primary Care
after the first CVD of discharge, days stayed in PPV: 69%
hospitalization the hospital, diagnosis type,
NPV: 82%
procedure of percutaneous
coronary intervention or
CABG, maximum complica-
tion and comorbidity level.
Random forest
 Predicting individ- Duan et al, 9361 Age, sex, ethnicity, SBP and C-statistic for ben- 250 No Individual participant 26

ual treatment ef- 2019 DBP, antihypertensive treat- efit: 0.60 bootstrap data from SPRINT and
fects for intensive ments, aspirin, statin, smok- samples ACCORD-BP trials.
Corrected 3-year
blood pressure ing status, cholesterol, BMI, of the data
RMST: 1068.71
therapy creatinine, triglycerides set, strati-
days.
fied by
Slope between treatment
predicted and ob- arm and
served ARR: 1.06. presence
of CVD
event

(Continued )

1110   April 2, 2021 Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106

 Padmanabhan et al Artificial Intelligence in Hypertension

HYPERTENSION COMPENDIUM
Table 2. Continued

Authors Sample Cross External Refer-

Outcome and year size Input variables Results validation validation Database ence

 Predict the occur- Ambale- 6814 Traditional risk factors, C-index: 81% No No Electronic Medical 50

rence of CVD Venkatesh demographics, anthropom- Research from the Multi-

Brier score: 0.067
et al, 2017 etry, site, medication use, Ethnic Study of Athero-
computed tomography, sclerosis
carotid ultrasonography,
MRI markers, questionnaire,
family history, alcohol use,
education level, economic
status/income, lab biomark-
ers, ECG.
 5-y survival pre- Samad et 171 510 Echocardiogram, age, sex, Accuracy: 96% 10-fold No Electronic Medical Re- 51

diction al, 2019 smoking status, height, cross vali- cords of a regional health
ROC-AUC: 80%
weight, heart rate, DBP dation system in Danville, PA
and SBP, cholesterol, 90
cardiovascular-relevant ICD-
10 codes, echocardiogram
date, date of death, or last
living encounter.
 Predict the risk Segar et 8756 Age, DBP, glycated hemo- C-index: 77% 1000 C-index: 74% The ACCORD trial data 52

of incident HF al, 2019 globin, serum creatinine, boot- set from 77 centers across
Hosmer-Lemeshow Hosmer-Leme-
hospitalization HDL-C, history of myocardial strapped the United States and
statistic: P = 0.29 show statistic: P
among patients infarction, presence of atrial replicates Canada
= 0.20
with diabetes fibrillation, fasting plasma
glucose.
 Predict the risk Dogan et 1704 Genome-wide DNA methyla- Accuracy: 74% No No The Offspring cohort from 53

for 5-y incident al, 2018 tion and DNA profiles, age, the Framingham Heart
Sensitivity: 70%
coronary heart sex, SBP, HDL cholesterol Study
disease level, total cholesterol level, Specificity: 74%
diabetes status, and self-
F1: 18%
reported smoking status.
Ensembles of multiple ML methods
 Identifying people Alaa et al, 423 604 473 variables divided into ROC-AUC: 77.4% Stratified No UK Biobank from 22 as- 54

at risk of CVDs 2019 9 categories: health and 10-fold sessment centers across
Downloaded from http://ahajournals.org by on August 11, 2023

Sensitivity: 70%
medical history, lifestyle and cross vali- England, Wales, and Scot-
environment, blood assays, PPV: 2.6% dation. land, from 2006 to 2010
physical activity, family his-
tory, physical measures,
psychosocial factors, dietary
and nutritional information,
and sociodemographics.
ML methods used: SVM,
random forest, neural net-
works, AdaBoost, and gradi-
ent boosting machines.
 Predict incident Alghamdi 32 555 62 attributes classified into ROC-AUC: 92% 10-fold No The FIT data set from 55

of diabetes et al, 2017 4 categories: demographic cross vali- the Henry Ford Affiliated
characteristics, disease his- dation Hospitals in metropolitan
tory, medication use history, Detroit, MI in the United
and stress test vital signs. States
ML methods used: Naïve
Bayes, Random Forest, and
Logistic regression
 Prediction of Ebrahimza- 106 Temporal features from 30- Sensitivity: 100% 10-fold No Atrial Fibrillation Prediction 56

paroxysmal atrial deh et al, min ECG segments cross vali- Database
Specificity: 96%
fibrillation 2018 dation
ML methods used: K-nearest Accuracy: 98%
neighbor, Support Vector
Machine, Neural Network

ACCORD indicates Action to Control Cardiovascular Risk in Diabetes; AKI, acute kidney injury; ARR, Absolute Risk Reduction; AUPRC, area under the precision-recall curve; BMI,
body mass index; CABG, Coronary Artery Bypass Graft; CAD, coronary artery disease; CHD, Coronary Heart Disease; CVD, cardiovascular disease; DBP, diastolic blood pressure;
FIT, Ford Exercise Testing; GFR, Glomerular Filtration Rate; HDL, high-density lipoprotein; HF, heart failure; ICD, International Classification of Diseases; LDL, low-density lipoprotein;
MAE, mean absolute error; MAPE, Mean Absolute Percentage Error; MCC, Matthews correlation coefficients; MRI, magnetic resonance imaging; NNT, number needed to treat; NPV,
negative predictive value; PPV, positive predictive value; RMSE, Root-Mean-Square Error; ROC-AUC, area under the receiver operating characteristic curve; SBP, systolic blood pres-
sure; and SPRINT, Systolic Blood Pressure Intervention Trial.

Predicting Incident Hypertension (retrospective-N=823 627 and prospective-N=680 810)

and validated a 1-year risk prediction model for incident
Ye et al21 extracted individual patient EHRs from
essential hypertension stratifying patients into 5-risk cat-
the Maine Health Information Exchange network
egories achieving predictive accuracies of 0.870 in the
Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106 April 2, 2021   1111
 Padmanabhan et al Artificial Intelligence in Hypertension
HYPERTENSION COMPENDIUM

validation cohort. Type 2 diabetes, lipid disorders, cardiovas- the clinical value of these findings still requires evidence
cular diseases, mental illness, clinical utilization indicators, from randomized trials to address possible confounding.
and socioeconomic determinants were recognized as driv-
ing or associated features of incident essential hypertension
in this model.21 This real-time predictive analytic model has Discovering Hypertension Genes
been deployed in the state of Maine for risk stratification. The rise of high-throughput -omic technologies has
Similarly, Kanegae et al22 analyzed health checkup resulted in the availability of large-scale genomic and
data from 18 258 individuals at hypertension diagno- other omic data sets for ML. ML has been used quite
sis and 2 preceding years and showed that clinic BP at extensively in genomics for predicting pathogenicity of
health checkups in the year or 2 before hypertension
coding variants,67 functional consequences of noncoding
diagnosis were the top 8 predictors of new-onset hyper-
variants,68 and prediction of some common complex phe-
tension achieving AUROCs of around 0.85.22
notypes and disease risk.69,70 In relation to hypertension,
Huan et al71 and Li et al28 used ML and network analysis
Diagnosing Secondary Hypertension to predict hypertension genes by assessing the interac-
Patients with secondary hypertension have a high risk of tion neighborhood of a gene. The research community
morbidity and mortality if not diagnosed and treated early. is beginning to apply ML to classify cardiovascular dis-
Secondary hypertension investigations are usually trig- eases based on a variety of -omics features, including
gered when patients present at a younger age and have the transcriptome and the microbiome,72,73 and these
suggestive signs and symptoms or present with an acute efforts are exploratory at best.
rise in BP from previously stable readings. One study The key message from these examples is the wide-
using data from EHRs of 7532 patients attempted to
ranging applications of ML algorithms on predominantly
predict 4 etiologies (renovascular hypertension, primary
observational data sets to develop clinical prediction
aldosteronism, thyroid dysfunction, aortic stenosis, and a
composite of any of the 4).66 Their model achieved an models and this is not exclusive to hypertension studies.
AUROC or 0.90 for the composite outcome.66 However, while some of the studies demonstrate pre-
diction accuracies that are many-fold higher than those
obtained using traditional statistical analysis methods,
Predicting Individual Treatment Effects
Downloaded from http://ahajournals.org by on August 11, 2023

these claims are contestable because of lack of external

By applying ML to SPRINT trial data, Duan et al26 found validation and clinical effectiveness studies. Almost all the
an individual patient’s predicted absolute benefit from papers focus on prediction accuracy as a performance
intensive treatment was not necessarily proportional to metric but rarely explain their prediction in a meaningful
their baseline cardiovascular disease risk, highlighting the
context. This has direct consequences on clinical use, as
clinical importance of heterogenous treatment effects for
the opaque nature of these predictions makes it diffi-
making individual treatment effect estimates. Modeling
individualized treatment effects can help clinicians identify cult for clinicians to accept and for regulators to approve.
which patients are more or less likely to benefit from treat- In simple terms, for clinicians to trust these algorithms,
ment, as compared to the average result from a random- they need to understand that the model makes the right
ized trial. Calculation of individualized treatment effects predictions for the right reasons and wrong predictions
rather than simply calculating baseline risk would poten- for the right reasons. It is increasingly recognized that
tially help direct therapy to patients most likely to benefit.26 interpretability of ML algorithms need to be considered
alongside prediction accuracies, and there are ongoing
Predicting Effective Antihypertensive Therapies efforts to tackle this.74

Koren et al40 used decision trees and neural networks

to analyze EHR data from 30 705 patients to predict Realizing the Dream of Clinical AI
the successful outcome defined as achieving BP lower While ML can classify a patient’s diagnosis or predict
than 140/90 mm Hg in at least one measurement within
a patient’s outcome accurately, real-world evidence that
90 days of starting treatment. The predictors used
deployment of these models has improved care and
were weight, age, body mass index, and smoking sta-
tus. Their results reflect current understanding that initial patient outcomes is lacking.75 Understanding the rea-
BP predicts lower success rates in attaining target BP. sons underlying this disparity between expectations and
They additionally show that concomitant treatment with reality is important in the design and development of
statins or proton pump inhibitors was associated with ML applications. We shall now detail the practical and
higher treatment success rates.40 While exemplifying logistic barriers to clinical implementation and potential
another avenue of interrogating data using ML, defining solutions (Table 3).

1112   April 2, 2021 Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106

 Padmanabhan et al Artificial Intelligence in Hypertension

HYPERTENSION COMPENDIUM
Table 3. Limitations of ML Models and Possible Solutions.

Limitations Solutions
Limited interpretability of ML models Using interpretable AI models (XGBoost, Decision Tree).
Performing post hoc explanation (ie, most important factors, representative exemplars).
The development of explainable AI models.
Counterfactual explanations.
Using a model with multistage architecture, allowing a clinician to inspect and gain potential insight into the
model’s decision process.
Lack in reasoning about cause-effect relations The development of causal AI, which can estimate the relationships between all variables.76
The incorporation of counterfactual reasoning into ML methods.34
Lack of external validation An initiative for data-sharing agreements to ensure that data is shared widely and rapidly.
Inadequate reporting of development and valida- The installment of rigorous reporting guidelines, such as TRIPOD-ML19 or SPIRIT-AI77 and CONSORT-AI.78
tion studies
Societal biases, risking unintended erroneous pre- Training and validation data sets used in algorithms construction should be representative of the target
dictions in minority subgroups population.
ML algorithms lack the capacity to self-monitor Human-in-the-loop system: incorporating human judgment into the AI systems to identify potential failures.
errors.
Developing flexible systems that can adapt and handle outlier events.

AI indicates artificial intelligence; CONSORT-AI, Consolidated Standards of Reporting Trials-AI; ML, machine learning; SPIRIT-AI, Standard Protocol Items: Recom-
mendations for Interventional Trials-AI; TRIPOD-ML, Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis-ML.

Technical Limitations ability of the network to generalize to other data, while

too few data will handicap the network’s capacity to solve
Besides their black-box nature, ML algorithms have sev-
the problem. However, with no reliable methods, many
eral technical limitations.79 Overfitting is a major issue
scientists rely on the use of systematic experimentation
common to all ML models. This can happen when a
to identify the best configurations for their specific data
model has become overly attuned to the training data,
sets. This strategy does not always guarantee the best
including minor random fluctuations, such that it does
result and can be time consuming.
not generalize to new data sets. This is the opposite of
Downloaded from http://ahajournals.org by on August 11, 2023

underfitting, where the model cannot fully capture the

predictive power of the data. Overfitting and underfit- Reporting AI Studies
ting can be overcome by making modifications to the The standard requirements for clinical implementation
training set or by optimizing of the parameters of the of any new diagnostic or therapeutic also apply to ML
model. For these reasons, training of a good model often algorithms. Thus, one of the roadblocks to clinical imple-
requires large data set, competent informatics skills, and mentation of ML algorithms is a failure to fulfill these
an adequate means of validation.79 Supervised learning standard requirements. Broadly, 2 key criteria have
algorithms specifically have limited capacity to only pre- been described as essential82: First, well-conducted and
dict predefined results and thus require manually labeled well-reported development and validation studies that
data. This means that the utility of supervised mod- describe an algorithm and its properties in detail, includ-
els is usually restricted to replicating, automating, and ing predictive accuracy in the target setting; second, well
standardizing human chores. Definition, evaluation, and conducted and transparently reported randomized clini-
judgment remain exclusive in the realm of human intel- cal trials that evaluate usefulness in the real world.82
ligence and insight.80 However, validation is a key issue Clinical prediction models using standard statistical
to unsupervised models. For an unsupervised model to methodologies undergo a scientifically rigorous process
be externally validated, knowledge of the true clusters to establish their diagnostic accuracy, which encom-
is required to be used as the basis for comparison with passes their safety, validity, reproducibility, usability,
the ML-generated clusters.81 However, such informa- and reliability with guidelines developed to improve the
tion is often not readily available. In that event, model robustnes and value of clinical prediction models through
validation remains solely on internal validation measures. the promotion of transparent and accurate reporting.20
However, the existing methods are not always reliable as In contrast to traditional statistics, whereby models are
they can be influenced by various data characteristics, explicitly programmed based on statistical theory and
including noise, variations in cluster densities, presence assumptions, ML models learn from examples without
of subclusters, and unequal cluster sizes.81 For neural the need for explicit rules to make decisions. Recent
networks, deciding the optimal number of hidden nodes systematic reviews have uniformly found that ML stud-
remains one of the most challenging aspects of network ies are rarely prospective, are at high risk of bias, lack
construction.79 Too many nodes will be detrimental to the data and code transparency and deviate from existing

Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106 April 2, 2021   1113

 Padmanabhan et al Artificial Intelligence in Hypertension
HYPERTENSION COMPENDIUM

reporting standards on or assessment, interpretation, and Interpretable ML are models where the inner mecha-
reproducibility.82,83 To rectify these shortcomings, guide- nism can be comprehended by humans to understand
lines are being established to standardize reporting of its decision-making logic.87 At present, a trade-off exists
ML studies: TRIPOD-ML (Transparent Reporting of a between performance and explainability (defined as the
Multivariable Prediction Model for Individual Prognosis degree to which an ML user can understand and inter-
or Diagnosis-ML),19 SPIRIT-AI (Standard Protocol Items: pret the prediction made by an ML model). The most
Recommendations for Interventional Trials-AI),77 and interpretable models are often associated with poor
CONSORT-AI (Consolidated Standards of Reporting performance (eg, linear regression and decision trees).
Trials-AI)78 are all intended to improve the transparency However, the algorithms of the best-forming mod-
and completeness of reporting of ML prediction mod- els involve nonlinearities with numerous interactions
els and clinical trials evaluating interventions involving between inputs. Even with the knowledge of the underly-
AI. The SPIRIT-AI and CONSORT-AI checklists contain ing mathematical principles, it is not always feasible to
15 and 14 new items, respectively, as extensions to the fully interpret the inner workings of models to under-
existing SPIRIT 201384 and CONSORT 201085 check- stand its decision-making logic and hence, these best
lists. The guidelines include requirements for reporting of performing models (eg, deep learning) are often the least
areas such as the quality and completeness of input data, explainable. Given the increasing use and superior per-
investigation of error cases, defining the clinical context, formance of black-box ML systems in health care, there
and the human-AI interaction involved. is a strong desire to understand the reasons behind their
decisions by improving the models’ explainability. This
has given rise to the subfield of explainable AI. The goal
Causal AI of explainable AI is to provide post hoc explanations for
Relying solely on predictive AI models in health care runs existing black-box models and thus communicate to the
the risk of devastating consequences if correlations are user why a black-box model made a particular decision.88
mistaken for causation. The importance of establishing The most common forms of explanation are either com-
causality is not new in clinical studies, and it is commonly municating the most important factors or features that
addressed through a randomized controlled trial. Ran- led to the decision or providing similar or representative
domized controlled trials, however, have their own limita- exemplars that support the decision. However, the post
tions, although the average treatment effect produced hoc explanations cannot perfectly simulate the original
by randomized controlled trials is a robust measure of model. Instead of providing concrete evidence, the expla-
Downloaded from http://ahajournals.org by on August 11, 2023

efficacy, it has limited utility in personalized treatment. nation model only attempts to replicate as closely as
Causal AI is a potential solution to deconvolute the possible the behavior of the original model. Even when
precise cause and effect relationships through power- then explanation model produces almost identical results
ful methods such as causal Bayesian networks, which to the initial predictions, it may use completely different
can estimate the relationships between all variables in features.88 As such, model-agnostic methods should be
a data set, resulting in an intuitive visual map showing used cautiously as they can provide a misleading expla-
which variables influence one another, as well as the nation of what the original model truly computes. Other
extent of their influence.76 The first-great advantage of strategies include counterfactual explanations89 or using
this approach is that these interactions do not need to or utilizing a model with multistage architecture, allowing
be specified a priori, making it a true discovery method. a clinician to inspect and gain potential insight into the
The second major advance is in counterfactual rea- model’s decision process.90
soning. This is exemplified by common clinical conun-
drums, such as What will happen if I take or do not take
an action? or What comorbid conditions could complicate Fairness in AI (Algorithmic Bias)
a treatment? Causal AI can potentially identify the root Blind spots in ML can reflect the worst societal biases,
causes of outcomes and also model interventions that risking unintended erroneous predictions in minority sub-
can change those outcomes by using algorithms to ask groups. Furthermore, there are growing concerns over the
what-if questions (counterfactuals).86 potential for amplifying biases present in the historical
data.91 Studies indicate that, in some contexts, AI systems
disproportionately affect groups that are already disadvan-
Explainable AI taged by factors, such as race, sex, and socioeconomic
The rise of sophisticated ML models has produced accu- background. In medicine, examples include hospital mor-
rate but obscure decision systems which hide their logic, tality prediction algorithms with varying accuracy by eth-
thus undermining transparency and trust. This has major nicity92 and algorithms which classify images of benign
consequences for the adoption of AI in clinical settings. and malignant moles underperforming on images of
A model’s transparency is brought together by 2 similar lesions in skin of color due to training on open data sets
but distinct concepts: interpretability and explainability. of predominantly fair-skinned patients.93 ML algorithms

1114   April 2, 2021 Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106

 Padmanabhan et al Artificial Intelligence in Hypertension

HYPERTENSION COMPENDIUM
are widely applied in the US health care systems to pre- frustrating both patients and doctors with inconsistent
dict which patients are likely to have the more complex results and a general lack of harmony with on-the-ground
medical needs and require preemptive measures. How- practices.98 The second example is the development of
ever, a recent study showed that this particular tool had an DeepMind’s AI for predicting acute kidney injury (AKI),
unintended consequence: Black patients who had more reported to be capable of detecting potentially fatal kid-
chronic illnesses than White patients were not flagged as ney injuries 48 hours before symptoms are recognized
needing extra care.94 A greater awareness of these issues by doctors.38 This study used recurrent neural networks
and empowering clinicians to participate critically in sys- to predict the probability of AKI from sequences of clini-
tem design and development will help guide researchers cal parameters from 703 782 adult patients from 1234
to ensure that the correct steps are taken to quantify bias clinical sites of the US Department of Veterans Affairs. In
before deploying models. There is growing recognition total, the data set comprised 6 billion entries and 620 000
that algorithms should be designed with the global com- features. For all defined severity levels of AKI, the model
munity in mind, and clinical validation should be performed correctly predicted ≈56% of all AKI episodes in a time
using a representative sample of the target population. interval of up to 48 hours, advancing the clinical diagnosis
at a chosen operating point with a precision of 33%. This
implies that of 3 AI-predicted cases true AKI was present
Health Care Stakeholders in only one patient and falsely predicted in the other 2.
In addition to clinicians, the health care system consists of Early prediction of AKI is considered a clinically valuable
multiple stakeholders with competing interests—patients, utility, but this premise was not supported when the real-
insurance companies, pharmaceutical firms, and the gov- life clinical impact of early prediction of AKI was assessed
ernment. Buy-in from all these stakeholders are essential through a mobile app implemented as a component of
to deploy an ML model. This requires a number of factors the DeepMind software Streams using an established
to be considered. First, it must be clear who is responsible UK National Health Service early detection algorithm for
for deploying an ML model and taking subsequent action AKI.99 Disappointingly, no improvement in the clinical out-
from the outcomes of the model.75,95 This person will need come was observed relative to a control group, although
to assess patient preferences regarding the use of a model the time to AKI recognition and treatment of nephrotoxic-
as well as ethical and medicolegal issues, including new ity was significantly shortened.99 These examples high-
obligations to treat.96 Second, it is still unclear where legal light the challenges to the deployment of ML applications
liability lies when ML algorithms result in harm and resolv-
Downloaded from http://ahajournals.org by on August 11, 2023

in the clinical routine and that despite the great promise

ing this will help clinical implementation of ML. Medicine is and potential they will have to clear substantial hurdles.
an ever-changing field, where clinical and operational prac-
tices in clinical settings constantly evolve. This may result
in changes in the input data and may no longer resemble AI or Natural Stupidity
the data that was used to train the model. To maintain per- The Dunning-Kruger effect is a cognitive bias that leads to
formance over time, the models should be constantly moni- inflated self-assessments: less experienced people fail to
tored for deteriorating performance and may need to be realize their mistakes while experts tend to be more self-
subjected to periodical recalibration or retraining. Data set critical and aware of their limitations.100 As AI becomes
shift may occur due to the technical differences between integrated into everyday activities, it is important to realize
institutions as well as variations in local clinical practices.97 that the Dunning-Kruger effect is not automatically fixed by
Thus, it can be challenging to implement an ML model technology. Like humans, AI systems do not always under-
in a different setting from where it was originally trained stand when they are wrong and can lead to devastating
but may be mitigated by site-specific training to adapt the consequences. Possible solutions include having humans
existing model to a new study population. integrated into AI systems, allowing one to identify those
potential failures (human-in-the-loop system) or creating
flexible systems that can adapt and handle outlier events.
Two Exemplars A common cautionary tale of AI is an AI system whose
We describe 2 examples of landmark ML studies whose optimum solution to achieving its task results in wiping
real-world clinical applications were underwhelming. One out humanity (this is illustrated by the canonical thought
of the biggest advances in ML has been in improving the experiment, the paperclip maximizer13). In a similar vein,
workflow of clinicians with predictive models for diabetic it is theoretically possible that an AI system tasked with
retinopathy, a leading cause of vision loss around the eradicating hypertension in the population may decide
world.39 A deep learning system for diabetic retinopathy that the most optimum solution to achieving this is to
created by Google Health failed to meet performance remove the human race. More realistic dangers are very
expectations when tested in a real-world setting across likely due to the unintended consequences of ML algo-
eleven clinics in Thailand.98 Despite high theoretical accu- rithms interacting with unpredictable humans. The Boe-
racy, the tool proved impractical in real-world testing, ing 737 Max tragedy highlights a simple oversight where

Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106 April 2, 2021   1115

 Padmanabhan et al Artificial Intelligence in Hypertension
HYPERTENSION COMPENDIUM

human override of the ML algorithm was not provided. 5. Padmanabhan S, Aman A, Dominiczak AF. Recent findings in the genetics
of blood pressure: how to apply in practice or is a moonshot required? Curr
Here, the AI system believed the plane was going up, Hypertens Rep. 2018;20:54. doi: 10.1007/s11906-018-0863-1
while the pilots could see it was going down but could 6. GBD 2015 Risk Factors Collaborators. Global, regional, and national com-
not override it, resulting in disastrous consequences. parative risk assessment of 79 behavioural, environmental and occupational,
and metabolic risks or clusters of risks, 1990-2015: a systematic analysis
Similarly, if there are no clear strategies for a physician or for the global burden of disease study 2015. Lancet. 2016;388:1659–1724.
health care provider to override a recommendation from 7. Padmanabhan S, Dominiczak AF. Genomics of hypertension: the road to
an ML algorithm or if such an override is not provided, precision medicine. Accessed November 20, 2020. https://www.nature.
com/articles/s41569-020-00466-4
then there is potential for harm. The risks of ML sys- 8. Loscalzo J. Precision medicine. Circ Res. 2019;124:987–989. doi: 10.1161/
tems in the health care world are real but very likely to be CIRCRESAHA.119.314403
9. Arnett DK, Blumenthal RS, Albert MA, Buroker AB, Goldberger ZD, Hahn
due to poor algorithm development and implementation
EJ, Himmelfarb CD, Khera A, Lloyd-Jones D, McEvoy JW, et al. 2019 ACC/
rather than due to a far-fetched malign AI system. AHA guideline on the primary prevention of cardiovascular disease: execu-
tive summary: a report of the American College of Cardiology/American
Heart Association Task Force on Clinical Practice Guidelines. Circulation.
2019;140:e563–e595. doi: 10.1161/CIR.0000000000000677
CONCLUSIONS 10. Williams B, Mancia G, Spiering W, Agabiti Rosei E, Azizi M, Burnier M, Clement
DL, Coca A, de Simone G, Dominiczak A, et al; ESC Scientific Document
AI and ML hold great promise in the future of health care. Group. 2018 ESC/ESH guidelines for the management of arterial hyperten-
Their clinical implementation is still evolving with some sion. Eur Heart J. 2018;39:3021–3104. doi: 10.1093/eurheartj/ehy339
aspects more advanced than others. The advantage of 11. Kallioinen N, Hill A, Horswill MS, Ward HE, Watson MO. Sources of inac-
curacy in the measurement of adult patients’ resting blood pressure in
using AI and ML is the pace of acquiring expertise. There clinical settings: a systematic review. J Hypertens. 2017;35:421–441. doi:
are challenges to clinical implementation, and deep collabo- 10.1097/HJH.0000000000001197
ration between physicians and ML experts is critical for suc- 12. Turing AM. I.—Computing machinery and intelligence. Mind. 1950;LIX:
433–460.
cessful, safe, and reliable applications of AI in health care 13. Bostrom N. Superintelligence. Paths, Dangers, Strategies. OUP; 2016.
that are validated through prospective studies. Human-in- 14. Honavar V. Symbolic artificial intelligence and numeric artificial neural net-
the-loop systems may result in quicker implementation. It is works: towards a resolution of the dichotomy. In: Sun R, Bookman LA, eds.
Computational Architectures Integrating Neural and Symbolic Processes: A
important to reiterate that ML is about correlations, whereas Perspective on the State of the Art. Springer US; 1995:351–388.
policy recommendations are based on causal inference. 15. Hinton GE. Connectionist learning procedures. In: Kodratoff Y, Michalski RS,
Ensuring diversity in the training of ML algorithms is criti- eds. Machine Learning. Morgan Kaufmann; 1990:555–610.
16. Montani S, Striani M. Artificial intelligence in clinical decision support:
cally important, otherwise this will amplify and propagate a focused literature survey. Yearb Med Inform. 2019;28:120–127. doi:
any distortions that currently exist in clinical practice. 10.1055/s-0039-1677911
Downloaded from http://ahajournals.org by on August 11, 2023

17. Bright TJ, Wong A, Dhurjati R, Bristow E, Bastian L, Coeytaux RR, Samsa
G, Hasselblad V, Williams JW, Musty MD, et al. Effect of clinical decision-
support systems: a systematic review. Ann Intern Med. 2012;157:29–43.
ARTICLE INFORMATION doi: 10.7326/0003-4819-157-1-201207030-00450
Affiliation 18. Ray S. A quick review of machine learning algorithms. Presented at the 2019
International Conference on Machine Learning, Big Data, Cloud and Parallel
BHF Glasgow Cardiovascular Research Centre, Institute of Cardiovascular and
Computing (COMITCon), Faridabad, India, February 14-16, 2019, and pub-
Medical Sciences, University of Glasgow.
lished in Comp Sci. 2019:35–39. doi: 10.1109/COMITCon.2019.8862451
19. Collins GS, Moons KGM. Reporting of artificial intelligence prediction models.
Acknowledgments
Lancet. 2019;393:1577–1579. doi: 10.1016/S0140-6736(19)30037-6
We are grateful to Clea du Toit (University of Glasgow), Bina Joe (University of To- 20. Moons KG, Altman DG, Reitsma JB, Ioannidis JP, Macaskill P, Steyerberg
ledo, Ohio), Sandeep Reddy (Deakin University, Australia) and three anonymous EW, Vickers AJ, Ransohoff DF, Collins GS. Transparent reporting of a
reviewers for their comments on this article. multivariable prediction model for Individual Prognosis or Diagnosis (TRI-
POD): explanation and elaboration. Ann Intern Med. 2015;162:W1–73. doi:
Sources of Funding 10.7326/M14-0698
S. Padmanabhan is funded by the Medical Research Council (MR/M016560/1), 21. Ye C, Fu T, Hao S, Zhang Y, Wang O, Jin B, Xia M, Liu M, Zhou X, Wu Q,
the British Heart Foundation (PG/12/85/29925, CS/16/1/31878, and et al. Prediction of incident hypertension within the next year: prospective
RE/18/6/34217), Health Data Research UK (HDR-5012), and Chief Scientist study using statewide electronic health records and machine learning. J
Office, Scotland. A.F. Dominiczak acknowledges funding from UK Research and Med Internet Res. 2018;20:e22. doi: 10.2196/jmir.9268
Innovation Strength in Places Fund. 22. Kanegae H, Suzuki K, Fukatani K, Ito T, Harada N, Kario K. Highly pre-
cise risk prediction model for new-onset hypertension using artificial intel-
Disclosures ligence techniques. J Clin Hypertens (Greenwich). 2020;22:445–450. doi:
None. 10.1111/jch.13759
23. Maxwell A, Li R, Yang B, Weng H, Ou A, Hong H, Zhou Z, Gong P, Zhang
C. Deep learning architectures for multi-label classification of intel-
REFERENCES ligent health risk prediction. BMC Bioinformatics. 2017;18:523. doi:
10.1186/s12859-017-1898-z
1. Topol EJ. High-performance medicine: the convergence of human and artificial 24. LaFreniere D, Zulkernine F, Barber D, Martin K. Using machine learning
intelligence. Nat Med. 2019;25:44–56. doi: 10.1038/s41591-018-0300-7 to predict hypertension from a clinical dataset. In: 2016 IEEE Symposium
2. Matheny ME, Whicher D, Thadaney Israni S. Artificial intelligence in Series on Computational Intelligence (SSCI). 2016:1–7.
health care: a report from the National Academy of Medicine. JAMA. 25. Poplin R, Varadarajan AV, Blumer K, Liu Y, McConnell MV, Corrado GS, Peng
2020;323:509–510. doi: 10.1001/jama.2019.21579 L, Webster DR. Prediction of cardiovascular risk factors from retinal fun-
3. Dzau VJ, Balatbat CA. Future of hypertension. Hypertension. 2019;74:450– dus photographs via deep learning. Nat Biomed Eng. 2018;2:158–164. doi:
457. doi: 10.1161/HYPERTENSIONAHA.119.13437 10.1038/s41551-018-0195-0
4. Mueller FB. AI (Artificial Intelligence) and hypertension research. Curr 26. Duan T, Rajpurkar P, Laird D, Ng AY, Basu S. Clinical value of pre-
Hypertens Rep. 2020;22:70. doi: 10.1007/s11906-020-01068-8 dicting individual treatment effects for intensive blood pressure

1116   April 2, 2021 Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106

 Padmanabhan et al Artificial Intelligence in Hypertension

HYPERTENSION COMPENDIUM
therapy. Circ Cardiovasc Qual Outcomes. 2019;12:e005010. doi: 10.1161/ 45. Pfohl S, Marafino B, Coulet A, Rodriguez F, Palaniappan L, Shah NH. Creat-
CIRCOUTCOMES.118.005010 ing fair models of atherosclerotic cardiovascular disease risk. Presented at
27. Mezzatesta S, Torino C, Meo P, Fiumara G, Vilasi A. A machine learning- the Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and
based approach for predicting the outbreak of cardiovascular diseases in Society, Honolulu, Hawaii, January 27-28, 2019, and published in AIES.
patients on dialysis. Comput Methods Programs Biomed. 2019;177:9–15. 2019;271-278. doi: 10.1145/3306618.3314278
doi: 10.1016/j.cmpb.2019.05.005 46. Arbabshirani MR, Fornwalt BK, Mongelluzzo GJ, Suever JD, Geise BD,
28. Li Y-H, Zhang G-G, Wang N. Systematic characterization and prediction of Patel AA, Moore GJ. Advanced machine learning in action: identifica-
human hypertension genes. Hypertension. 2017;69:349–355. tion of intracranial hemorrhage on computed tomography scans of the
29. Yakovlev A, Metsker O, Kovalchuk S, Bologova E. Prediction of in-hospital head with clinical workflow integration. NPJ Digit Med. 2018;1:9. doi:
mortality and length of stay in acute coronary syndrome patients using 10.1038/s41746-017-0015-z
machine-learning methods. J Am Coll Cardiol. 2018;71:A242. 47. Moradi M, Madani A, Karargyris A, Syeda-Mahmood TF. Chest x-ray gen-
30. Frizzell JD, Liang L, Schulte PJ, Yancy CW, Heidenreich PA, Hernandez AF, eration and data augmentation for cardiovascular abnormality classifica-
Bhatt DL, Fonarow GC, Laskey WK. Prediction of 30-day all-cause read- tion. Presented at the Medical Imaging Conference 2018, Houston, Texas,
missions in patients hospitalized for heart failure: comparison of machine February 10-15, 2018, and published in abstract form. SPIE. 2018. doi.
learning and other statistical approaches. JAMA Cardiol. 2017;2:204–209. org/10.1117/12.2293971
doi: 10.1001/jamacardio.2016.3956 48. Khalid SG, Zhang J, Chen F, Zheng D. Blood pressure estimation using
31. Ivanov I, Lomaev Y, Barkovskaya A. Automatic calculation of left ventricular photoplethysmography only: comparison between different machine learn-
volume in magnetic resonance imaging using an image-based clustering ing approaches. J Healthc Eng. 2018;2018:1548647. doi: 10.1155/
approach. IOP Conf Ser Mater Sci Eng. 2019;537:042046. 2018/1548647
32. Wu X, Yuan X, Wang W, Liu K, Qin Y, Sun X, Ma W, Zou Y, Zhang H, Zhou X, 49. Hu F, Warren J, Exeter DJ. Predicting lipid-lowering medication persistence
et al. Value of a machine learning approach for predicting clinical outcomes after the first cardiovascular disease hospitalization. Methods Inform Med.
in young patients with hypertension. Hypertension. 2020;75:1271–1278. 2020;59:61–74.
doi: 10.1161/HYPERTENSIONAHA.119.13404 50. Ambale-Venkatesh B, Yang X, Wu CO, Liu K, Hundley WG,
33. Motwani M, Dey D, Berman DS, Germano G, Achenbach S, Al-Mallah MH, McClelland R, Gomes AS, Folsom AR, Shea S, Guallar E, et al. Car-
Andreini D, Budoff MJ, Cademartiri F, Callister TQ, et al. Machine learn- diovascular event prediction by machine learning: the multi-ethnic
ing for prediction of all-cause mortality in patients with suspected coronary study of atherosclerosis. Circ Res. 2017;121:1092–1101. doi:
artery disease: a 5-year multicentre prospective registry analysis. Eur Heart 10.1161/CIRCRESAHA.117.311312
J. 2017;38:500–507. 51. Samad MD, Ulloa A, Wehner GJ, Jing L, Hartzel D, Good CW, Williams BA,
34. Al’Aref SJ, Maliakal G, Singh G, van Rosendael AR, Ma X, Xu Z, Alawamlh Haggerty CM, Fornwalt BK. Predicting survival from large echocardiog-
OAH, Lee B, Pandey M, Achenbach S, et al. Machine learning of clinical vari- raphy and electronic health record datasets. JACC Cardiovasc Imaging.
ables and coronary artery calcium scoring for the prediction of obstructive 2019;12:681–689.
coronary artery disease on coronary computed tomography angiography: 52. Segar MW, Vaduganathan M, Patel KV, McGuire DK, Butler J, Fonarow GC,
analysis from the CONFIRM registry. Eur Heart J. 2020;41:359–367. doi: Basit M, Kannan V, Grodin JL, Everett B, et al. Machine learning to predict
10.1093/eurheartj/ehz565 the risk of incident heart failure hospitalization among patients with diabe-
35. Wallert J, Tomasoni M, Madison G, Held C. Predicting two-year survival ver- tes: the WATCH-DM risk score. Diabetes Care. 2019;42:2298–2306. doi:
sus non-survival after first myocardial infarction using machine learning and 10.2337/dc19-0587
Swedish national register data. BMC Med Inform Decis Mak. 2017;17:99. 53. Dogan M, Beach S, Simons R, Lendasse A, Penaluna B, Philibert R. Blood-
doi: 10.1186/s12911-017-0500-y based biomarkers for predicting the risk for five-year incident coronary
Downloaded from http://ahajournals.org by on August 11, 2023

36. Huang C, Murugiah K, Mahajan S, Li SX, Dhruva SS, Haimovich JS, Wang heart disease in the Framingham Heart Study via machine learning. Genes.
Y, Schulz WL, Testani JM, Wilson FP, et al. Enhancing the prediction of 2018;9:641.
acute kidney injury risk after percutaneous coronary intervention using 54. Alaa AM, Bolton T, Di Angelantonio E, Rudd JHF, van der Schaar M.
machine learning techniques: a retrospective cohort study. PLoS Med. Cardiovascular disease risk prediction using automated machine learn-
2018;15:e1002703. doi: 10.1371/journal.pmed.1002703 ing: a prospective study of 423,604 UK Biobank participants. PLoS One.
37. Tabassian M, Sunderji I, Erdei T, Sanchez-Martinez S, Degiovanni A, Marino 2019;14:e0213653. doi: 10.1371/journal.pone.0213653
P, Fraser AG, D’hooge J. Diagnosis of heart failure with preserved ejec- 55. Alghamdi M, Al-Mallah M, Keteyian S, Brawner C, Ehrman J, Sakr S. Pre-
tion fraction: machine learning of spatiotemporal variations in left ven- dicting diabetes mellitus using SMOTE and ensemble machine learn-
tricular deformation. J Am Soc Echocardiogr. 2018;31:1272–1284.e9. doi: ing approach: the Henry Ford ExercIse Testing (FIT) project. PLoS One.
10.1016/j.echo.2018.07.013 2017;12:e0179805. doi: 10.1371/journal.pone.0179805
38. Tomašev N, Glorot X, Rae JW, Zielinski M, Askham H, Saraiva A, Mottram A, 56. Ebrahimzadeh E, Kalantari M, Joulani M, Shahraki RS, Fayaz F, Ahmadi
Meyer C, Ravuri S, Protsyuk I, et al. A clinically applicable approach to con- F. Prediction of paroxysmal atrial fibrillation: a machine learning based
tinuous prediction of future acute kidney injury. Nature. 2019;572:116–119. approach using combined feature vector and mixture of expert classifica-
doi: 10.1038/s41586-019-1390-1 tion on HRV signal. Comput Methods Programs Biomed. 2018;165:53–67.
39. Ting DSW, Cheung CY, Lim G, Tan GSW, Quang ND, Gan A, Hamzah H, 57. Koshimizu H, Kojima R, Kario K, Okuno Y. Prediction of blood pressure vari-
Garcia-Franco R, San Yeo IY, Lee SY, et al. Development and validation of ability using deep neural networks. Int J Med Inform. 2020;136:104067. doi:
a deep learning system for diabetic retinopathy and related eye diseases 10.1016/j.ijmedinf.2019.104067
using retinal images from multiethnic populations with diabetes. JAMA. 58. Whelton PK, Carey RM, Aronow WS, Casey DE Jr, Collins KJ,
2017;318:2211–2223. doi: 10.1001/jama.2017.18152 Dennison Himmelfarb C, DePalma SM, Gidding S, Jamerson KA, Jones DW,
40. Koren G, Nordon G, Radinsky K, Shalev V. Machine learning of big data in et al. 2017 ACC/AHA/AAPA/ABC/ACPM/AGS/APhA/ASH/ASPC/
gaining insight into successful treatment of hypertension. Pharmacol Res NMA/PCNA guideline for the prevention, detection, evaluation, and man-
Perspect. 2018;6:e00396. doi: 10.1002/prp2.396 agement of high blood pressure in adults: executive summary: a report
41. Weng SF, Reps J, Kai J, Garibaldi JM, Qureshi N. Can machine-learning of the American College of Cardiology/American Heart Association Task
improve cardiovascular risk prediction using routine clinical data? PLoS Force on Clinical Practice Guidelines. Circulation. 2018;138:e426–e483.
One. 2017;12:e0174944. doi: 10.1371/journal.pone.0174944 doi: 10.1161/CIR.0000000000000597
42. Acharya UR, Fujita H, Oh SL, Hagiwara Y, Tan JH, Adam M, Tan RS. Deep 59. Liu JE, Roman MJ, Pini R, Schwartz JE, Pickering TG, Devereux RB. Car-
convolutional neural network for the automated diagnosis of congestive diac and arterial target organ damage in adults with elevated ambulatory
heart failure using ECG signals. Appl Intell. 2018;49:16–27. and normal office blood pressure. Ann Intern Med. 1999;131:564–572. doi:
43. Zhao J, Feng Q, Wu P, Lupu RA, Wilke RA, Wells QS, Denny JC, Wei W-Q. 10.7326/0003-4819-131-8-199910190-00003
Learning from longitudinal data in electronic health record and genetic data 60. Mancia G, Facchetti R, Bombelli M, Grassi G, Sega R. Long-term risk of
to improve cardiovascular event prediction. Sci Rep. 2019;9:717. mortality associated with selective and combined elevation in office, home,
44. Elhaj FA, Salim N, Ahmed T, Harris AR, Swee TT. Hybrid classification and ambulatory blood pressure. Hypertension. 2006;47:846–853. doi:
of bayesian and extreme learning machine for heartbeat classification 10.1161/01.HYP.0000215363.69793.bb
of arrhythmia detection. Presented at the 2017 6th ICT International 61. Booth JN 3rd, Diaz KM, Seals SR, Sims M, Ravenell J, Muntner P, Shimbo
Student Project Conference (ICT-ISPC), Johor, Malaysia, May 23-24, D. Masked hypertension and cardiovascular disease events in a prospective
2017, and published in abstract form in IEEE. 2017;1-4. doi: 10.1109/ cohort of blacks: the Jackson Heart Study. Hypertension. 2016;68:501–
ICT-ISPC.2017.8075320 510. doi: 10.1161/HYPERTENSIONAHA.116.07553

Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106 April 2, 2021   1117

 Padmanabhan et al Artificial Intelligence in Hypertension
HYPERTENSION COMPENDIUM

62. Martín-Rioboó E, Pérula de Torres LA, Banegas JR, Lobos-Bejarano JM, 81. Liu Y, Li Z, Xiong H, Gao X, Wu J. Understanding of internal clustering vali-
Brotons Cuixart C, García Criado EI, Martin-Carrillo P, Martín-Rabadán Muro dation measures. Presented at the 2010 IEEE International Conference on
M, Egocheaga Cabello MI, Maiques Galán A; MAMPA, PAPPS Study Groups. Data Mining, Sydney, Australia, December 13-17, 2010, and published in
Knowledge, availability, and use of ambulatory and home blood pres- abstract form. IEEE. 2010:911–916. doi: 10.1109/ICDM.2010.35.
sure monitoring in primary care in Spain: the MAMPA study. J Hypertens. 82. Nagendran M, Chen Y, Lovejoy CA, Gordon AC, Komorowski M, Harvey
2018;36:1051–1058. doi: 10.1097/HJH.0000000000001673 H, Topol EJ, Ioannidis JPA, Collins GS, Maruthappu M. Artificial intel-
63. Kronish IM, Kent S, Moise N, Shimbo D, Safford MM, Kynerd RE, O’Beirne ligence versus clinicians: systematic review of design, reporting stan-
R, Sullivan A, Muntner P. Barriers to conducting ambulatory and home blood dards, and claims of deep learning studies. BMJ. 2020;368:m689. doi:
pressure monitoring during hypertension screening in the United States. J Am 10.1136/bmj.m689
Soc Hypertens. 2017;11:573–580. doi: 10.1016/j.jash.2017.06.012 83. Yusuf M, Atal I, Li J, Smith P, Ravaud P, Fergie M, Callaghan M, Selfe J.
64. Stergiou GS, Kollias A, Destounis A, Tzamouranis D. Automated blood pres- Reporting quality of studies using machine learning models for medi-
sure measurement in atrial fibrillation: a systematic review and meta-analysis. cal diagnosis: a systematic review. BMJ Open. 2020;10:e034568. doi:
J Hypertens. 2013;31:215–216. doi: 10.1097/HJH.0b013e32835ae9b3 10.1136/bmjopen-2019-034568
65. van der Steen MS, Lenders JW, Thien T. Side effects of ambulatory 84. Chan AW, Tetzlaff JM, Altman DG, Laupacis A, Gøtzsche PC,
blood pressure monitoring. Blood Press Monit. 2005;10:151–155. doi: Krleža-Jerić K, Hróbjartsson A, Mann H, Dickersin K, Berlin JA, et al.
10.1097/00126097-200506000-00007 SPIRIT 2013 statement: defining standard protocol items for clini-
66. Diao X, Huo Y, Yan Z, Wang H, Yuan J, Wang Y, Cai J, Zhao W. An applica- cal trials. Ann Intern Med. 2013;158:200–207. doi: 10.7326/0003-
tion of machine learning to etiological diagnosis of secondary hyperten- 4819-158-3-201302050-00583
sion: retrospective study using electronic medical records. JMIR Med Inform. 85. Schulz KF, Altman DG, Moher D; CONSORT Group. CONSORT 2010
2021;9:e19739. doi: 10.2196/19739 statement: updated guidelines for reporting parallel group random-
67. Quang D, Chen Y, Xie X. DANN: a deep learning approach for annotating ized trials. Ann Intern Med. 2010;152:726–732. doi: 10.7326/0003-
the pathogenicity of genetic variants. Bioinformatics. 2015;31:761–763. doi: 4819-152-11-201006010-00232
10.1093/bioinformatics/btu703 86. Richens JG, Lee CM, Johri S. Improving the accuracy of medical diag-
68. Zhou J, Troyanskaya OG. Predicting effects of noncoding variants with deep nosis with causal machine learning. Nat Commun. 2020;11:3923. doi:
learning-based sequence model. Nat Methods. 2015;12:931–934. doi: 10.1038/s41467-020-17419-7
10.1038/nmeth.3547 87. Barredo Arrieta A, Díaz-Rodríguez N, Del Ser J, Bennetot A, Tabik S,
69. Zhou J, Theesfeld CL, Yao K, Chen KM, Wong AK, Troyanskaya OG. Deep Barbado A, Garcia S, Gil-Lopez S, Molina D, Benjamins R, et al. Explainable
learning sequence-based ab initio prediction of variant effects on expres- artificial intelligence (XAI): concepts, taxonomies, opportunities and chal-
sion and disease risk. Nat Genet. 2018;50:1171–1179. doi: 10.1038/ lenges toward responsible AI. Inform Fusion. 2020;58:82–115.
s41588-018-0160-6 88. Rudin C. Stop explaining black box machine learning models for high
70. Sundaram L, Gao H, Padigepati SR, McRae JF, Li Y, Kosmicki JA, Fritzilas N, stakes decisions and use interpretable models instead. Nat Mach Intell.
Hakenberg J, Dutta A, Shon J, et al. Predicting the clinical impact of human 2019;1:206–215.
mutation with deep neural networks. Nat Genet. 2018;50:1161–1170. doi: 89. Guidotti R, Monreale A, Giannotti F, Pedreschi D, Ruggieri S, Turini F.
10.1038/s41588-018-0167-z Factual and counterfactual explanations for black box decision making.
71. Huan T, Meng Q, Saleh MA, Norlander AE, Joehanes R, Zhu J, Chen BH, IEEE Intell Syst. 2019;34:14–23.
Zhang B, Johnson AD, Ying S, et al; International Consortium for Blood 90. De Fauw J, Ledsam JR, Romera-Paredes B, Nikolov S, Tomasev N,
Pressure GWAS (ICBP). Integrative network analysis reveals molecular Blackwell S, Askham H, Glorot X, O’Donoghue B, Visentin D, et al. Clinically
mechanisms of blood pressure regulation. Mol Syst Biol. 2015;11:799. doi: applicable deep learning for diagnosis and referral in retinal disease. Nat
Downloaded from http://ahajournals.org by on August 11, 2023

10.15252/msb.20145399 Med. 2018;24:1342–1350. doi: 10.1038/s41591-018-0107-6

72. Alimadadi A, Manandhar I, Aryal S, Munroe PB, Joe B, Cheng X. Machine learn- 91. Crawford K, Calo R. There is a blind spot in AI research. Nature.
ing-based classification and diagnosis of clinical cardiomyopathies. Physiol 2016;538:311–313. doi: 10.1038/538311a
Genomics. 2020;52:391–400. doi: 10.1152/physiolgenomics.00063.2020 92. Chen IY, Johansson FD, Sontag D. Why is my classifier discriminatory?
73. Aryal S, Alimadadi A, Manandhar I, Joe B, Cheng X. Machine learn- Presented at the Proceedings of the 32nd International Conference on
ing strategy for gut microbiome-based diagnostic screening of cardio- Neural Information Processing Systems, Montréal, Canada, December 2-8,
vascular disease. Hypertension. 2020;76:1555–1562. doi: 10.1161/ 2018, and published in abstract form. NeurIPS. 2018:3543–3554.
HYPERTENSIONAHA.120.15885 93. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S.
74. Elshawi R, Al-Mallah MH, Sakr S. On the interpretability of machine learn- Dermatologist-level classification of skin cancer with deep neural net-
ing-based model for predicting hypertension. BMC Med Inform Decis Mak. works. Nature. 2017;542:115–118. doi: 10.1038/nature21056
2019;19:146. doi: 10.1186/s12911-019-0874-0 94. Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial
75. Emanuel EJ, Wachter RM. Artificial intelligence in health care: will the bias in an algorithm used to manage the health of populations. Science.
value match the hype? JAMA. 2019;321:2281–2282. doi: 10.1001/ 2019;366:447–453. doi: 10.1126/science.aax2342
jama.2019.4914 95. Verghese A, Shah NH, Harrington RA. What this computer needs is a phy-
76. Pearl, Judea. Theoretical Impediments to Machine Learning With Seven sician: humanism and artificial intelligence. JAMA. 2018;319:19–20. doi:
Sparks from the Causal Revolution. Presented at the WSDM '18: Proceed- 10.1001/jama.2017.19198
ings of the Eleventh ACM International Conference on Web Search and 96. Char DS, Shah NH, Magnus D. Implementing machine learning in health
Data Mining, Los Angeles, California, February 5-9, 2018, and published in care - addressing ethical challenges. N Engl J Med. 2018;378:981–983.
abstract form. ACM. 2018:3. doi: 10.1145/3159652.3176182 doi: 10.1056/NEJMp1714229
77. Cruz Rivera S, Liu X, Chan AW, Denniston AK, Calvert MJ; SPIRIT-AI 97. Kelly CJ, Karthikesalingam A, Suleyman M, Corrado G, King D. Key chal-
and CONSORT-AI Working Group; SPIRIT-AI and CONSORT-AI Steer- lenges for delivering clinical impact with artificial intelligence. BMC Med.
ing Group; SPIRIT-AI and CONSORT-AI Consensus Group. Guidelines 2019;17:195. doi: 10.1186/s12916-019-1426-2
for clinical trial protocols for interventions involving artificial intelligence: 98. Beede E, Baylor E, Hersch F, Iurchenko A, Wilcox L, Ruamviboonsuk P,
the SPIRIT-AI extension. Nat Med. 2020;26:1351–1363. doi: 10.1038/ Vardoulakis LM. A human-centered evaluation of a deep learning system
s41591-020-1037-7 deployed in clinics for the detection of diabetic retinopathy. Presented
78. Liu X, Cruz Rivera S, Moher D, Calvert MJ, Denniston AK; SPIRIT-AI and at the Proceedings of the 2020 Chi Conference on Human Factors in
CONSORT-AI Working Group. Reporting guidelines for clinical trial reports Computing Systems. Association for Computing Machinery, Honolulu,
for interventions involving artificial intelligence: the CONSORT-AI extension. Hawaii, April 25-30, 2020, and published in abstract form. ACM. 2020:1–
Nat Med. 2020;26:1364–1374. doi: 10.1038/s41591-020-1034-x 12. doi: 10.1145/3313831.3376718
79. Handelman GS, Kok HK, Chandra RV, Razavi AH, Huang S, Brooks M, Lee 99. Connell A, Montgomery H, Martin P, Nightingale C, Sadeghi-Alavijeh O,
MJ, Asadi H. Peering into the black box of artificial intelligence: evaluation King D, Karthikesalingam A, Hughes C, Back T, Ayoub K, et al. Evaluation
metrics of machine learning methods. AJR Am J Roentgenol. 2019;212:38– of a digitally-enabled care pathway for acute kidney injury manage-
43. doi: 10.2214/AJR.18.20224 ment in hospital emergency admissions. NPJ Digit Med. 2019;2:67. doi:
80. Krittanawong C, Johnson KW, Rosenson RS, Wang Z, Aydar M, Baber U, 10.1038/s41746-019-0100-6
Min JK, Tang WHW, Halperin JL, Narayan SM. Deep learning for cardiovas- 100. Kruger J, Dunning D. Unskilled and unaware of it: how difficulties in recog-
cular medicine: a practical primer. Eur Heart J. 2019;40:2058–2073. doi: nizing one’s own incompetence lead to inflated self-assessments. J Pers
10.1093/eurheartj/ehz056 Soc Psychol. 1999;77:1121–1134. doi: 10.1037//0022-3514.77.6.1121

1118   April 2, 2021 Circulation Research. 2021;128:1100–1118. DOI: 10.1161/CIRCRESAHA.121.318106