TYPE Review

PUBLISHED 13 October 2022

DOI 10.3389/fcvm.2022.945726

Artificial intelligence in
OPEN ACCESS cardiology: Hope for the future
EDITED BY
Savvas Loizos,
Hygeia Hospital, Greece
and power for the present
REVIEWED BY
Rickey Carter, Loucia Karatzia1 , Nay Aung2,3 and Dunja Aksentijevic1*
Mayo Clinic Florida, United States
1
Jack J. Miller, Centre for Biochemical Pharmacology, William Harvey Research Institute, Barts and The London
Aarhus University, Denmark School of Medicine and Dentistry, Queen Mary University of London, London, United Kingdom,
2
Centre for Advanced Cardiovascular Imaging, William Harvey Research Institute, Barts
*CORRESPONDENCE and The London School of Medicine and Dentistry, Queen Mary University of London, London,
Dunja Aksentijevic United Kingdom, 3 National Institute for Health and Care Research (NIHR) Barts Biomedical
[email protected] Research Centre, William Harvey Research Institute, Barts and The London School of Medicine
SPECIALTY SECTION and Dentistry, Queen Mary University of London, London, United Kingdom
This article was submitted to
Cardiovascular Imaging,
a section of the journal
Frontiers in Cardiovascular Medicine
Cardiovascular disease (CVD) is the principal cause of mortality and morbidity
RECEIVED 16 May 2022
ACCEPTED 09 September 2022
globally. With the pressures for improved care and translation of the latest
PUBLISHED 13 October 2022 medical advances and knowledge to an actionable plan, clinical decision-
CITATION making for cardiologists is challenging. Artificial Intelligence (AI) is a field in
Karatzia L, Aung N and Aksentijevic D
computer science that studies the design of intelligent agents which take the
(2022) Artificial intelligence
in cardiology: Hope for the future best feasible action in a situation. It incorporates the use of computational
and power for the present. algorithms which simulate and perform tasks that traditionally require human
Front. Cardiovasc. Med. 9:945726.
doi: 10.3389/fcvm.2022.945726 intelligence such as problem solving and learning. Whilst medicine is arguably
COPYRIGHT the last to apply AI in its everyday routine, cardiology is at the forefront
© 2022 Karatzia, Aung and of AI revolution in the medical field. The development of AI methods for
Aksentijevic. This is an open-access
article distributed under the terms of
accurate prediction of CVD outcomes, non-invasive diagnosis of coronary
the Creative Commons Attribution artery disease (CAD), detection of malignant arrythmias through wearables,
License (CC BY). The use, distribution
and diagnosis, treatment strategies and prediction of outcomes for heart
or reproduction in other forums is
permitted, provided the original failure (HF) patients, demonstrates the potential of AI in future cardiology.
author(s) and the copyright owner(s) With the advancements of AI, Internet of Things (IoT) and the promotion of
are credited and that the original
publication in this journal is cited, in precision medicine, the future of cardiology will be heavily based on these
accordance with accepted academic innovative digital technologies. Despite this, ethical dilemmas regarding the
practice. No use, distribution or
reproduction is permitted which does implementation of AI technologies in real-world are still unaddressed.
not comply with these terms.
KEYWORDS

artificial intelligence, cardiology, machine learning, cardiac imaging, cardiac MR

(CMR), cardiovascular diagnostic

Highlights
- Artificial intelligence is a computer science field that studies the problem of building
agents which take the best possible course of action in a specific situation.
- Cardiology is at the forefront of artificial intelligence revolution in medicine.

Frontiers in Cardiovascular Medicine 01 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

- AI allows accurate prediction of cardiovascular outcomes, intelligent if a human were so behaving” (5). In 1968, the
non-invasive diagnosis of coronary artery disease, detection cognitive scientist Marvin Minsky outlined AI as the “science
of malignant arrhythmias and diagnosis, treatment, and of making machines do things that would require intelligence if
prediction of outcomes for heart failure patients. one by men” (6). More recently, in his published paper “What
- The advancements of artificial intelligence, Internet of Things, is Artificial Intelligence,” John McCarthy gives the following
and precision medicine will lead to future innovations in the explanation regarding AI: “It is the science and engineering
field of cardiovascular research. of making intelligent machines, especially intelligent computer
- Artificial intelligence in cardiology is limited by ethical and programs. It is related to the similar task of using computers to
data privacy concerns, which are still to be addressed. understand human intelligence, but AI does not have to confine
- Regulations are required to be implemented for the safe use of itself to methods that are biologically observable” (7). Kaplan
artificial intelligence in cardiology and medicine in the future. and Haenlein in 2019 summarised AI as “a system’s ability
to interpret data correctly, to learn from such data, and to
use those learnings to achieve specific goals and tasks through
flexible learning” (3). Stuart Russell and Peter Norvig, authors
Introduction of “Artificial Intelligence: A Modern Approach,” have defined
AI by dividing it into four goal-based categories (Table 1; 8).
Cardiovascular disease (CVD) is the principal cause of The definitions are laid out in two scopes. One dimension is
mortality and morbidity globally. The diagnosis and treatment whether the goal of AI is to match human performance or ideal
of CVD relies on data in several forms, such as patient rationality. The other aspect is whether the goal is to build
history, physical examination, laboratory data, invasive and systems that think or systems that act (9).
non-invasive imaging techniques. With the pressures for Traditionally, statistics has been the standard method used
improved care and translation of the latest medical advances in medical research to show the benefit of new treatments,
and knowledge to an actionable plan, clinical decision-making identify risk factors for a disease and predict prognosis.
for cardiologists is challenging (1). On the other hand, emerging Traditional medical research proposes a hypothesis, which
new technologies and the growth of artificial intelligence (AI) is then tested with statistical analysis. Statistics analyses a
and machine learning (ML) in the last few decades have offered given dataset using mathematical equations and discovers
physicians opportunities to conduct more efficient and data- relationships between data points and outcomes. It is focused
driven research. The availability of large-volume data from on validation of the hypothesis and understating of the causality
electronic health records (EHRs), mobile health devices and and the mechanisms (10). AI is data driven and does not require
imaging data enables the rapid development of AI algorithms in the formulation of a hypothesis. It makes predictions with high
medicine. Cardiology has been one of the few medical specialties accuracy, without needing to interpret the data given (1). Its goal
in which AI technologies have been examined systematically (2). is to identify hidden patterns in the data and predict new data.
AI is able to use very complex nonparametric models from a
vast amount of data in comparison to simple parametric models
Artificial intelligence requiring a suitable-sized data set used in statistics (10).

It is difficult to determine the exact year that AI was born.

The English mathematician Alan Turing, named by some as Machine learning
the father of AI, developed the famous code breaking machine
The Bombe for the British government, which broke the Enigma Machine learning (ML) is a subfield of AI (Figure 1).
code, used by the German army in the Second World War. ML allows a system to learn from data rather than through
The Bombe was considered the first working electro-mechanical explicit programming. It uses algorithms that learn from
computer (3). In 1950, inspired by his achievement, Turing data, identify specific patterns, and make decisions/predict
published his article “Computing Machinery and Intelligence,” outcomes based on the learned model (10). Once the ML
where he proposed the famous question “Can machines think?” algorithm is trained with data, the ML model will be provided
and recommended definitions for the terms machine and think with an input. The output will be a predictive model,
(4). He also outlined the world known Turing Test–which is based on the data that trained the model (11). In ML,
considered today as the standard method to identify intelligence the machine learns from the data through three different
of an artificial system. According to the Turing Test, if a methods: supervised, unsupervised, and reinforcement learning
human is interacting with another human and a machine and (Figure 2).
cannot distinguish the machine from the human, then the Supervised learning is the most widely used technique today.
machine is considered to be intelligent (3). In 1955, during the The model is provided both the inputs and the outputs at the
Dartmouth Research Project, AI was defined as the problem time of training. It then goes on to produce a prediction model
of “making a machine behave in ways that would be called through categorising future events or finding which variables

Frontiers in Cardiovascular Medicine 02 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

TABLE 1 Definitions of Artificial Intelligence (AI) by four goal based categories (8).

Acting humanly Acting rationally

“The art of creating machines that perform functions that require intelligence “Computational intelligence is the study of the design of intelligent agents.” (108)
when performed by people.” (107)

Thinking humanly Thinking rationally

“The exciting new effort to make computers think. . . machines with minds, in “The study of mental faculties through the use of computational models.” (110)
the full and literal sense.” (109)

are most applicable to the outcome (1). Most commonly used the data states. This is used in cases where the amount of data
supervised learning tasks include classification (identification available is small for the purpose of training a model. Decision
of the group a new measurement belongs to) and regression tree algorithms use a branching like structure to map possible
(prediction of a continuous value of a new observation) (10). outcomes of a decision, with a percentage assigned to each
In unsupervised learning the model learns patterns from the node, depending on the chance of the outcome occurring (11).
input, without feedback provided from humans (8). Therefore, A random forest algorithm consists of the output of multiple
the algorithm needs to explore the data and find hidden patterns. decision trees, to reach a single result. Support vector machine
The most common tasks in unsupervised learning are clustering algorithms draw a boundary line which increases the margins
(dividing objects in groups with similar characteristics) and from each class and new observations are classified based on
dimensionality reduction (reducing the number of variables of this line (10). Ensemble learning is another ML method, in
data through keeping initial variables that explain the data), which multiple weak algorithms are combined to obtain a
which can also be applied in supervised learning (10). good prediction. Bagging, boosting, and stacking are the three
Reinforcement learning involves receiving an output approaches of ensemble learning. In bagging, multiple weak
variable to be amplified and a sequence of choices that can be learners (algorithms) are trained in parallel, and the results of
taken in order to influence the output (3). The model is not told each learner are combined to produce a final output. In boosting,
which actions to take, but it needs to discover which actions multiple weak learners are combined in series and trained
will lead to the highest level of reward, by trying them. Trial- consequently, considering errors from previous algorithms, in
and-error-search and delayed-reward are the two features that order to reduce bias. In stacking, the results of weak learners are
distinguish reinforcement learning (12). used as input for another ML algorithm (10).
Machine learning operates via a variety of algorithms,
which serve different tasks. Understanding the different types
Deep learning
of ML algorithms aids the researcher to choose the best type of
algorithm for their project. One of the most known examples is
Deep learning (DL) is a subfield of ML (Figure 1).
the Bayesian algorithm. It allows the scientist to encode previous
It is inspired by the complexity of the human brain in
beliefs of what a model should look like, irrespectively of what
handling data and generating patterns, used for decision
making (1). DL is comprised of deep neural networks.
A neural network has three or more layers: an input
layer, one or many hidden layers, and an output layer.
Data is consumed via the input layer. Then the hidden
layers extract the salient features from the input data to
produce an output that closely approximates the ground-
truth. A neural network may comprise of millions of simple
processing nodes which are firmly unified (11). In DL there
are multiple hidden tiers of artificial neural networks that can
create automated forecasts from training datasets. DL requires
complex data for training but is not required to extract features
from the input data. Once implemented it is self-directed,
thus eliminating manual human interaction. It can extract
important results from vast amount of data through iterative
FIGURE 1
Relationship between Artificial Intelligence (AI), Machine processing (10).
Learning (ML), and Deep Learning (DL). Neural networks can be used for various tasks such
as classification, clustering, dimensionality reduction, pattern

Frontiers in Cardiovascular Medicine 03 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

FIGURE 2
Overview of the role of AI in cardiovascular medicine. Abbreviations: EHRs, electronic health records; CMR, cardiac magnetic resonance; CT,
computed tomography; IoT, internet of things; SPECT, single photon-emission computerised tomography.

recognition, natural language processing, computer vision, and In Multiplayer Perceptron (MLP), input data travels through
predictive analysis (13). Neural networks consist of multiple various layers of artificial neurons. It is a fully connected neural
layers of interconnected artificial neurons. A neuron receives network, as all nodes are connected to all the neurons in the next
inputs multiplied with random weights, to which a bias layer. Input and output layers and multiple hidden layers (three
value is then added. An activation function is then applied or more) are present, and propagation is bi-directional (forward
and defines the final value to be given out of the neuron. and backward). MLP is used in speech recognition, machine
There are different types of activation functions, depending translation and complex classification. In a Recurrent Neural
on the input values (14). Neural networks are classified Network (RNN), the output of a layer becomes the input to the
depending on their structure, data flow, neurons used and
next layer—which is the only layer in the network. Therefore, the
their density. The most important types of neural networks
output of a layer becomes an input to itself and forms a feedback
involve:
loop. This means then network has internal memory, which
influences the current output. RNNs are used in tasks such as
1. Feed Forward Neural Network
text processing and text to speech processing. An example of
2. Multilayer Perceptron
this is when typing in Google, it automatically completes the
3. Radial Basis Function Neural Network
sentence for us! The Radial Basis Function Network (RBFN) is
4. Recurrent Neural Network
5. Modular Neural Network based on the radial basis function (activation function), which
6. Convolutional Neural Network is included in the hidden layer. The input is designated to
a centre and the output combines the outputs of the radial
Feed Forward Neural Networks (FFNNs) are the simplest basis function and weight parameters to perform classification
form of neural networks, as data travels in just one direction, or inference. It is used for prediction analysis and function
passing from input and exiting through output nodes (hidden approximation. A Modular Neural Network (MNN) is consisted
layers may or may not be present). They can be either of multiple different networks which work independently and
single-layered or multi-layered FFNNs and the number of perform different tasks, towards achieving the output. It is
layers depend on how complex the function is. FFNNs are usually used in stock market prediction systems and in cases of
usually applied in face recognition or simple classification. compression of high level input data (13).

Frontiers in Cardiovascular Medicine 04 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

FIGURE 3
Demonstration of a Convolutional Neural Network (CNN) architecture. A CNN is composed of several blocks which include convolutional
layers, pooling layers, and fully connected layers.

Convolutional neural networks When the features extracted by the convolution layers and sub-
sampled by the pooling layers are created, they are mapped
Convolutional Neural Networks (CNNs) are a group of by fully connected layers to the final outputs of the network,
deep neural networks, used in various fields including face same as the probability for each class in classification tasks. An
recognition, speech processing and computer vision. They are activation function is also applied to the last fully connected
a powerful tool in DL, as they necessitate minimal amount layer, depending on the task. For multiclass classification task,
of pre-processing information (15). The CNN’s architecture is the softmax function converts the output real values from the
inspired by neurons in human and animal brains. It consists last fully connected layer to aim class probabilities, ranging
of multiple stacks (blocks) of convolution layers and pooling from 0 to 1 and with a sum of 1 (17). Multiple different CNN
layers, followed by a fully connected layer and a normalising architectures have been created over the last few years. Model
layer (Figure 3). The convolutional layer is the most important architecture is vital in performance improvement of different
component of the CNN architecture. Its convolutional topology applications (Table 2).
allows CNNs to perform dimensionality reduction, effective The aim of a CNN’s training is to find kernels (filters)
automated feature extraction (in contrast to traditional in convolution layers and weights in fully connected layers
algorithms’ labour hand-crafted feature extraction) and perform which will reduce the difference between output predictions and
operations from 2D and 3D images. Its most important ground truth labels on a training dataset. The data available
characteristics, weight sharing (all neurons of neighbouring is split into a training set, a validation set and a test set. The
layers share the same weight) and local connectivity [neurons training set is used to train the network, in two phases. In
in one layer are connected to neurons in the next layer that are forward propagation, the input is passed completely through the
spatially close to them, thus keeping the ones that carry the most network, where loss values are calculated. In backpropagation,
important information (memory-effective)], make the CNN’s each layer will receive the gradient of loss with respect to
training process more simplified and efficient, as a small number its outputs and return the gradient of loss with respect to
of parameters is utilised with minimal human effort (16). its inputs, leading to update of the learnable parameters. The
A collection of filters (kernels), which are part of the validation set evaluates the model during the training process
convolutional layer, perform convolution (a sliding window and performs model selection. The test set is used at the end
across an input, creating one averaged output for each stride to evaluate the performance of the final model selected in the
the window takes) of the input image to generate output feature previous steps (17). The CNN’s architecture and function, makes
maps. The outputs are passed to an activation function. The it attractive for the purpose of solving classification problems
most common non-linear activation function currently used for in a high dimensional space. The reason for this is that CNNs
CNNs is ReLU. The large feature maps created by convolutional are universal, meaning they can be used to approximate any
operations, are then sub-sampled by the pooling layers. The continuous function to an arbitrary accuracy when the depth of
basic idea of down-sampling is that only the most important the neural network is large enough. CNNs and DL in general,
information of the feature map is maintained. In the fully work well in exceptionally large networks with a vast number of
connected layer, each neuron is connected to all neurons of the parameters, as they are surprisingly good at extrapolating when
previous layer, thus forming a fully connected neural network. fed data similar to what they were trained on (18).

Frontiers in Cardiovascular Medicine 05 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

TABLE 2 Most used Convolutional Neural Network (CNN) architectures.

Model Main task Dataset Error rate Input size Year

AlexNet Uses Dropout and ReLu ImageNet 16.4 227 × 227 × 3 2012
VGG Small filter size, increased depth ImageNet 7.3 224 × 224 × 3 2014
GoogleNet Different filter size, increased depth, block concept, concatenation concept ImageNet 6.7 224 × 224 × 3 2015
ResNet Robust against overfitting due to symmetry mapping-based skip links ImageNet 3.57 224 × 224 × 3 2016
DenseNet Blocks of layers—layers connected to each other CIFAR-10, CIFAR100, 3.46, 17.18, 5.54 224 × 224 × 3 2017
ImageNet

Generative adversarial networks Overfitting is one of the most important issues that need to
be addressed when building an AI model. Effectively overfitting
Generative adversarial networks (GANs) were introduced violates the principle of parsimony, a well-known law in
by Goodfellow et al. in, as a new framework for the creation statistical analysis. Parsimony highlights that a problem should
of synthetic data, which aim to mimic the real dataset (19). be stated in the simplest possible terms and explained with
GANs are an unsupervised learning algorithm and consist of the fewest assumptions possible. With overfitting, the model
two neural networks. The generator network generates new tries to fit the training data entirely and ends up memorising
examples, and the discriminator network evaluates whether the irrelevant data patterns, noise and random fluctuations and
generated examples belong to the real training dataset (classifies performs less well in a subsequent unseen dataset. This
them as real or fake). The two models are trained at the can be recognised in cases where the model performs well
same time, until the generator is generating realistic examples. on the training set but not on the validation set. Cross-
GANs have been mainly used to generate images and healthcare validation can detect overfitting by identifying how well the
records data in the medical field (20). For example, Amirrajab model can generalise to other datasets. Regular monitoring
et al. used a heart stimulator and combined it with a GAN to of the loss and accuracy on the training and validation
generate synthetic short-axis cine CMR images at multiple slice sets, can lead to early recognition of overfitting. Obtaining a
locations (21). In a different study, a GAN was used to convert larger training dataset, data augmentation and reduction of
CMR images to computed tomography, for better visualisation architectural complexity are some of the ways overfitting can be
of calcified structures, which are difficult to detect on CMR (22). mitigated (17).
GANs have shown to improve the performance of predictive Outliers are another important issue in AI applications.
models by filling the missing data. Che et al., added GAN- These are values that look different (are in the extremes) from
generated synthetic data to real patient data, leading to an other values in the data, may carry undue weight on their
improved CNN-based risk prediction model (23). final classification and therefore mislead the training process
and produce less accurate models. The ML algorithm should
be able to deal with outliers and similar technical problems
Limitations of artificial intelligence if challenged. Detection of outliers through visualisation
models techniques (i.e., box plot) and mathematical functions and
prevention of them via larger datasets, can solve this
Machine learning models are best trained and have higher problem (24).
accuracy when using big data. Big data is any kind of data When dealing with large datasets, it is important to be
source that has the following characteristics: exceptionally large aware of the risks when calculating the effect size and the
amount of data, the capability to transfer that data at a great statistical significance. Effect size is essentially the quantification
velocity of speed, a mounting range of data sources and validity of the size of difference between two groups. Statistical
so that data sources reflect the truth (11). A large dataset significance provides the likelihood that the difference between
allows for subsampling of the data for bootstrapping approaches two groups could be an accident of sampling and is calculated
(thus providing measures of robustness of an approach) and with the p-value. The main issue of statistical significance
computational reasons (a model structure can be developed is that the p-value depends on the size of the effect and
on a subset of a large dataset). In the case of a small DL the size of the sample. With larger datasets, everything
dataset, techniques such as data augmentation (modification of becomes statistically significant, even if practically is not
training data through random transformations so the model significant. Less emphasis should be given on p-values and more
does not see the same inputs during the training iterations) importance should be attributed to the effect size calculation
and transfer learning (features learned on a large dataset can be along with a margin error/confidence interval when involving
shared in a similar target dataset) can be applied, to train the big data, as larger datasets produce more accurate effect
model efficiently. size (25).

Frontiers in Cardiovascular Medicine 06 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

The widespread use of Big Data in the field of AI, does Electrocardiography
not come without challenges. As a dataset grows larger, it
can lead to class imbalance. The ML model’s performance The electrocardiogram (ECG) is considered the first-
decreases, as datasets include data from classes with various line non-invasive diagnostic investigation for the evaluation
probabilities of occurrence (26). Different methods have been of cardiovascular pathology. However, its interpretation can
applied to solve this issue, including down-sampling large be time-consuming and challenging at times. Automated
classes or up-sampling small classes, and constructing models ECG interpretation, via digital ECG machines nowadays is
for every hierarchical level (17). As the volume of data almost universal. Despite the significant progress made in
increases, variance and bias also increases. Variance involves the development of computerised interpretation of the ECG,
the consistency of a learner’s ability to predict random things, various limitations are still present, with systematic over-reading
whilst bias is the learner’s ability to learn the wrong thing. of the ECG deemed necessary (31). AI methods are increasingly
If bias is introduced in the data, generalisation of the data used with aim to improve the accuracy of automated ECG
is compromised. Regularisation techniques are well-established interpretation and aid patient stratification and prognostication.
methods in ML which improve generalisation (26). Modern ML models identify the P and T waves and the QRS
Adversarial attacks compromise the reliability and complexes and calculate parameters such as the heart rate (HR),
robustness of DL methods and their safe application in the cardiac axis, different interval lengths of a patient’s ECG,
medicine. They encompass mildly altered images, which ST-changes and common rhythm abnormalities such as atrial
resemble original images, but they are maliciously designed fibrillation (AF) (32). In a recent study, a 34-layer DNN (33
to confuse pre-trained models. This can lead to a completely convolutional layers followed by a linear output layer into a
different prediction for the image the neural network analyses. softmax) has been developed, which detects various arrythmias
Various methods are being proposed for defence against and outperforms a board-certified cardiologist in recall and
such attacks, but none has been proven safe enough yet precision (33). In another study, three neural networks—back
(27). Lastly, what a neural network considers meaningful propagation, self-organising map and radial basis function—
information for extraction from the data presented to it, were used to categorise ECG indicators of cardiac patients into
remains an unaddressed question. Attention mapping is a three states (normal, abnormal, and life-threatening) and were
scalar matrix, which aims to augment the significant image found to be accurate in 99% of test cases (34). Zhao et al., used
regions and suppress the irrelevant information in other a support vector machine (SVM) and identified five common
regions, with respect to the target task. It amplifies the arrythmias from ECG tracings of a large dataset. Sinus rhythm
importance of input variables in terms of their impact on (SR), left bundle branch block (LBBB), right bundle branch
outcomes (28). Whilst attention mechanisms are potentially block (RBBB), premature ventricular contraction (PVC), and
able to boost the performance of a neural network, they premature atrial contraction (PAC) were classified with accuracy
are not without limitations. They require vast amount of of 100, 98.66, 100, 99.66, and 100, respectively (35).
data for training, are not robust when generalised in other Atrial fibrillation is one of the most common arrythmias.
tasks other than the one they were trained for, cannot Subclinical AF can cause strokes, which can lead to disability
control spurious correlations in the data and no research and premature death. Efforts to diagnose asymptomatic AF are
has been undertaken to compare different attention models’ increasingly gaining momentum. Recently, a risk prediction
performances (29, 30). model (baseline and time-varying neural networks) was used
for AF diagnosis from a cohort of 604,135 patients in a
retrospective study. At the end of the follow up period
Applications of artificial (8 years), 3% of subjects had been diagnosed with AF, with
intelligence in cardiology the algorithm achieving an area under the curve (AUC) [the
integral of the ROC (receiver operator characteristic) curve,
Artificial intelligence applications in cardiovascular thus the proportion of correctly classified outcomes] of 0.87
research are increasingly becoming more popular over the in differentiating between patients with AF and those without
last decade (Figure 2). AI algorithms have been broadly AF (36). In a different study, CNNs were used to screen 12-
used for diagnosis from an image, image segmentation and lead ECGs for features not noticeable by the physician and
reconstruction, image quality control, patient prognostication, detected subclinical paroxysmal AF from ECGs with normal
phenogrouping, and gaining of scientific insight. Patient rhythm (SR). 454,789 ECGs from 126,526 individuals were
meta-data (demographics, co-morbidity) has been used to included in the training dataset, 64,340 ECGs were included
improve the performance of ML algorithms. AI based software in the internal validation set and 130,802 ECGs in the
devices and risk assessment tools have also been adapted in testing set (37).
the field of Cardiology. Table 3 demonstrates examples of such The decision to start antithrombotic therapy for patients
achievements. with newly diagnosed AF relies on the balance between two risk

Frontiers in Cardiovascular Medicine 07 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

TABLE 3 Overview of the use of AI in cardiology.

Application/Task Model Data Training set Testing set Accuracy/specificity/ References

sensitivity

Identification of arrythmia (SR, SVM ECG - - 100% (SR), 98.66% (15)

LBBB, RBBB, PVC, PAC) (LBBB), 100% (RBBB),
99.66% (PVC), and 100%
(PAC) accuracy
Discrimination of HCM from Ensemble ML (SVM, Clinical, - - 87% sensitivity and 82% (32)
ATH RF, ANNs) echocardiography specificity
Prediction of ACM in patients Boosted ensemble Clinical, CTA - 10,030 subjects AUC 0.79 (40)
with suspected CAD undergoing algorithm
CTA
Arrythmia detection (prediction DNN ECG - - 99% accuracy (13)
of 12 types of arrythmia
compared to cardiologist)
Detection of subclinical AF CNN ECG 454,789 images 130,801 images AUC 0.90, sensitivity (17)
82.3%, specificity 83.4%,
accuracy 83.3%
Identification of ventricular CNN ECG, 44,959 subjects 52,870 subjects AUC 0.93, sensitivity (52)
dysfunction (EF 35%) echocardiography 86.3%, specificity 85.7%,
and accuracy 85.7%
Phenogroup HF patients and Unsupervised ML Clinical, - 1,106 subjects - (60)
identification of responders to (Multiple Kernel echocardiography
CRT implantation Learning and
K-means clustering)
Automated analysis of cardiac CNN CMR 599 subjects 110 subjects - (47)
structure and function (left
ventricular chamber volumes,
mass and EF)
Prediction of CAD on CTA Boosted ensemble Clinical, CTA - 13,054 subjects AUC 0.881 (38)
algorithm (CACS)
Prediction of ACM for 1, 2-, 3-, RF Clinical, ECG, 2,282 subjects 1,510 subjects AUC 0.768 (1 year), 0.793 (59)
4-, and 5-years post CRT echocardiography (2 years), 0.785 (3 years),
implantation 0.776 (4 years), 0.803
(5 years)
Prediction of early coronary Ensemble LogitBoost Clinical, SPECT - 1,980 subjects AUC 0.81 (44)
revascularisation within 90 days algorithm
after SPECT MPI
Identification of patients with Tensor based ML CMR 200 subjects 1,122 subjects AUC 0.92 (48)
PAH algorithm
(multilinear
subspace learning)

ACM, all-cause mortality; AF, atrial fibrillation; ANNs, artificial neural networks, ATH, athlete’s heart; AUC, area under the curve [integral of the ROC (receiver operator characteristic)
curve]; CACS, coronary artery calcium score; CAD, coronary artery disease; CMR, cardiovascular magnetic resonance imaging; CNN, convolutional neural network; CRT, cardiac
resynchronisation therapy; CTA, Cardiac Computed Tomography Angiography; DNN, deep neural network; ECG, electrocardiogram; EF, ejection fraction; HCM, hypertrophic
cardiomyopathy; HF, heart failure; LBBB, left bundle branch block; ML, machine learning; MPI, myocardial perfusion imaging; PAC, premature atrial contraction; PAH, pulmonary
arterial hypertension; PVC, premature ventricular contraction; RBBB, right bundle branch block; RF, random forest; SPECT, single-proton emission computerised tomography; SR, sinus
rhythm; SVM, support vector machine.

stratification methods. The CHA2DS2-VASc score measures used to predict a patient’s risk of ischemic stroke, major
the possibility of a future ischemic stroke and the HAS- bleeding and death, and were compared to clinical risk
BLED score predicts a patient’s bleeding risk. They both scores by the AUC. GBM, the best performing ML algorithm
aid in the decision-making process of AF treatment on an
of all, showed modest performance improvement for stroke
individual basis. In a recent retrospective cohort study of
compared to CHA2DS2-VASc (AUC = 0.685 vs. AUC = 0.652),
9,670 patients, diagnosed with non-valvular AF and followed
but significant improvement in predicting major bleeding in
up for up to 1 year, multilabel ML methods were compared
to the currently used risk scores for prediction of outcomes comparison to HAS-BLED (AUC = 0.709 vs. AUC = 0.522)
in AF patients. SVM, gradient boosting machine (GBM) and and death in comparison to CHA2DS2-VASc (AUC = 0.765 vs.
multi-layer neural networks (MLNN) were the ML algorithms AUC = 0.606) (38).

Frontiers in Cardiovascular Medicine 08 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

Artificial intelligence and machine skill, a voice application which can provide various medical
learning medical devices guidelines including on cardio-pulmonary resuscitation. The
CardioCube voice application enhances paperless medical
Atrial fibrillation detection can be a difficult task as the history taking, in an outpatient cardiology clinic in Los Angeles.
current diagnostic methods (pulse palpation, ECG, ambulatory Patients answer verbally a set of pre-prepared questions and
Holter monitoring) all have limitations. Today various mobile produce high accuracy reports in the her system. Another
devices can be used for detection of AF. These include application by CardioCube (FCNcare), was implemented at a
smartphones, smart bands or smartwatches, earlobe sensors, family care network in Belligham (USA) and allowed HF and
and handheld electrocardiogram devices. These devices diabetic patients to update their medical status observations
are characterised by their non-invasive nature, safety, and from home. Telemedicine nurses obtained instant reports
instantaneous access to patients (39). from the produced EHRs and were able to triage the patients
Wearable devices are user friendly and allow uninterrupted accordingly. Voice applications can therefore be used as digital
monitoring and instantaneous individual analysis of ECG screening tools and red-flagging systems for patients with
signals. The Apple Watch and AliveCor are the most chronic diseases (44).
distinguished examples of wearable. The KardiaBand from Artificial intelligence and ML based medical devices
AliverCor is an example of a smartphone application, based on undergo a premarket review by the Food and Drug
ML for the recognition of AF from an ECG (40). In a RCT of AF Administration (FDA), before their widespread commercial
screening, the AliveCor Kardia monitor connected to a WiFi- use. The FDA has three levels of clearance for AI/ML based
enabled iPod attained iECGs from 1,001 ambulatory patients. algorithms. 510(k) clearance is granted when the algorithm has
The study showed that screening with a twice weekly single been shown to be at least as effective as another similar legally
lead iECG and remote analysis in ambulant patients aged 65 marketed algorithm. Premarket approval is granted for Class III
and above at high risk of stroke, was considerably more likely medical devices, which may have a vast impact in human health
to detect AF in comparison to routine monitoring over a 12- and require more thorough evaluation, to define their safety. De
month period (41). In another RCT the occurrence of recurrent novo pathway approval involves those novel medical devices
AF or atrial flutter via daily ECG self-recordings and the time for which there is no similar legally marketed product but
to treatment of the recurring arrythmia in patients undertaking have shown to be safe and effective for use. Table 4 illustrates
catheter radiofrequency ablation or direct current cardioversion examples of FDA approved medical devices used in the field of
for AF or atrial flutter, were evaluated. The chance of recurrence cardiovascular medicine (45).
identification was higher in the group which used the AliveCor
KardiaMobile ECG monitor (intervention group). The time
from detection to treatment was also shorter for that group (42). Transthoracic echocardiography
The Apple Heart Study showed that the utilisation
of smartphones was effective in identifying patients with Echocardiogram is characterised by its easy application and
subclinical paroxysmal AF. It included data from 420,000 widespread availability. It is an imaging modality that provides
participants, with a median follow up time of 117 days. a real-time imaging of the heart and instant identification
It detected 0.5% of patients with possibly irregular pulse, of any structural defects. AI enables the enhancement of
34% of which were diagnosed with AF confirmed by ECG. imaging measurements’ accuracy, via the decrease in inter-
The notification group (set informed of irregular pulse) had and intra-operator inconsistency and by offering further
higher possibility of commencing anticoagulant or antiplatelet information, subtle to be detected by the human eye (46). ML
treatment. Also, from the patients diagnosed with AF, 24% algorithms have been broadly used in the field of transthoracic
underwent cardioversion, 3% received an implantable loop echocardiography, with aim of diagnosis from an image, image
recorder, 20% started anti-arrhythmic medication, and 18% segmentation and patient prognostication.
undertook ablation (43). An innovative 2-dimensional echocardiographic image
Voice technology has been increasingly utilised for analysis system used AI-learned pattern recognition and
mainstream use via voice assistants, such as Amazon’s Alexa automatically calculated left ventricular EF (LVEF) (measure
or Google Assistant. These advanced software architectures of contractile function). The results of the study were similar
are based on neural network techniques undertaking the task to the results from the standard manual estimation (biplane
of speech recognition. They interpret a complex conversation Simpson’s method) and had less variability than visual EF
and generate human like responses. They are available either (47). A multicentre study investigated the possibility of a
on smart speakers or on smartphones. Voice assistants are fully automated computer vision software (AutoLV) using ML-
emerging tools for remote monitoring and undertaking of enabled image analysis, for measurements of left ventricular
medical services. For example, they can be used for educational volumes and EF, and average biplane longitudinal strain (LS) (a
purposes, such as in the case of the Mayo Clinic First Aid technique for evaluation of the left ventricular function). The

Frontiers in Cardiovascular Medicine 09 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

TABLE 4 FDA approved AI/ML based medical technologies/software.

Device name Parent company Description FDA approval Type of FDA Date
name number approval

Atrial fibrillation history Apple Inc. Detection of atrial fibrillation K213971 510(k) premarket 03/06/2022
feature notification
LINQ II Insertable Cardiac Medtronic, Inc. Arrhythmia detector and alarm (including K210484 510(k) premarket 11/06/2021
Monitor, Zelda AI ECG ST-segment measurement and alarm) notification
Classification System
Gili Pro Biosensor (Also Continuse Biometrics Hardware and software for optical DEN200038 De novo pathway 01/04/2021
Known as Gili Biosensor Ltd. camera-based measurement of pulse rate,
System) heart rate, breathing rate, and/or
respiratory rat
Analytic for Hemodynamic Fifth Eye Inc. Adjunctive hemodynamic indicator with DEN200022 De novo pathway 01/03/2021
Instability (AHI) decision point
ECG 2.0 App Apple Inc. Ambulatory ECG rhythm assessment K201525 510(k) premarket 08/10/2020
notification
Bodyguard Remote Preventice Technologies, Arrhythmia detector and alarm (Including K192732 510(k) premarket 26/03/2020
Monitoring System Inc. ST-segment measurement and alarm) notification
AI-ECG Tracker Shenzhen Carewell Assessment of arrhythmias using ECG K200036 510(k) premarket 20/03/2020
Electronics Co., Ltd data acquired from adults without notification
pacemakers
Eko Analysis Software Eko Devices Inc. Support in the evaluation of patients’ heart K192004 510(k) premarket 15/01/2020
sounds and ECG’s notification
FFRangio CathWorks Ltd. Analysis of previously acquired K192442 510(k) premarket 09/12/2019
angiography DICOM data for patients notification
with coronary artery disease
EchoGo Core Ultromics Ltd. Quantification and reporting of results of K191171 510(k) premarket 13/11/2019
cardiovascular function notification
AI-Rad Companion Siemens Medical Image processing software for analysis K183268 510(k) premarket 10/09/2019
(Cardiovascular) Solutions USA, Inc. from CT images to support physicians in notification
evaluation and assessment of
cardiovascular diseases
EMurmur ID CSD Labs GmbH Detection and identification of heart K181988 510(Kk) premarket 17/04/2019
murmurs notification
KardiaAI AliveCor, Inc. Ambulatory ECG rhythm assessment K181823 510(k) premarket 11/03/2019
notification
EchoMD Automated Ejection Bay Labs, Inc. Provides automated estimation of left K173780 510(k) premarket 14/06/2018
Fraction Software ventricular ejection faction based on notification
acquired transthoracic cardiac ultrasound
images
Acumen Hypotension Edwards Lifesciences Detection support for future hypotensive DEN160044 De novo 16/03/2018
Prediction Index (HPI) LLC events
Feature Software
Arterys Cardio DL Arterys Inc. Analysis of cardiovascular images K163253 510(k) premarket 05/01/2017
acquired from MR scanners notification
Steth IO Stratoscientific, Inc. Detection and amplification of sounds K160016 15/07/2016
from the heart and lungs

ECG, electrocardiogram; CT, computed tomography; MR, magnetic resonance.

automated measurements were achievable in 98% of the studies, and achieve detection of 23 viewpoints and segmentation of the
with an average analysis time of 8± s per patient. The study cardiac chambers across 5 common views. The CNN was a VGG
demonstrated a rapid and efficient way of assessment LVEF network which took a fixed-sized input of grayscale images,
and LS (48). passed it through 10 convolution layers, five max-pool layers,
Another ground-breaking publication concerns the and three fully connected layers. The output was fed into a 23-
fully automated echocardiogram interpretation in clinical way softmax layer, to represent 23 different echocardiography
practice, via ML methods. 14,035 echocardiograms with 70,000 views. Training data comprised of 10 random frames from
pre-processed images were used to train and evaluate a CNN each manually labelled echocardiographic video. The study

Frontiers in Cardiovascular Medicine 10 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

found that the automated measurements were comparable Cardiac computed tomography
or superior to the manual measurements across 11 internal angiography
consistency metrics (49). In another study, a CNN (consisting
of six convolutional layers, two fully connected layers, and a Coronary artery disease (CAD) risk assessment is
softmax classifier) was trained and validated on 200,000 images, fundamental in the efforts to reduce future cardiovascular
and was able to classify 15 standard echocardiography views, events. Traditional prediction models have limitations,
based on labelled still images and videos. The model classified including variations among the validation cohorts, a small
among 12 video views with a 97.8% overall test accuracy. It also number of predictors, and the absence of important variables.
exceeded board-certified echocardiographers’ accuracy among The need for robust prediction tools for accurate prediction
15 views on single low-resolution images (91.7 vs. 70.2–84, of CAD burden and the recent advancements in AI, led to the
respectively) (50). development of ML-based risk prediction models (55). Cardiac
From a cardiomyopathy point of view, ML algorithms computed tomography angiography (CTA) is a non-invasive
were trained from clinical, conventional echocardiography data imaging investigation that permits a direct evaluation of the
and speckle tracking echocardiography variables, to distinguish patency of coronary arteries and has been vital in ascertaining
constrictive pericarditis from restrictive cardiomyopathy. The the incidence of coronary artery disease (CAD) and consequent
associative memory classifier (AMC) was found to be the prognostication (56). Coronary artery calcium scoring (CACS)
best performing algorithm with AUC of 89.2%. This method with or without CTA can provide qualitative and quantitative
was found to be superior to the use of commonly used details on atherosclerosis, whilst CTA can determine the
echocardiography variables, for the differentiation between stenosis of an atherosclerotic lesion (32). ML methods are
these two diseases which carry many similarities (51). applied in CTAs, to maximise information extraction via image
Similarly, an ensemble ML algorithm model encompassed acquisition, and improve diagnostic accuracy and prognostic
by three algorithms (SVM, RF, and MLP with back propagation), outcomes via precision risk stratification.
incorporated speckle-tracking echocardiographic data, Data from 8,844 patients from a multi-centre registry
to automatically distinguish the condition of inherited were used to compare the AUC for conventional CTA risk
hypertrophic cardiomyopathy (HCM) from hypertrophy scores in comparison to a score created using a boosted
physiologically seen in athletes. The model demonstrated ensemble algorithm for risk stratification. With a mean follow-
increased sensitivity and specificity in comparison to standard up time of 4.6±1.5 years, the AUC was considerably better
diagnostic variables (52). Valvular disease can also be for the ML based approach, indicating that ML can improve
assessed using AI methodology. SVM classifiers were used risk stratification, compared to the current CTA risk scores
for classification and determination of the severity of mitral (57). In another multicentre study, 13,054 participants with
regurgitation (MR), a common valve disease. The method suspected or previously established CAD, underwent CACS
achieved sensitivity of 99.38% and specificity of 99.63% for the measurements. The CACS was used in a gradient boosting ML
identification of the severity of MR in normal subjects (53). algorithm (XGBoost) (boosting tree-based ensemble algorithm),
The most recent advancement of AI in echocardiography in combination with clinical risk factors, for assessment of
concerns a video-based DL algorithm, which exceeded human potential improvement of risk stratification. The study showed
experts’ performance in tasks such as EF estimation, assessment around a 9% increase in the ability to approximate pre-test
of cardiomyopathies and left ventricle segmentation. The probability of obstructive CAD, when adding CACS in the
variance in predictions from this algorithm is equivalent to or baseline model. In the subgroup of younger patients (less
less than measurements of cardiac function by human experts. than 65 years old) this was increased to around 17% (58).
EchoNet-Dynamic is an end-to-end deep learning approach. Another study investigated ML based risk stratification in
It uses the standard apical four-chamber view echocardiogram an asymptomatic healthy population. 85,945 asymptomatic
videos as input. Spatiotemporal convolutions with residual participants underwent a CTA scan with CACS, with 66
connections are used for prediction of EF of each cardiac cycle. available parameters. A ML algorithm was used to predict
Weak supervision from expert human tracing is used to generate moderate (CACS > 100) and high-risk (CACS > 400) CAD
frame-level semantic segmentations of the left ventricle. The patients and was compared a conventional risk prediction score.
outputs are combined to create beat-to-beat predictions of 8.4 and 2.4% of the population had indication for moderate and
the EF and the presence of HF with reduced EF (via AUC). high-risk CAD, respectively.
EchoNet-Dynamic was created by using 10.030 apical four- The study showed that the ML algorithm was superior to
chamber echocardiogram videos during training of the model. the conventional risk prediction score, in both the moderate
It is the first video-based DL model for echocardiogram and and high risk for CAD groups (59). From a clinical perspective,
its performance in measuring EF is better than previous DL the use of ML for prediction of CAD for all patient subgroups
models. It can rapidly identify subtle changes in EF and aid the leads to effective precision risk stratification (exploration of all
precise diagnosis of CVD in real time (54). available information for calculation of each individual’s risk),

Frontiers in Cardiovascular Medicine 11 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

less exposure radiation (as CTA with CACS has not been found Single-photon emission computed
to be superior of ML with CACS alone) and a more automated tomography
and accurate selection process for further diagnostic evaluation
of the appropriate candidates with better clinical outcomes. The Assessment of myocardial perfusion correlates to the
use of ML for CAD prediction aims to create risk stratification existence of obstructive CAD. This can be implemented with
models that are more accurate and cost and time efficient in single-photon emission computed tomography (SPECT) stress
clinical practice, compared to conventional models (58). testing (32). AI methodology has been applied in this modality,
Motwani et al., studied 10,030 patients with suspected CAD aiming to improve tasks such as image acquisition, image
during a 5-year follow-up from an international multicentre reconstruction and automated quantitation. An ensemble-
study. All patients underwent clinically indicated CTA. 25 boosting ML algorithm, LogiBoost, incorporated from clinical
clinical and 44 CTA parameters were measured. The ML data and quantitative image features, compared the diagnostic
approach concerned automated feature selection by information accuracy of the model for estimation of obstructive CAD to
ranking, model building with a boosted ensemble algorithm the standard quantification [total perfusion deficit (TPD)] and
(LogitBoost) and 10-fold stratified cross-validation, through the visual analysis by two experienced readers. The accuracy of the
whole process. The primary outcome of the study was all-cause ML algorithm was similar to Expert 1 and superior to combined
mortality (ACM). 745 patients died during the 5-year follow supine/prone TPD and expert 2, showing an improvement in
up. The ML approach was identified as a significantly better the diagnostic performance of myocardial perfusion imaging
predictor of a 5-year ACM, in comparison to the clinical or (MPI), by integration of ML (62). The automatic prediction of
CTA measures alone, as indicated by the higher AUC in all obstructive CAD from myocardial perfusion imaging (MPI) by
comparisons (60). DL, compared to TPD, was assessed in a more recent multicentre
Most cases of acute coronary syndrome (ACS) are caused study. 1,638 patients without known CAD underwent stress MPI
by unstable but non-obstructive atherosclerotic plaques. The and invasive coronary angiography (ICA) within 6 months of
current available non-invasive diagnostic tests which detect MPI. AUC for disease prediction (obstructive disease defined
coronary artery stenosis or stress-induced myocardial ischemia, as > 70% of narrowing of coronary arteries) by DL was
are unable to detect these unstable non-obstructive plaques. It higher than TPD, showing the potential for improvement
of automatic interpretation of MPI, by AI methods (63). In
is established that vascular inflammation causes atherosclerotic
another multicentre study, 1,980 patients with suspected CAD,
plaque formation and rupture, leading to ACS. The perivascular
underwent stress MPI with novel SPECT scanners. All patients
fat attenuation index (FAI) is an AI-derived imaging biomarker,
had subsequent ICA within 6 months. LogiBoost, was also
which captures the alterations in perivascular fat attenuation,
utilised in this study to forecast early coronary revascularisation
caused by vascular inflammation. Two independent cohorts
within 90 days after the SPECT MPI and was compared to
including 1,872 patients undergoing CTA, investigated the
standard quantitative analysis (TPD) and expert interpretation.
prognostic value of perivascular fat attenuation mapping for
The ML outperformed both TPD and expert analysis, with an
all-cause and cardiac modality. Perivascular fat attenuation
AUC of 0.79 (64).
mapping was performed around the three major coronary
arteries. In both the derivation and validation cohorts, high
perivascular FAI values around the proximal right coronary
Cardiac magnetic resonance imaging
artery and left anterior descending artery, were projecting
of all cause and cardiac mortality. A cut-off of −70 for Cardiac MR is an imaging modality utilised for non-invasive
the perivascular FAI was determined, above which sharp assessment of CVD. It evaluates the cardiac morphology,
rise in cardiac mortality was observed. This ground-breaking function, perfusion, and quantitative myocardial tissue
study showed that perivascular FAI, an AI-derived biomarker, measurement (65). It is considered the gold standard for
provides a quantitative measure of coronary inflammation non-invasive evaluation of EF and left ventricular volume (66).
and increases cardiac risk prediction and reclassification CMR is broadly used for the diagnosis of cardiomyopathies,
over current-state-of-the-art valuation via CTA. Non-invasive congenital heart disease, valvular heart disease, IHD, pericardial
detection of coronary inflammation via FAI can lead to timely lesions, and cardiac tumours. However, it requires acquiring
and aggressive initiation of primary prevention for patients with images characterised by high temporal and spatial resolution,
no visible CAD but unstable atherosclerotic plaques that can different contrasts and/or whole-heart coverage, leading to a
potentially lead to myocardial infarction if untreated. Prompt lengthy scanning time. ML incorporation to CMR, can lead to
prevention will lead to reduced incidence of clinically diagnosed a more efficient scanning and accurate interpretation process.
CAD and requirement for further intervention. FAI can also DL based MRI reconstruction is based on a model which learns
guide future trials in assessing novel but affordable therapeutic the factors of the reconstruction procedure beforehand, so
agents that target inflammation (61). that it can be applied to all new data as a simple operation.

Frontiers in Cardiovascular Medicine 12 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

Most important tasks utilised with DL are image construction, of batch normalisation after each hidden convolutional layer
image segmentation and image quality control in the field to stabilise the training and the use of dropout regularisation
of CMR (65). after each concatenating operation to avoid over-fitting. The
Whilst CMR is executed at high resolution, analysis of the U-net was tested on 600 subjects from the UK biobank for the
scan by the clinician remains variable, time consuming and purpose of intra-domain testing and on 699 subjects from two
prone to errors. Deep learning methodology has been used to other sets (ACDC dataset and BSCMR-AS dataset) for cross-
overcome the challenge of automated derivation of information domain testing. Chen’s proposed method was compared to Bai’s
from CMR images (67). An automated 2-dimensional CNN automated methodology. Despite both achieving comparable
(adapted from the VGG-16 network) was used to take CMR dice scores on the intra-domain UK biobank test set with high
image as input, learn image features from fine to coarse accuracy, Chen’s method achieved superior mean dice scores
scales through a series of convolutions, concatenate multi- for all of the three structures (LV, RV myocardium) on the two
scale features and predict a pixelwise image segmentation (67). cross-domain datasets. The proposed method achieved better
The CNN was trained on 599 independent multicentre disease overall segmentation accuracy with lower variance on the three
cases and subsequently was compared to an expert cardiologist datasets, improving CNN-based model generalisability for the
and a trained junior cardiologist for the identification of left CMR image segmentation task across different scanners and
ventricular chamber volumes, mass, and EF, from 110 patients sites (70).
who underwent scan: rescan CMR within a week. The study Late gadolinium enhancement (LGE) CMR imaging is the
showed that clinicians can detect a 9% change in LVEF with cornerstone of non-invasive myocardial tissue characterisation.
the greatest source of human error being attributed to the An important example consists of the relationship between the
observer. The precision of the CNN was similar to human presence and extend of LGE and adverse outcomes, in patients
analysis, but its performance was 186 times faster (68). In with HCM. However, LGE necessitates the administration of
another study, a ML approach was used for the identification an intravenous gadolinium-based contrast agent, which should
of diagnostic features in pulmonary arterial hypertension be used cautiously in patients with severe renal failure or
(PAH) using CMR. A multilinear principal component analysis allergy to gadolinium-based contrast. In a recent RCT of
(MPCA) algorithm was utilised to extract low-dimensional 1,348 patients with HCM, a new DL driven CMR technology
features from high dimensional input to tensor representations named virtual native enhancement (VNE), was used to generate
of data. The algorithm distinguished patients with and without images identical to the standard LGE, without the need of a
PAH with higher accuracy in comparison to manually drawn gadolinium-based contrast agent. The DL algorithm consisted
CMR measurements with an AUC of 0.92. Additionally, the of three parallel CNNs streams which processed and enhanced
diagnosis that used the ML approach was less time consuming signals in native T1 maps (pixel-wise maps of tissue T1
(within 10 s) and had less variability (69). relaxation times) and cine imaging (sequence of images at
Bai et al., trained a 16-layer CNN (adapted from the VGG- different cardiac phases) of cardiac structure and function.
16 network) on a 4,875-subject dataset from the UK biobank, Each stream had an encoder-decoder U-net architecture. The
to automatically analyse CMR images. Its performance was encoder computed image features from fine to coarse and
assessed using technical [dice coefficient (metric of the similarity produced a multiscale feature representation, which the decoder
of the two segmentations)] and clinical [left ventricle (LV) end- combined to produce final feature maps. The feature maps
diastolic volume (LVEDV) and end-systolic volume (LVESV), from U-nets were concatenated and input into a further
LV mass (LVM); right ventricle (RV) end-diastolic volume neural network, to produce a final VGE image. The neural
(RVEDV) and end-systolic volume (RVESV)] parameters. The networks were trained with the use of a modified conditional
automated method achieved great performance in segmentation GAN approach. When compared, the VNE imaging achieved
of the LV and RV on short-axis CMR images (dice metric better image quality than LGE and was in high agreement
of 0.94 and 0.90 accordingly) and the left atrium (LA) and with it in visuospatial distribution and myocardial lesion
right atrium (RA) on long-axis images (dice metric of 0.93 quantification. Overall, VNE resembles conventional LGE,
and 0.96 accordingly), from an intra-domain UK biobank test but does not require intravenous access or administration of
set of 600 subjects. The automated method using the CNN contrast, can be repeated if required to confirm the imaging
was comparable to human inter-observer variability (67). CNNs findings without the consequences of giving contrast, and can be
perform segmentation tasks with great precision if the training completed within 15 min as uses native imaging. Its advantages
and test images originate from the same scanner or site. make it an attractive technology which can potentially be
However, their performance decreases when the test images extended for diagnosis of other myocardial pathologies in the
come from different scanners or sites. Chen et al., trained a future (71).
2D CNN (U-net) on 3,975 subjects from the UK biobank. The In addition to image construction and segmentation, DL
U-net architecture has been the most widely used architecture methodology has recently been utilised for image quality control
for biomedical segmentation. In Chen’s paper, the network was purposes in the field of CMR. In a retrospective study of 3,827
identical to the original except of two differences, application subjects (including healthy and pathological hearts), a set of

Frontiers in Cardiovascular Medicine 13 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

algorithms including CNNs were used for the development of (EF 35%). The CNN was composed of three parts. Six
a framework for the automatic detection and quality-controlled convolutional blocks (convolution, batch normalisation, Relu,
selection of cine images, used for cardiac function analysis from max pooling) extracted temporal features, one convolutional
routine clinical CMR exams. The framework encompassed three block (convolution, batch normalisation, Relu) extracted spatial
steps. The first pre-processing step excluded still images. In features and two fully connected layers (fully connected, batch
the second step, one CNN classified images in standard cine normalisation, Relu, dropout) regressed the features to a
views and a second CNN classified images depending on the softmax activated output. The DL algorithm was tested on a
image quality and orientation. The final algorithm selected one set of 52.870 patients and showed AUC, sensitivity, specificity
good image of each class, which was then used for analysis and accuracy of 0.93, 86.3, 85.7, and 85.7%, respectively.
of cardiac function. The classification CNN achieved accuracy Patients without ventricular dysfunction with a positive AI
between 0.989 and 0.998, whilst the quality control CNN screen were four times more probable to develop future
achieved accuracies of 0.861 for 2-chamber, 0.806 for 3-chamber, ventricular dysfunction, in comparison to patients with a
and 0.859 for 4-chamber views. The complete framework also negative screen. An inexpensive, non-invasive test such as AI
achieved accuracies of 89.7, 93.2, and 93.9% for 2-, 3-, and 4- screening from ECG data, can be a powerful future tool for
chamber acquisition from each study, respectively. This study screening asymptomatic individuals (76).
demonstrates the future potential for high quality automated Readmission rates after hospitalisation with HF remain high
cine CMR analysis from the scanner to report (72). and lead to increased disease burden and costs for healthcare
systems. The performance of current predictive methods for
the likelihood of HF readmissions is modest. Accumulating
Heart failure publications show positive results from ML-driven methods for
predicting readmission of patients with HF (77–79). In one
Heart failure (HF) affects 1–2% of the adult population in of the studies, EHRs were used to enrol 1,653 patients within
developed countries and more than 10% of patients > 70 years 30 days of their discharge after an index admission for HF.
of age (73). Early diagnosis or prediction of HF has a high impact ML algorithms were compared to the traditional method of
on successful treatment and prolongation of life expectancy for logistic regression (LR), for effectiveness in predicting 30 and
patients. ML applications have been used for the early detection 180 days all cause readmissions and readmissions due to HF.
of HF, classification, severity estimation, and prediction of For the 30-day all-cause readmission prediction, random forest
adverse events (e.g., 30-day re-hospitalisation) (32). (RF) showed a 17.8% improvement over LR. For readmissions
A multi-level risk assessment for developing HF, with because of HF, boosting showed a 24.9% improvement over
prediction of five risk levels (no risk, low, moderate, high, and LR. Lastly, the ML models stipulated enhanced recognition of
extremely high risk), via the use of a decision tree classifier, was groups at low and high risk for readmission, by increasing the
established in a study by Aljaaf et al. They also added three predictive range compared with LR (80). A retrospective study
new risk factors (obesity, physical activity, and smoking) in a applied DL methodology [deep unified networks (DUNs)] to
previously used dataset and enhanced the accuracy of predicting data from EHRs of 11,510 patients, in order to generate a risk
HF. This was the first study with a multi-level prediction of prediction model to forecast 30-day readmissions in patients
HF, in contrast to the binary outcomes from previous studies. with HF. DUNs consist of a mesh-like network structure which
The predictive model showed an improvement from existing avoid overfitting. The DUNs’ AUC (0.705) had the best result
studies with a sensitivity of 86.5% and specificity of 95.5% (74) of 10-fold cross-validation, compared to LR (0.664), gradient
A SVM was trained on clinical parameters from 289 patients boosting (0.650) and maxout networks (activation function used
and triaged patients into three categories (HF, HF-prone, and in neural networks) (0.695). The DUNs model also showed an
healthy). The overall classification accuracy was 74.4%, with accuracy of 76.4% at the classification threshold with greatest net
precisions of 78.79, 87.5, and 65.85% for ascertaining the healthy savings for the hospital (81).
group, HF-prone group and HF group, respectively. The scoring Cardiac resynchronisation therapy (CRT) is fundamental
model showed improved accuracy in classification of HF, in to the management of symptomatic HF with left ventricular
comparison to clinical practice criteria (25–50% accuracy) (75). systolic dysfunction and intraventricular conduction delay
More recently, a DL approach was used to screen (reduced EF and wide QRS complex). Conventionally, patients
individuals for asymptomatic left ventricular dysfunction eligible for CRT implantation, should have an ECG morphology
(ALVD). ALVD is prevalent in 1.4–2.2% of the population and with LBBB and QRS duration ≥ 150 ms. Patients with these ECG
if left undiagnosed, it can lead to increased morbidity and characteristics have greater benefit on reduction of mortality
mortality. Early identification of ALVD and commencement and readmissions after receiving CRT. However, around 30%
of treatment, can prevent its progression to symptomatic HF of patients meeting these criteria and receiving an implant, do
and reduce mortality. A CNN was trained on 12-lead-ECG and not experience clinical benefit from CRT. Therefore, predicting
echocardiography information, including LVEF, from 44.959 a patient’s outcome after CRT is an essential step in the decision-
individuals to identify patients with ventricular dysfunction making procedure pre-implantation. In a retrospective study,

Frontiers in Cardiovascular Medicine 14 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

ML models were developed for prediction of ACM or HF Machine learning application in the field of HF, has
hospitalisation at 12 months post-CRT. Clinical characteristics enabled the accurate prediction of HF, ALVD, and low EF in
and ECG features were used in model development. A RF asymptomatic individuals. Early identification of patients at risk
algorithm was found to be the best performing model, when of developing HF or at an early onset of the condition, can
compared to other ML models and the traditional ECG lead to prompt and aggressive primary prevention/initiation of
prediction methodology (LBBB and QRS ≥ 150 ms). Whilst the treatment and more rigorous follow up of these patients, with
ECG morphology did not reach significance for differentiation improved clinical outcomes. Prompt detection of patients who
of survival difference across subgroups (p = 0.08), the RF model might require re-admission after hospitalisation with HF, can
formed quartiles of patients with an 8-fold change in survival initiate a well-structured outpatient pathway for this cohort,
between those with the maximum and minimum predicted which will aim to keep these patients in the community and
probability for events (p < 0.0001). Furthermore, the RF model reduce re-admissions and subsequent morbidity and mortality
achieved better discrimination of the risk of the composite end and hospital costs. This could be achieved, through liaison
point of ACM and HF readmission, than the ECG morphology- of the secondary care team with the community heart failure
based subgroup analysis (82). team and the patients’ GP and maximisation of their treatment
Another retrospective study concerned data from 1,510 in the community, via more regular reviews at home, at the
patients who underwent CRT implantation. ML models were GP practice or in an outpatient clinic or ambulatory setting.
trained from 33 pre-implant clinical features, to predict 1–5- Moreover, accurate and personalised prediction of the cohort of
year ACM. The best performing ML model (highest AUC for patients which would have a good outcome if having a CRT-D
inserted, can lead to the reduction of unnecessary procedures
the prediction of all-cause mortality at 1, 2, 3, 4, and 5 year
(and subsequently reduced hospital costs and resources) and
follow up), a RF model, was chosen for further assessment
the associate medical risks for those patients who would not
and it was referred as the SEMMELWEIS-CRT score. This
have the same outcome. Overall, the incorporation of ML
was compared to pre-existing scores and showed significantly
methodology into the field of HF aims the early detection of
better response prediction and improved discrimination of
those patients most at risk of developing the disease, correct
mortality. An online calculator was developed, which will enable
classification of patients based on their personalised risk and
a personalised calculation of predicted mortality in patients
prompt intervention which can be beneficial for both the
undergoing CRT implantation (83). Similarly, in another study
patients (improved morbidity and mortality via early initiation
a ML-based approach was used to phenogroup a HF cohort and
of treatment) and secondary care (via shifting treatment and
identify responders to CRT. 1,106 patients from a multicentre
follow up in the community and reducing hospital admissions).
trial, were randomised into two groups [CRT with a defibrillator
Lastly, an important aspect of ML models and their application
(CRT-D) or an implantable cardioverter defibrillator (ICD)].
in clinical practice is the myriad of signals they can highlight
An unsupervised ML algorithm, via dimensionality reduction
within the data, which can potentially aid in the better
and clustering, classified patients into groups, based on clinical understanding of a particular aspect of the disease (which would
parameters, left ventricular volume, and deformation traces not be noticeable either way) and lead to further scientific
at baseline. The treatment effect of CRT-D on the primary discoveries in the future.
outcome (all cause death/HF event and on volume response)
was compared among the different groups. From the four
phenogroups identified, two had a greater proportion of known Other applications
clinical characteristics prognostic of CRT response and were
linked to an improved treatment effect of CRT-D on the primary An algorithm for heart murmur detection was developed
outcome (84). in a virtual clinical trial, with aim to enhance the precision of
Machine learning has enabled physicians to use data from screening for valvular and congenital heart diseases. 3,180 heart
ECGs and draw specific echocardiography results, without the sounds recordings (pathologic murmur, innocent murmur, no
use of the echocardiogram. A randomised controlled trial (RCT) murmur), from 603 outpatient visits were chosen from a large
aimed to identify patients with low ejection fraction (EF), via AI- database. Algorithm assessment of heart rate (HR) showed great
enabled ECGs. Low EF is an important marker of heart failure similarity to the gold standard. Pathologic cases were identified
which can be effectively treated to improve survival if recognised with sensitivity of 93%, specificity of 81%, and accuracy of
early. The study showed that the use of an AI algorithm 88%. This trial was the first to objectively evaluate an AI-based
(using neural networks) based on ECGs, led to the diagnosis of murmur detection algorithm, making it a potentially useful
patients with low EF at an early stage in the setting of routine screening tool for heart disease (88).
primary care (85). Other studies enabled the calculation of other Various studies have been carried out to demonstrate
parameters such as left ventricular hypertrophy (LVH) and left effectiveness of AI-driven phenogrouping, in all fields of
ventricular diastolic function (LVDF) based on ECG features cardiology. Such landmark study concerns the development
and ML methods (86, 87). of a phenomapping-derived tool, for selection of anatomical

Frontiers in Cardiovascular Medicine 15 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

or functional testing in patients with stable chest pain. train a CNN. If more than 50% of the subject’s rPPG segments
The decision support tool named ASSIST (Anatomical vs. were identified as AF rhythm by the model, the participant
Stress testing decision Support Tool), was developed using would be classified as AF. The accuracy of the DL model for
data from the PROMISE (PROspective Multicentre Imaging discrimination of AF from NSR and other ECG abnormalities,
Study for Evaluation of Chest pain) trial. Data from 9,572 was compared to the standard 12 lead ECG. The DL model
patients undergoing anatomical (4,734) or functional (4,838) achieved a 90 and 97.1% accuracy in detecting AF in 30 s and
imaging were used to create a topological presentation of the 10-min recordings, respectively (91).
study populations, based on 57 pre-randomisation variables.
Individual patient-centred hazard ratios for MACE were
calculated with Cox regression models, within each patient’s Internet of things
5% topological neighbourhood, leading to heterogeneity in the
map and distinction of phenotypic neighbourhoods favouring Internet of Things (IoT) is described as “a network of devices
either anatomical or functional imaging. A gradient boosting interacting with each other via machine to machine (M2M)
algorithm was used in 80% of the PROMISE population, in communications, enabling collection and exchange of data” (92).
order to predict a personalised outcome if using anatomical IoT can be applied in medicine in fields such as remote
or functional testing and create the ASSIST tool. The ASSIST health monitoring, chronic diseases management, obedience to
tool was tested in the rest 20% of the PROMISE population treatment and medication at home and elderly care (93). When
and in an external validation cohort (from the SCOT-HEART it comes to cardiology applications, IoT can be applied for the
trial), undergoing anatomical or functional testing as first identification of cardiac emergencies remotely. Wolgast et al.,
assessment. The testing stagey recommended by ASSIST showed designed a body area network for measurement of an ECG signal
a significantly lower incidence of each study’s endpoint and and its transmission via Bluetooth to a smartphone for data
of ACM or non-fatal AMI. The personalised novel tool can analysis. The user’s own smartphone would process the data
support physicians in the decision to proceed with anatomical and built-in communications could be used to raise an alarm
or functional testing when evaluating patients with stable chest if a heart attack was identified (94). In another study, subjects
pain (89). were observed for a period of 3 months, using a wearable sensor
An interesting example of the use of AI methodology, is a which documented physiological data. Data were uploaded
recent study by a Chinese group of scientists who developed a constantly via a smartphone to a cloud analytics platform. A ML
CNN (50-layer ResNet classification network) which detected model was used to design a prognostic algorithm which detected
CAD (stenosis > 50% documented by angiography), via HF exacerbation and predicted rehospitalisation after a HF
analysing the patient’s facial photo. 5,796 patients were divided admission (95).
to a training (5,216) and validation sets (580) for the algorithm
development. The AI algorithm’s AUC was 0.730 and was found
to be higher than the standard prediction scores. Sensitivity Ethical dilemmas
was 80% and specificity was 54%. Further studies would need
to be conducted, as the study had several limitations including Despite its huge potential, AI is still something new,
the geographical characteristic of the cohort (only Chinese unfamiliar, and sometimes difficult to comprehend. It therefore
population). Significant CAD was defined based on coronary carries various ethical dilemmas and limitations that need to
angiogram or CTA data, which led to small selection bias, be addressed. Firstly, the design of studies based in AI and
which could have potentially altered the algorithm outcome. the training and validation process of the new technology,
Lastly, in the visualisation tests, the cheek, forehead, and nose can be flawed. Most of the studies reporting AI applications
contributed more to the algorithm than other areas of the face. have retrospective design and small sample size, which can
This could be the result of the extraction of features by the DL potentially lead to bias. More importantly, AI-driven studies
algorithm that are associated with CAD but are not obvious to can have selection bias, which includes sampling and observer
human observers. Despite its limitations, this example shows the selection bias (32). From another perspective, since AI-driven
vast advancements and the future potential of AI applications technologies achieve their results from existing features and
in cardiology, with the generation of results from a simple dynamics of the populations they analyse, this can lead to
intervention such as taking a selfie (90)! reproduction, amplification of patterns of marginalisation,
In a recent prospective, single-centre study, contactless inequalities and discrimination that exists in these populations.
facial video recording was used to train a 12-layer DNN for Again, the features of the data chosen to train the algorithms are
the detection of asymptomatic AF. A camera based remote chosen by the investigators and the AI-driven application can
photoplethysmography (rPPG) was used during a 10-min facial replicate the investigator’s preconceptions and biases (96).
video recording of 453 study participants. Its signals were Foundation models are AI systems that are trained on
extracted and segmented into 30 s clips, which were used to broad data and can be adapted to a variety of downstream

Frontiers in Cardiovascular Medicine 16 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

tasks (33). An example of such model is AlphaFold, which the model’s accuracy and transparency. XAI ensures AI systems
is an AI system developed by DeepMind and can predict a meet regulatory standards and adopt good practice towards
protein’s 3D structure from its amino acid sequence. These accountability, making them easier and faster to deploy in
predictions are easily available to the scientific community and businesses and high risk environments such as the medical
can provide individual downloads for the human proteome field (101).
and the proteomes of 47 other key organisms in research The rapidly increasing use of smart medical devices and
(97). Such systems require vast amount of data for training digital health applications through IoT and AI, imposes a danger
and large amount of computing resources to effectively use of dehumanisation of medicine. More and more intelligent
that data. Consequently, the ownership of the data and their applications replace the work of physicians in various sectors.
models is often centralised, giving power and decision rights For example, the detection of new AF via a wearable device,
to organisations with the most resources (big tech companies) aids the diagnosis of a potential arrythmia, but takes away
and reducing opportunities for others, leading to inequalities, or the clinician’s opportunity to demonstrate their knowledge and
one can say anti-democratic situations. Despite efforts to build practice their skills (e.g., detect the arrythmia from a real time
models through distributed training, it is highly likely that this ECG for the first time). On the other hand, the potential of AI in
large gap will remain between the two (98). reducing admin burden for physicians (e.g., analysing EHRs),
Since ML models learn on high dimensional correlations can create the opportunity for having more interaction and
which exceed the interpretive abilities of humans, the rationale quality time with their patients. Balance is what is required, in
behind algorithmically produced outcomes which affect order to maintain healthy physician to patient relationship, with
decision making for patients, can remain unjustifiable. The integration of AI technologies when needed to relieve admin
absence of a familiar logic behind its output, might lead the burden (102).
clinician who is interpreting it to pause. Also, when decisions, Introduction of IoT and AI-driven tools for medical
predictions or classifications are made based on AI systems, monitoring of various parameters in individuals has generated
individuals are unable to hold direct accountability for these ethical concerns (102). AI technologies can utilise such personal
outcomes. In the case of harm of the patients, this accountability data, without obtaining the proper consent of the data subject
gap can affect their autonomy and violate their rights (96). or handle it in a manner personal information is revealed.
The question of trusting a system, which might not even be One can say that the ability to lead a private life, could be
understood by the decision maker, is raised. In the end, could jeopardised (96). The concern of personal data privacy is raised,
such systems beat a clinician’s judgement, if there is a conflict as most data protection laws are based on principles established
in management plans? For example, would the automated in 1980, which might not be reflecting the current reality.
diagnosis of an ECG-reading algorithm saying that a patient As per current laws, personal data should be collected and
has a STEMI (ST elevation myocardial infarction) surpass a used for a specific purpose. Also, data should be sufficient,
clinician’s view who is aware that the ECG changes are due to relevant, and restricted to what is required in the context of the
long-standing LBBB? The answer is no. Recently, it was found purpose of its use. However, when AI is concerned, neither of
that DL models made incorrect decisions by using cautiously the above can be guaranteed. The AI algorithms are complex,
engineered inputs, raising concerns that such systems are not always understood by their programmer, can generate
not yet ready for mainstream use (99). The same concern is surprisingly different results from what was expected and can
raised with non-robust CNNs and ML models under various lead to a change in the purpose, through the learning and
circumstances, such as in the previously discussed case of development process. Data can neither be restricted or deleted
adversarial attacks. Physicians should have the last and most after its original use, as keeping data is vital for the models’
important say when AI is applied for decision-making in optimal performance. It is now more essential than ever that the
the medical field. data protection laws are re-visited and adjusted to work better
Another recent development, which aims to mitigate the towards data privacy issues arising from the vastly growing fields
famous issue of “black-box” AI methodologies, is explainable of AI, big data and IoT (103).
AI (XAI). As AI becomes more advanced, it is less understood
by humans. Whilst lower performance systems such as ML
learning are more understandable, higher performance models
such as DL techniques are difficult to comprehend even from Translation of artificial intelligence
the engineers or data scientists who created the algorithms, since to future clinical practice
they are directly created from data. In safety critical situations,
such as in medicine, the non-transparency of these techniques Despite the landmark studies exhibiting the potential of AI
can lead to wrong decision making and pose serious danger in transforming medicine, the ethical dilemmas concerning its
to human lives (100). The aim of XAI is to allow humans to real-life implementation are still unaddressed. AI systems can be
comprehend how the algorithm works, trust its results and the flawed and their generalisability to new populations and settings,
output it produces, unmask potential biases, and characterise may produce bad outcomes and lead to poor decision-making.

Frontiers in Cardiovascular Medicine 17 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

Going forwards, education of scientists, physicians but plans. AI will be a part of every cardiologist’s daily routine to
also of the public regarding AI and the logic behind its provide the opportunity for effective phenotyping of patients
applications is vital. This can lead to better understanding and design of predictive models for different diseases. It will
and improved engagement in commercialisation of AI enhance the use of non-invasive diagnostics and reduce the need
applications. Medical engineering has been incorporated in for costly and complicated invasive tests, for the diagnosis of
several universities’ curriculum. Subjects such as computational CAD. Future cardiologists will be able to tell an asymptomatic
sciences, coding, and algorithmics, should also be incorporated patient, whether they will develop a lethal arrythmia or an MI
in the curriculum. Universities have also started providing short and what needs to be done to avoid this. Cardiologists should
courses and postgraduate level degrees on AI in healthcare. educate themselves in the development of AI and take part in AI
Educated physicians in AI, could aid adoption of innovative innovations and utilise them in their practice. However, they will
applications, but also raise awareness when ethical and privacy need to take into consideration the ethical dilemmas generated
issues are risen. in areas where AI is replacing human and aim to integrate
Another important aspect is the achievement of robust their knowledge and AI-derived suggestions, for a mature and
regulation and quality control of AI systems. As AI is a new accurate decision making in every step in the decision process.
and rapidly evolving innovative field, it carries significant risks
if underperforming and unregulated. As previously mentioned,
the FDA has recently released a regulatory framework with Author contributions
aim to establish safe and effective AI- based medical devices,
which can progress for patient use (104). The European Union All authors listed have made a substantial, direct, and
has also proposed a regulatory framework on the use of AI, intellectual contribution to the work, and approved it for
with plan to come into force in the second half of 2022, in a publication.
transitional period (105). Regulations should also be established
for upgrades of AI products, throughout the lifespan of the
product. Some AI systems have been built with continuous Funding
updates, but this could potentially result to drift with time.
Periodical updates after a complete evaluation of the clinical DA acknowledged Wellcome Trust Award (221604/Z/20/Z)
significance of the AI product are preferred. Guidelines should and Bart’s Charity Large Project Grant (G-002145). NA
also be developed for the purpose of evaluation of the product’s acknowledges the support from an Academy of Medical Sciences
performance and the detection of deficits over time (106). Starter Grant for Clinical Lecturers (SGL024/1024).
Due to the various limitations and ethical dilemmas AI
carries and its potential harm to the public, it is necessary to
incorporate AI ethics and safety, during the development of AI Acknowledgments
systems. AI ethics “is a set of values, principles, and techniques
that employ widely accepted standards of right and wrong to guide NA acknowledges the National Institute for Health Research
moral conduct in the development and use of AI technologies.” (NIHR) Integrated Academic Training Programme which
An ethical platform is required for the responsible delivery of supports his Academic Clinical Lectureship post. Figures
an AI project. This necessitates cooperation from all the team created in Biorender.
members of the multidisciplinary team, in order to maintain a
culture of responsibility and execute a governance architecture
that will adopt ethically practices at every point in the innovation Conflict of interest
and implementation lifecycle. Overall, the AI project needs to be
ethically acceptable, fair and non-discriminatory, justifiable and The authors declare that the research was conducted in the
worthy of public trust (96). absence of any commercial or financial relationships that could
be construed as a potential conflict of interest.

Conclusion
Publisher’s note
In our fast-paced world, time is precious and limited.
Healthcare is facing a crisis of understaffed departments and All claims expressed in this article are solely those of the
more informed patients who demand the best treatment. There authors and do not necessarily represent those of their affiliated
is an unmet need for the effective triage of patients, efficient organizations, or those of the publisher, the editors and the
clinical evaluation and incorporation of clinical expertise reviewers. Any product that may be evaluated in this article, or
with evidence-based medicine and the latest technologies and claim that may be made by its manufacturer, is not guaranteed
accurate decision making for the right diagnostics and treatment or endorsed by the publisher.

Frontiers in Cardiovascular Medicine 18 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

References
1. Itchhaporia D. Artificial intelligence in cardiology. Trends Cardiovasc Med. 24. Muralidhar K. Outlier Detection Methods in Machine Learning. Towards
(2022) 32:34–41. doi: 10.1016/j.tcm.2020.11.007 Data Science. (2022). Available online at: https://towardsdatascience.com/outlier-
detection-methods-in-machine-learning-1c8b7cca6cb8 (accessed July 2022).
2. Seetharam K, Shrestha S, Sengupta PP. Artificial intelligence in cardiovascular
medicine. Curr Treat Options Cardiovasc Med. (2019) 21:25. doi: 10.1007/s11936- 25. Coe R. It’s the effect Size, Stupid. What effect size is and why it is important.
019-0728-1 Paper Presented at the British Educational Research Association Annual Conference.
Exeter: (2002).
3. Kaplan A, Haenlein M. Siri, Siri, in my hand: who’s the fairest in the land?
On the interpretations, illustrations, and implications of artificial intelligence. Bus 26. L’Heureux A, Grolinger K, Elyamany HF, Capretz MAM. Machine learning
Horiz. (2019) 62:15–25. with big data: challenges and approaches. IEEE Access. (2017) 5:7776–97. doi:
10.1109/ACCESS.2017.2696365
4. Turing AM. Computing machinery and intelligence. Mind. (1950) 59:433–60.
doi: 10.1093/mind/LIX.236.433 27. Xu M, Zhang T, Li Z, Liu M, Zhang D. Towards evaluating the robustness of
deep diagnostic models by adversarial attack. Med Image Anal. (2021) 69:101977.
5. McCarthy J, Minsky M, Rochester N, Shannon CE. A proposal for the
doi: 10.1016/j.media.2021.101977
dartmouth summer research proiect on articial intelligence. AI Magaz. (2006)
27:12. doi: 10.1609/aimag.v27i4.1904 28. Jetley S, Lord NA, Namhoon L, Torr PHS. Learn to pay attention.
International conference on learning representations. arXiv [Preprint]. (2018).
6. Stonier, T. (1992). The Evolution of Machine Intelligence. Beyond Information,
doi: 10.48550/arXiv.1804.02391
1st Edn. London: Springer, p. 221.
29. Hassanin M, Anwar S, Radwan I, Khan F, Mian A. Visual attention methods
7. McCarthy J. What is Artificial Intelligence? (2004). Available online at: http:
in deep learning: an in-depth survey. arXiv [Preprint]. (2022). doi: 10.48550/arXiv.
//www-formal.stanford.edu/jmc/whatisai/whatisai.html (accessed July 2022)
2204.07756
8. Russell SJ, Norvig P. Artificial Intelligence: A Modern Approach. 4th ed. Boston,
30. Kim J, Lee S, Hwang E, Ryu KS, Jeong H, Lee JW, et al. Limitations of deep
MA: Pearson (2021). p. 1166
learning attention mechanisms in clinical research: empirical case study based
9. Bringsiord S, Govindaraiulu NS. Artificial Intelligence. The Stanford on the korean diabetic disease setting. J Med Internet Res. (2020) 22:e18418.
Encyclopedia of Philosophy. (2022). Available online at: https://plato.stanford.edu/ doi: 10.2196/18418
entries/artificial-intelligence/#Bib (accessed July 2022).
31. Schlapfer J, Wellens HJ. Computer-interpreted electrocardiograms: benefits
10. Kagiyama N, Shrestha S, Fario PD, Sengupta PP. Artificial intelligence: and limitations. J Am Coll Cardiol. (2017) 70:1183–92. doi: 10.1016/j.jacc.2017.07.
practical primer for clinical research in cardiovascular disease. J Am Heart Assoc. 723
(2019) 8:e012788. doi: 10.1161/JAHA.119.012788
32. Al’Aref SJ, Anchouche K, Singh G, Slomka PJ, Kolli KK, Kumar A, et al.
11. Hurwitz J, Kirsch D. Machine Learning for Dummies. IBM Limited ed. Clinical applications of machine learning in cardiovascular disease and its
Hoboken, NJ: John Wiley & Sons (2018). p. 432 relevance to cardiac imaging. Eur Heart J. (2019) 40:1975–86. doi: 10.1093/
eurheartj/ehy404
12. Sutton RS, Barto AG. Reinforcement Learning : An Introduction. Adaptive
Computation and Machine Learning. 2nd ed. Cambridge, MA: The MIT Press 33. Hannun AY, Rajpurkar P, Haghpanahi M, Tison GH, Bourn C, Turakhia
(2018). p. 548 MP, et al. Cardiologist-level arrhythmia detection and classification in ambulatory
electrocardiograms using a deep neural network. Nat Med. (2019) 25:65–9. doi:
13. Mahmood A, Shresta A. Review of deep learning algorithms and
10.1038/s41591-018-0268-3
architectures. IEEE Access. (2019) 7:53040–65. doi: 10.1109/ACCESS.2019.
2912200 34. Kannathal N, Acharya UR, Lim CM, Sadasivan P, Krishnan S. Classification
14. ProjectPro. Top 10 Deep Learning Algorithms in Machine Learning. (2022). of cardiac patient states using artificial neural networks. Exp Clin Cardiol. (2003)
Available online at: https://www.projectpro.io/article/deep-learning-algorithms/ 8:206–11.
443#mcetoc_1g5it6rql26 (accessed July 2022) 35. Zhao Q, Zhang L. ECG feature extraction and classification using wavelet
15. Dilsizian ME, Siegel EL. Machine meets biology: a primer on artificial transform and support vector machines. Proceedings of the 2005 International
intelligence in cardiology and cardiac imaging. Curr Cardiol Rep. (2018) 20:139. Conference on Neural Networks and Brain. Beijing (2005). p. 1089–92. doi: 10.
doi: 10.1007/s11886-018-1074-8 1109/ICNNB.2005.1614807
16. Srinivas S, Sarvadevabhatla RK, Mopuri KR, Prabhu N, Kruthiventi SSS, Babu 36. Sekeli S, Sandler B, Johnston E, Pollock KG, Hill NR, Gordon J, et al.
VA. Taxonomy of deep convolutional neural nets for computer vision. Front Robot Detecting undiagnosed atrial fibrillation in UK primary care: validation of a
AI. (2016) 2:36. doi: 10.3389/frobt.2015.00036 machine learning prediction algorithm in a retrospective cohort study. Eur J Prev
Cardiol. (2021) 28:598–605. doi: 10.1177/2047487320942338
17. Alzubaidi L, Zhang J, Humaidi AJ, Al-Dujaili A, Duan Y, Al-Shamma O, et al.
Review of deep learning: concepts, CNN architectures, challenges, applications, 37. Attia ZI, Noseworthy PA, Lopez-Jimenez F, Asirvatham SJ, Deshmukh
future directions. J Big Data. (2021) 8:53. doi: 10.1186/s40537-021-00444-8 AJ, Gersh BJ, et al. An artificial intelligence-enabled ECG algorithm for
the identification of patients with atrial fibrillation during sinus rhythm: a
18. Zhou D-X. Universality of deep convolutional neural networks. Appl Comput
retrospective analysis of outcome prediction. Lancet. (2019) 394:861–7. doi: 10.
Harm Analy. (2019) 48:787–94. doi: 10.1016/j.acha.2019.06.004
1016/S0140-6736(19)31721-0
19. Goodfellow IJ, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S,
38. Lu J, Hutchens R, Hung J, Bennamoun M, McQuillan B, Briffa T, et al.
et al. Generative adversarial nets. Advances in neural information processing
Performance of multilabel machine learning models and risk stratification
systems. arXiv [Preprint]. (2014). doi: 10.48550/arXiv.1406.2661
schemas for predicting stroke and bleeding risk in patients with non-valvular atrial
20. Skandarani Y, Lalande A, Afilalo J, Jodoin PM. Generative adversarial fibrillation. arXiv [Preprint]. (2022). doi: 10.48550/arXiv.2202.01975
networks in cardiology. Can J Cardiol. (2022) 38:196–203. doi: 10.1016/j.cjca.2021.
39. Lopez Perales CR, Van Spall HGC, Maeda S, Jimenez A, Latcu DG, Milman A,
11.003
et al. Mobile health applications for the detection of atrial fibrillation: a systematic
21. Amirrajab S, Abbasi-Sureshjani S, Al Khalil Y, Lorenz C, Weese J, Pluim review. Europace. (2021) 23:11–28. doi: 10.1093/europace/euaa139
J, et al. XCAT-GAN for synthesizing 3d consistent labeled cardiac MR images
40. Minchole A, Rodriguez B. Artificial intelligence for the electrocardiogram.
on anatomically variable XCAT phantoms. In: Medical Image Computing and
Nat Med. (2019) 25:22–3. doi: 10.1038/s41591-018-0306-1
Computer Assisted Intervention - MICCAI 2020 editor. MICCAI 2020. Lecture
Notes in Computer Science. (Vol. 12264), Cham: Springer (2020). doi: 10.1007/ 41. Halcox JPJ, Wareham K, Cardew A, Gilmore M, Barry JP, Phillips C, et al.
978-3-030-59719-1_13 Assessment of remote heart rhythm sampling using the alivecor heart monitor
to screen for atrial fibrillation: the REHEARSE-AF study. Circulation. (2017)
22. Zhang Z, Yang L, Zheng Y. Translating and segmenting multimodal medical
136:1784–94. doi: 10.1161/CIRCULATIONAHA.117.030583
volumes with cycle- and shape-consistency generative adversarial network.
Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern 42. Goldenthal IL, Sciacca RR, Riga T, Bakken S, Baumeister M, Biviano AB, et al.
Recognition. Salt Lake City, UT (2018). p. 9242–51. doi: 10.1109/CVPR.2018. Recurrent atrial fibrillation/flutter detection after ablation or cardioversion using
00963 the AliveCor KardiaMobile device: iHEART results. J Cardiovasc Electrophysiol.
(2019) 30:2220–8. doi: 10.1111/jce.14160
23. Che Y, Cheng Y, Zhai S, Sun Z, Liu Y. Boosting deep learning risk prediction
with generative adversarial networks for electronic health records. Proceedings of 43. Perez MV, Mahaffey KW, Hedlin H, Rumsfeld JS, Garcia A, Ferris T, et al.
the 2017 IEEE International Conference on Data Mining (ICDM). New Orleans, Large-scale assessment of a smartwatch to identify atrial fibrillation. N Engl J Med.
LA (2017). p. 787–92. doi: 10.1109/ICDM.2017.93 (2019) 381:1909–17. doi: 10.1056/NEJMoa1901183

Frontiers in Cardiovascular Medicine 19 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

44. Koulaouzidies G, Jadczyk T, Iakovidis DK, Koulaouzidis A, Bisnaire M, 63. Betancur J, Commandeur F, Motlagh M, Sharir T, Einstein AJ, Bokhari S, et al.
Charisopoulou D. Artificial intelligence in cardiology—a narrative review of Deep learning for prediction of obstructive disease from fast myocardial perfusion
current status. J Clin Med. (2022) 11:3910. doi: 10.3390/jcm11133910 SPECT: a Multicenter study. JACC Cardiovasc Imaging. (2018) 11:1654–63. doi:
10.1016/j.jcmg.2018.01.020
45. Benjamens S, Dhunnoo P, Mesko B. The state of artificial intelligence-based
FDA-approved medical devices and algorithms: an online database. NPJ Digit 64. Hu L-H, Betancur J, Sharir T, Einstein AJ, Bokhari S, Fish MB, et al. Machine
Med. (2020) 3:118. doi: 10.1038/s41746-020-00324-0 learning predicts per-vessel early coronary revascularization after fast myocardial
perfusion SPECT: results from multicentre REFINE SPECT registry. Eur Heart J
46. Alsharqi M, Woodward WJ, Mumith JA, Markham DC, Upton R, Leeson P.
Cardiovasc Imaging. (2020) 21:549–59. doi: 10.1093/ehjci/jez177
Artificial intelligence and echocardiography. Echo Res Pract. (2018) 5:R115–25.
doi: 10.1530/ERP-18-0056 65. Bustin A, Fuin N, Botnar RM, Prieto C. From compressed-sensing to artificial
intelligence-based cardiac MRI reconstruction. Front Cardiovasc Med. (2020)
47. Cannesson M, Tanabe M, Suffoletto MS, McNamara DM, Madan S, Lacomis
7:17. doi: 10.3389/fcvm.2020.00017
JM, et al. A novel two-dimensional echocardiographic image analysis system using
artificial intelligence-learned pattern recognition for rapid automated ejection 66. Seetharam K, Brito D, Farjo PD, Sengupta PP. The role of artificial intelligence
fraction. J Am Coll Cardiol. (2007) 49:217–26. doi: 10.1016/j.jacc.2006.08.045 in cardiovascular imaging: state of the art review. Front Cardiovasc Med. (2020)
7:618849. doi: 10.3389/fcvm.2020.618849
48. Knackstedt C, Bekkers SC, Schummers G, Schreckenberg M, Muraru D,
Badano LP, et al. Fully automated versus standard tracking of left ventricular 67. Bai W, Sinclair M, Tarroni G, Oktay O, Rajchl M, Vaillant G, et al. Automated
ejection fraction and longitudinal strain: the FAST-EFs multicenter study. J Am cardiovascular magnetic resonance image analysis with fully convolutional
Coll Cardiol. (2015) 66:1456–66. doi: 10.1016/j.jacc.2015.07.052 networks. J Cardiovasc Magn Reson. (2018) 20:65. doi: 10.1186/s12968-018-
0471-x
49. Zhang J, Gajjala S, Agrawal P, Tison GH, Hallock LA, Beussink-Nelson L, et al.
Fully automated echocardiogram interpretation in clinical practice. Circulation. 68. Bhuva AN, Bai W, Lau C, Davies RH, Ye Y, Bulluck H, et al. A multicenter,
(2018) 138:1623–35. doi: 10.1161/CIRCULATIONAHA.118.034338 scan-rescan, human and machine learning CMR study to test generalizability
and precision in imaging biomarker analysis. Circ Cardiovasc Imaging. (2019)
50. Madani A, Arnaout R, Mofrad M, Arnaout R. Fast and accurate view
12:e009214. doi: 10.1161/CIRCIMAGING.119.009214
classification of echocardiograms using deep learning. NPJ Digit Med. (2018) 1:6.
doi: 10.1038/s41746-017-0013-1 69. Swift AJ, Lu H, Uthoff J, Garg P, Cogliano M, Taylor J, et al. A machine
learning cardiac magnetic resonance approach to extract disease features and
51. Sengupta PP, Huang YM, Bansal M, Ashrafi A, Fisher M, Shameer K,
automate pulmonary arterial hypertension diagnosis. Eur Heart J Cardiovasc
et al. Cognitive machine-learning algorithm for cardiac imaging: a pilot study
Imaging. (2021) 22:236–45. doi: 10.1093/ehjci/jeaa001
for differentiating constrictive pericarditis from restrictive cardiomyopathy. Circ
Cardiovasc Imaging. (2016) 9:e004330. doi: 10.1161/CIRCIMAGING.115.004330 70. Chen C, Bai W, Davies RH, Bhuva AN, Manisty CH, Augusto JB,
et al. Improving the generalizability of convolutional neural network-based
52. Narula S, Shameer K, Salem Omar AM, Dudley JT, Sengupta PP. Machine-
segmentation on CMR images. Front Cardiovasc Med. (2020) 7:105. doi: 10.3389/
learning algorithms to automate morphological and functional assessments in 2D
fcvm.2020.00105
echocardiography. J Am Coll Cardiol. (2016) 68:2287–95. doi: 10.1016/j.jacc.2016.
08.062 71. Zhang Q, Burrage MK, Lukaschuk E, Shanmuganathan M, Popescu IA,
Nikolaidou C, et al. Toward replacing late gadolinium enhancement with artificial
53. Moghaddasi H, Nourian S. Automatic assessment of mitral regurgitation
intelligence virtual native enhancement for gadolinium-free cardiovascular
severity based on extensive textural features on 2D echocardiography videos.
magnetic resonance tissue characterization in hypertrophic cardiomyopathy.
Comput Biol Med. (2016) 73:47–55. doi: 10.1016/j.compbiomed.2016.03.026
Circulation. (2021) 144:589–99. doi: 10.1161/CIRCULATIONAHA.121.054432
54. Ouyang D, He B, Ghorbani A, Yuan N, Ebinger J, Langlotz CP, et al.
72. Vergani V, Razavi R, Puyol-Anton E, Ruijsink B. Deep learning for
Video-based AI for beat-to-beat assessment of cardiac function. Nature. (2020)
classification and selection of cine CMR images to achieve fully automated quality-
580:252–6. doi: 10.1038/s41586-020-2145-8
controlled CMR analysis from scanner to report. Front Cardiovasc Med. (2021)
55. Krittanawong C, Virk HUI, Bangalore S, Wang Z, Johnson KW, Pinotti R, 8:742640. doi: 10.3389/fcvm.2021.742640
et al. Machine learning prediction in cardiovascular diseases: a meta-analysis. Sci
73. Bazoukis G, Stavrakis S, Zhou J, Bollepalli SC, Tse G, Zhang Q, et al. Machine
Rep. (2020) 10:16057. doi: 10.1038/s41598-020-72685-1
learning versus conventional clinical methods in guiding management of heart
56. Dey D, Gaur S, Ovrehus KA, Slomka PJ, Betancur J, Goeller M, et al. failure patients- a systematic review. Heart Fail Rev. (2021) 26:23–34.
Integrated prediction of lesion-specific ischaemia from quantitative coronary CT
74. Aljaaf AJ, Al-Jumeily D, Hussain AJ, Dawson T, Fergus P, Al-Jumaily M.
angiography using machine learning: a multicentre study. Eur Radiol. (2018)
Predicting the likelihood of heart failure with a multi level risk assessment
28:2655–64. doi: 10.1007/s00330-017-5223-z
using decision tree. Proceedings of the 2015 Third International Conference
57. van Rosendael AR, Maliakal G, Kolli KK, Beecy A, Al’Aref SJ, Dwivedi A, on Technological Advances in Electrical, Electronics and Computer Engineering
et al. Maximization of the usage of coronary CTA derived plaque information (TAEECE). Beirut: (2015). p. 101–6. doi: 10.1109/TAEECE.2015.7113608
using a machine learning based algorithm to improve risk stratification; insights
75. Yang G, Ren Y, Pan Q, Ning G, Gong S, Cai G, et al. A heart failure diagnosis
from the CONFIRM registry. J Cardiovasc Comput Tomogr. (2018) 12:204–9.
model based on support vector machine. Proceedings of the 2010 3rd International
doi: 10.1016/j.jcct.2018.04.011
Conference on Biomedical Engineering and Informatics. Yantai: (2010). p. 1105–8.
58. Al’Aref SJ, Maliakal G, Singh G, van Rosendael AR, Ma X, Xu Z, et al. doi: 10.1109/BMEI.2010.5639619
Machine learning of clinical variables and coronary artery calcium scoring for
76. Attia ZI, Kapa S, Lopez-Jimenez F, McKie PM, Ladewig DJ, Satam G, et al.
the prediction of obstructive coronary artery disease on coronary computed
Screening for cardiac contractile dysfunction using an artificial intelligence-
tomography angiography: analysis from the CONFIRM registry. Eur Heart J.
enabled electrocardiogram. Nat Med. (2019) 25:70–4. doi: 10.1038/s41591-018-
(2020) 41:359–67. doi: 10.1093/eurheartj/ehz565
0240-2
59. Han D, Kolli KK, Gransar H, Lee JH, Choi SY, Chun EJ, et al.
77. Shameer K, Johnson KW, Yahi A, Miotto R, Li LI, Ricks D, et al. Predictive
Machine learning based risk prediction model for asymptomatic individuals
modeling of hospital readmission rates using electronic medical record-wide
who underwent coronary artery calcium score: comparison with traditional risk
machine learning: a case-study using mount sinai heart failure cohort. Pac Symp
prediction approaches. J Cardiovasc Comput Tomogr. (2020) 14:168–76. doi: 10.
Biocomput. (2017) 22:276–87. doi: 10.1142/9789813207813_0027
1016/j.jcct.2019.09.005
78. Vedomske M, Brown D, Harrison J. Random forests on ubiquitous data
60. Motwani M, Dey D, Berman DS, Germano G, Achenbach S, Al-Mallah
for heart failure 30-day readmissions prediction. Proceedings of the 2013 12th
MH, et al. Machine learning for prediction of all-cause mortality in patients
International Conference on Machine Learning and Applications. Miami, FL:
with suspected coronary artery disease: a 5-year multicentre prospective
(2013). p. 415–21. doi: 10.1109/ICMLA.2013.158
registry analysis. Eur Heart J. (2017) 38:500–7. doi: 10.1093/eurheartj/eh
w188 79. Zolfaghar K, Meadem N, Teredesai A, Roy SB, Chin S-C, Muckian B. Big data
solutions for predicting risk-of-readmission for congestive heart failure patients.
61. Oikonomou EK, Marwan M, Desai MY, Mancio J, Alashi A, Hutt Centeno
Proceedings of the 2013 IEEE International Conference on Big Data. Silicon Valley,
E, et al. Non-invasive detection of coronary inflammation using computed
CA: (2013). p. 64–71. doi: 10.1109/BigData.2013.6691760
tomography and prediction of residual cardiovascular risk (the CRISP CT study):
a post-hoc analysis of rospective outcome data. Lancet. (2018) 392:929–39. doi: 80. Mortazavi BJ, Downing NS, Bucholz EM, Dharmarajan K, Manhapra
10.1016/S0140-6736(18)31114-0 A, Li SX, et al. Analysis of machine learning techniques for heart failure
readmissions. Circ Cardiovasc Qual Outcomes. (2016) 9:629–40. doi: 10.1161/
62. Arsanjani R, Xu Y, Dey D, Vahistha V, Shalev A, Nakanishi R, et al. Improved
CIRCOUTCOMES.116.003039
accuracy of myocardial perfusion SPECT for detection of coronary artery disease
by machine learning in a large population. J Nucl Cardiol. (2013) 20:553–62. 81. Golas SB, Shibahara T, Agboola S, Otaki H, Sato J, Nakae T, et al. A machine
doi: 10.1007/s12350-013-9706-2 learning model to predict the risk of 30-day readmissions in patients with heart

Frontiers in Cardiovascular Medicine 20 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

failure: a retrospective analysis of electronic medical records data. BMC Med IEEE Antenn Propagat Magaz. (2016) 58:84–92. doi: 10.1109/MAP.2016.259
Inform Decis Mak. (2018) 18:44. doi: 10.1186/s12911-018-0620-z 4004
82. Kalscheur MM, Kipp RT, Tattersall MC, Mei C, Buhr KA, DeMets DL, 95. Stehlik J, Schmalfuss C, Bozkurt B, Nativi-Nicolau J, Wohlfahrt P,
et al. Machine learning algorithm predicts cardiac resynchronization therapy Wegerich S, et al. Continuous wearable monitoring analytics predict heart
outcomes: lessons from the COMPANION trial. Circ Arrhythm Electrophysiol. failure hospitalization: the LINK-HF Multicenter study. Circ Heart Fail. (2020)
(2018) 11:e005499. doi: 10.1161/CIRCEP.117.005499 13:e006513. doi: 10.1161/CIRCHEARTFAILURE.119.006513
83. Tokodi M, Schwertner WR, Kovacs A, Toser Z, Staub L, Sarkany A, et al. 96. Leslie D. Understanding Artificial Intelligence Ethics and Safety: A Guide for
Machine learning-based mortality prediction of patients undergoing cardiac the Responsible Design and Implementation of AI Systems in the Public Sector.
resynchronization therapy: the SEMMELWEIS-CRT score. Eur Heart J. (2020) London: The Alan Turing Institute (2019). doi: 10.5281/zenodo.3240529
41:1747–56. doi: 10.1093/eurheartj/ehz902
97. AlphaFold. AlphaFold Protein Structure Database. (2022). Available online at:
84. Cikes M, Sanchez-Martinez S, Claggett B, Duchateau N, Piella G, Butakoff https://alphafold.ebi.ac.uk/ (accessed July 2022).
C, et al. Machine learning-based phenogrouping in heart failure to identify
98. Bommasani R, Hudson DA, Adeli E, Altman R, Arora S, von Arx S, et al.
responders to cardiac resynchronization therapy. Eur J Heart Fail. (2019) 21:74–
On the opportunities and risks of foundation models. arXiv [Preprint]. (2018).
85. doi: 10.1002/ejhf.1333
doi: 10.48550/arXiv.2108.07258
85. Yao X, Rushlow DR, Inselman JW, McCoy RG, Thacher TD, Behnken
99. Nguyen A, Yosinski J, Clune J. Deep neural networks are easily fooled: high
EM, et al. Artificial intelligence-enabled electrocardiograms for identification of
confidence predictions for unrecognizable images. Proceedings of the 2015 IEEE
patients with low ejection fraction: a pragmatic, randomized clinical trial. Nat
Conference on Computer Vision and Pattern Recognition (CVPR). Boston, MA:
Med. (2021) 27:815–9. doi: 10.1038/s41591-021-01335-4
(2015). p. 427–36. doi: 10.1109/CVPR.2015.7298640
86. Hopkins CB, Suleman J, Cook C. An artificial neural network for the
100. Gunning D, Stefik M, Choi J, Miller T, Stumpf S, Yang GZXAI. Explainable
electrocardiographic diagnosis of left ventricular hypertrophy. Crit Rev Biomed
artificial intelligence. Sci Robot. (2019) 4:eaay7120. doi: 10.1126/scirobotics.
Eng. (2000) 28:435–8. doi: 10.1615/critrevbiomedeng.v28.i34.140
aay7120
87. Kagiyama N, Piccirilli M, Yanamala N, Shrestha S, Farjo PD, Casaclang-
101. Samek W, Montavon G, Vedaldi A, Hansen LK, Muller K-R. Explainable AI:
Verzosa G, et al. Machine learning assessment of left ventricular diastolic function
Interpreting, Explaining and Visualizing Deep Learning. Berlin: Springer Nature
based on electrocardiographic features. J Am Coll Cardiol. (2020) 76:930–41.
(2019).
doi: 10.1016/j.jacc.2020.06.061
102. Briganti G, Le Moine O. Artificial intelligence in medicine: today and
88. Thompson WR, Reinisch AJ, Unterberger MJ, Schriefl AJ. Artificial
tomorrow. Front Med. (2020) 7:27. doi: 10.3389/fmed.2020.00027
intelligence-assisted auscultation of heart murmurs: validation by virtual clinical
trial. Pediatr Cardiol. (2019) 40:623–9. doi: 10.1007/s00246-018-2036-z 103. Cate FH, Lynskey Kuner C, Lynskey O, Millard C, Ni Loideain N,
Svantesson DJB. Expanding the artificial intelligence-data protection debate. Int
89. Oikonomou EK, Van Dijk D, Parise H, Suchard MA, de Lemos J, Antoniades
Data Privacy Law. (2019) 8:289–92.
C, et al. A phenomapping-derived tool to personalize the selection of anatomical
vs. functional testing in evaluating chest pain (ASSIST). Eur Heart J. (2021) 104. Food and Drug Administration. Proposed Regulatory Framework for
42:2536–48. doi: 10.1093/eurheartj/ehab223 Modifications to Artificial Intelligence/Machine Learning (AI/ML)-Based Software
as a Medical Device (SaMD) - Discussion Paper and Request for Feedback. Silver
90. Lin S, Li Z, Fu B, Chen S, Li X, Wang Y, et al. Feasibility of using deep
Spring, MD: Food and Drug Administration (2019).
learning to detect coronary artery disease based on facial photo. Eur Heart J.
(2020) 41:4400–11. doi: 10.1093/eurheartj/ehaa640 105. European Commission. Regulatory Framework Proposal on Artificial
Intelligence. Brussels: European Commission (2022).
91. Sun Y, Yang Y-Y, Wu B-J, Huang P-W, Cheng S-E, Wu B-F, et al.
Contactless facial video recording with deep learning models for the detection 106. Kelly CJ, Karthikesalingman A, Suleyman M, Corrado G, King D. Key
of atrial fibrillation. Sci Rep. (2022) 12:281. doi: 10.1038/s41598-021-03 challenges for delivering clinical impact with artificial intelligence. BMC Med.
453-y (2019) 17:195. doi: 10.1186/s12916-019-1426-2
92. Baker SB, Xiang W, Atkinson I. Internet of things for smart healthcare: 107. Kurzweil, R. (Ed.) (1992). The Age of Intelligent Machines. Cambridge, MA:
technologies, challenges, and opportunities. IEEE Access. (2017) 5:26521–44. doi: MIT press, p. 580.
10.1109/ACCESS.2017.2775180 108. Poole, D., Mackworth, A. K., and Goebel, R. (1998). Computational
93. Islam SMR, Kwak D, Kabir MH, Hossain M, Kwak KS. The internet of Intelligence: A Logical Approach, 1st Edn. New York, NY: Oxford University Press.
things for health care: a comprehensive survey. IEEE Access. (2015) 3:678–708. 109. Haugeland, J. (1985). Artificial Intelligence: The Very Idea, 1st Edn.
doi: 10.1109/ACCESS.2015.2437951 Cambridge, MA: MIT press, p. 287.
94. Wolgast G, Ehrenborg C, Israelsson A, Helander J, Johansson E, Manefjord 110. Charniank, E., and McDermott, D. (1985). Introduction to Artificial
H. Wireless body area network for heart attack detection [Education Corner]. Intelligence, 1st Edn. Boston, MA: Addison-Wesley, p. 720.

Frontiers in Cardiovascular Medicine 21 frontiersin.org

Karatzia et al. 10.3389/fcvm.2022.945726

Glossary
ACM, all-cause mortality; ACS, acute coronary syndrome; AI, artificial intelligence; AF, atrial fibrillation; AHA, American
heart association; ALVD, asymptomatic left ventricular dysfunction; AMI, acute myocardial infarction; AUC, area under the
curve; CACS, coronary artery calcium scoring; CAD, coronary artery disease; CNN, convolutional neural network; CRT, cardiac
resynchronisation therapy; CT, computerised tomography; CTA, cardiac computed tomography angiography; CTP, computed
tomography myocardial perfusion; CVD, cardiovascular disease; DD, diastolic dysfunction; DL, deep learning; DUN, deep unified
network; ECG, electrocardiogram; ECR, early coronary revascularisation; EF, ejection fraction; EHR, electronic health record; ESC,
European society cardiology; FAI, fat attenuation index; FDA, food and drug administration; FNNN, feed forward neural network;
HCM, hypertrophic cardiomyopathy; HF, heart failure; HFpEF, heart failure with preserved ejection fraction; HR, heart rate; ICA,
invasive coronary angiography; IoT, internet of things; LBBB, left bundle branch block; LR, logistic regression; LS, longitudinal
strain; LV, left ventricle; LVDF, Left ventricular diastolic function; LVEDV, left ventricle end diastolic volume; LVEF, left ventricular
ejection fraction; LVESV, left ventricle end systolic volume; LVH, left ventricular hypertrophy; LVM, left ventricle mass; MACE, major
adverse cardiac event; ML, machine learning; MLP, multiplayer perceptron; MNN, modular neural network; MPI, myocardial perfusion
imaging; MR, mitral regurgitation; PAH, pulmonary arterial hypertension; PCI, percutaneous coronary intervention; PVC, premature
ventricular contraction; RBBB, right bundle branch block; RBFN, radial basis function network; RCT, randomised controlled trial;
RF, random forests; RNN, recurrent neural network; ROC, receiver operator characteristic; RV, right ventricle; RVEDV, right ventricle
end diastolic volume; RVESV, right ventricle end systolic volume; SPECT, single-photon emission computed tomography; SR, sinus
rhythm; SVM, support vector machine; TPD, total perfusion deficit.

Frontiers in Cardiovascular Medicine 22 frontiersin.org