Deep Learning For Cardiovascular Medicine: A Practical Primer

European Heart Journal (2019) 00, 1–15 CLINICAL REVIEW
doi:10.1093/eurheartj/ehz056 Frontiers in cardiovascular medicine
Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

Deep learning for cardiovascular medicine: a
practical primer
Chayakrit Krittanawong1,2*, Kipp W. Johnson3, Robert S. Rosenson2, Zhen Wang4,5,
Mehmet Aydar6, Usman Baber2, James K. Min7, W.H. Wilson Tang8,9,10,
Jonathan L. Halperin2, and Sanjiv M. Narayan11*
1
Department of Internal Medicine, Icahn School of Medicine at Mount Sinai, 1 Gustave L. Levy Pl, New York, NY 10029, USA; 2Department of Cardiovascular Diseases, Icahn
School of Medicine at Mount Sinai, Mount Sinai Hospital, Mount Sinai Heart, New York, NY 10029, USA; 3Department of Genetics and Genomic Sciences, Institute for Next
Generation Healthcare, Icahn School of Medicine at Mount Sinai, New York, NY 10029, USA; 4Robert D. and Patricia E. Kern Center for the Science of Health Care Delivery,
Mayo Clinic, Rochester, MN 55905, USA; 5Division of Health Care Policy and Research, Department of Health Sciences Research, Mayo Clinic, Rochester, MN 55905, USA;
6
Department of Computer Science, Kent State University, Kent, OH 44240, USA; 7Department of Radiology, New York-Presbyterian Hospital and Weill Cornell Medicine, New
York, NY 10065, USA; 8Department of Cardiovascular Medicine, Heart and Vascular Institute, Cleveland Clinic, OH 44195, USA; 9Department of Cellular and Molecular
Medicine, Lerner Research Institute, Cleveland, OH 44195, USA; 10Center for Clinical Genomics, Cleveland Clinic, Cleveland, OH 44195, USA; and 11Cardiovascular Institute
and Department of Cardiovascular Medicine, Stanford University Medical Center, Stanford, CA 94035, USA
Received 29 September 2018; revised 2 November 2018; editorial decision 10 January 2019; accepted 22 January 2019
Deep learning (DL) is a branch of machine learning (ML) showing increasing promise in medicine, to assist in data classification, novel disease
phenotyping and complex decision making. Deep learning is a form of ML typically implemented via multi-layered neural networks. Deep learning
has accelerated by recent advances in computer hardware and algorithms and is increasingly applied in e-commerce, finance, and voice and image
recognition to learn and classify complex datasets. The current medical literature shows both strengths and limitations of DL. Strengths of DL in-
clude its ability to automate medical image interpretation, enhance clinical decision-making, identify novel phenotypes, and select better treatment
pathways in complex diseases. Deep learning may be well-suited to cardiovascular medicine in which haemodynamic and electrophysiological in-
dices are increasingly captured on a continuous basis by wearable devices as well as image segmentation in cardiac imaging. However, DL also
has significant weaknesses including difficulties in interpreting its models (the ‘black-box’ criticism), its need for extensive adjudicated (‘labelled’)
data in training, lack of standardization in design, lack of data-efficiency in training, limited applicability to clinical trials, and other factors. Thus, the
optimal clinical application of DL requires careful formulation of solvable problems, selection of most appropriate DL algorithms and data, and
balanced interpretation of results. This review synthesizes the current state of DL for cardiovascular clinicians and investigators, and provides
technical context to appreciate the promise, pitfalls, near-term challenges, and opportunities for this exciting new area.
...................................................................................................................................................................................................
Keywords Big data • Artificial intelligence • Deep learning • Cardiovascular medicine • Precision medicine
.. heterogeneous phenotypes.1–3 Deep learning (DL) is a branch of

Introduction ..
.. artificial intelligence (AI) that combines computer science, statis-
..
The practice of cardiovascular medicine routinely requires man- .. tics and decision theory to find patterns in complex and often volu-
agement of conditions as complex as heart failure with reduced .. minous data. In general, DL is a type of machine learning (ML) that
..
(HFrEF) or preserved (HFpEF) ejection fraction, multivessel cor- .. typically utilizes multi-layered neural networks (Figure 1). Deep
onary disease, complex arrhythmias, sudden cardiac arrest, cardio- .. learning has been already shown to outperform experts and other
..
vascular diseases (CVDs) during pregnancy, or congenital heart .. ML strategies in areas as diverse as voice recognition, image classifi-
disease. Despite advances in each of these areas, significant clinical .. cation, commerce and game playing,4–7 and there is anticipation in
..
challenges remain. Many challenges relate to the complexity of .. that DL could similarly disrupt clinical decision-making by integrat-
integrating data from multiple modalities, making actionable pre- .. ing complex data streams, making ‘intelligent inferences’ and ultim-
..
dictions and distilling these solutions to individual patients with .. ately personalizing therapy.
.
* Corresponding authors. Tel: þ1 212 523 4000, Fax: þ1 212 523 8605, Email: [email protected]; Tel: þ1 650 724 1850, Fax: þ1 650 725 7568,
Email: [email protected]
C The Author(s) 2019. For permissions, please email: [email protected].
Published on behalf of the European Society of Cardiology. All rights reserved. V
2 C. Krittanawong et al.

Figure 1 Relationship of deep learning to clinical and translational medicine. Venn diagrams show deep learning as one type of machine learning,
within the scope of artificial intelligence. Statistical methods are applied across clinical and translational science, and the form known as statistical
learning theory has overlap with machine learning. Automated decision making is often used in clinical practice. Deep learning may extend statistical
approaches in some key areas by analysing large multivariate datasets, which often show complex interactions, in which simple hypotheses are diffi-
cult to formulate. Deep learning has been successful in medical image recognition (e.g. electrocardiogram, echocardiogram, and magnetic resonance
imaging) and holds the promise of enhancing clinic decision making.
There are several parallels, as well as divergences, between trad-

.. evaluate the possibility of overfitting, a limitation of DL is that it
..
itional statistical methods, ML and DL. Statistical approaches typic- .. relies to a much greater extent on empirical validation.
..
ally test hypotheses or estimate parameters and emphasize .. To date, cardiovascular applications of DL have been promis-
inference based upon statistical sampling of a population. In these .. ing8–10 although many challenges remain. In particular, it is critical to
..
cases, traditional statistics can be as effective as DL even in large .. select the right tools for each specific problem and dataset in CVD.
‘big data’ applications; for instance in genome-wide association .. This practical review is designed to enable the reader to understand
..
studies (GWAS), a single loci meeting a Bonferroni-adjusted P- .. and evaluate applications of DL to cardiovascular medicine or re-
value (e.g. P < 5 10-8) can identify important traits. However, if .. search. We discuss the historical development of DL, definitions, re-
..
simple hypotheses are less readily formulated due to complex .. view the current literature to recognize optimal applications,
interactions, for instance if GWAS yields multiple concurrent ‘hits’, .. summarize the design and interpretation of DL studies, discuss cur-
..
ML may be better suited because it does not require specific .. rent challenges and pitfalls, and future directions.
hypotheses, and can analyse varied data elements with complex ..
..
interactions. Deep learning strategies generally attempt to use as ..
..
much information as is available in a dataset (e.g. every pixel in .. History and definitions
echocardiography images) in order to generate novel features to ..
be used for downstream analysis. Statistical methods and DL are .. Artificial intelligence is the field of computer science broadly focused
..
both influenced by aberrations in sample data and may suffer from .. on teaching computers to learn complex tasks and make predictions.
overfitting. While statistical methods include approaches to
.. Early AI applications focused on hand-developing complex decision
Deep learning for cardiovascular medicine 3

Figure 2 Types of machine learning in cardiovascular science. (A) Supervised learning uses inputs (e.g. electrocardiograms) each with a label
(‘ground truth’, and a diagnosis of atrial fibrillation or not atrial fibrillation). Machines are iteratively ‘trained’, using direct feedback for multiple inputs,
until their output matches the ground truth. Trained machines can then classify unknown (test) electrocardiograms. One misclassification is shown.
(B) Unsupervised learning uses unlabelled data, ideally in large quantities, to identify novel patterns. In this example, QRS indices identified novel phe-
notypes (‘clusters’) for hypertrophic cardiomyopathy with distinct outcomes (Ref: Lyon et al.26). (C) Reinforcement learning uses models developed
from psychological training applied to gaming, but infrequently to medicine. An agent, e.g. a clinical decision-making tool, performs an action At (e.g.
which therapy for non-ST-segment elevation myocardial infarction best reduces mortality? (1) non-invasive, (2) early invasive, and (3) mixed) that
alters the environment (e.g. biomarker response or patient outcomes). A Rt reward is then given (e.g. higher survival rate) that alters the state St. This
process is iterated with the intention of moving State Stþ1 closer to the desired goal (i.e. improved outcomes).
rules for computers to follow but, due to the complexity of human

.. and can identify patterns and make decisions. Ideally, this can happen
..
decision making, this was by-and-large not successful. Instead ML, and .. with minimal human intervention. Machine learning can be further
particularly DL, have emerged as more promising. Machine learning
.. subdivided into supervised, unsupervised, or reinforcement learning.
..
analyses data in ways that automate the construction of analytical .. Figure 2 summarizes categories of ML and DL in the context of
..
models and decision rules, developing systems that learn from data . emerging applications in cardiovascular science. Supervised learning

Figure 3 Neural network design to classify atrial fibrillation from the electrocardiogram. Continuous electrocardiogram voltage points (red dots,
arrows) are fed to ‘input neurons’ (x0, x1, x2, . . ., xm), which are coded as software objects. Hidden neurons within this three-layer network (h0, h1,
h2, . . ., hn) connect input and output layer neurons (here, two neurons) by numerical weights (w). Deep learning typically uses multiple hidden layers,
as shown here. The output indicates atrial fibrillation (y1; correct, red) or non-atrial fibrillation (y0). If the output is correct for that electrocardiogram
input, weights are strengthened; else they are reduced. This process is iterated during training on multiple input electrocardiograms. The trained net-
work can then be tested on new (unseen) electrocardiograms. Other designs could accept categorical variables (age, gender) or mixed data types.
..
identifies patterns in large amounts of data that are typically anno- .. Implementing deep learning in
tated (‘labelled’) by humans, such as the presence or absence of ..
.. cardiovascular applications
reduced systolic function on an echocardiogram or atrial fibrillation ..
(AF) on an electrocardiogram (ECG). Supervised learning may be ..
.. Hardware and software considerations
implemented using neural networks, long-used for medical pattern ..
recognition in cardiology,11,12 neuroscience,13–15 and other fields yet .. Historically, ML was computationally expensive and performed by
.. scientists using supercomputers or high-end workstations with multi-
still limited in clinical use. Unsupervised learning, on the other hand, ..
analyses large amounts of unlabelled data to identify hidden patterns
.. core processors. However, due to the highly parallelizable nature of
.. DL algorithms, GPUs designed for gaming have enabled DL to be
or natural structure in data,16 which greatly increases the volume of ..
data that can be analysed (e.g. from large electronic medical records)
.. performed on desktop machines. Although professional-quality
.. GPUs are still relatively expensive, DL can be performed using cloud
at the potential cost of data quality and interpretability. ..
Reinforcement learning trains software to make decisions that maxi-
.. services by services such as Amazon AWS or Google Cloud. Deep
.. learning software packages are almost uniformly open-source, which
mize a ‘reward’ function,17 which may address a clinical problem (e.g. ..
improve ST-segment elevation myocardial infarction outcomes, or
.. means they are freely available with few constraints for academic
.. research. Furthermore, many DL pre-trained models can be down-
reduce error in ECG diagnosis). ..
.. loaded and repurposed for new tasks, including models such as
Deep learning is a specific type of ML inspired by the way that .. AlexNet,20 VGG Network,21 InceptionNet,22 and ResNet,23 and will
the human brain processes data, and enabled by hardware advan- ... be further described in following sections. In fact, downloading pre-
ces such as graphics processing units (GPUs),16,18,19 vast cata- ..
.. trained models and repurposing them for new datasets avoids much
logues of labelled data, and advances in computer science theory. .. of the time consuming and computationally expensive steps of DL.
To date, most implementations in medicine have used convolu- ..
.. Table 1 demonstrates step-by-step an example DL process in cardio-
tional neural networks (CNNs). Following the ‘AI winter’ from ..
the 1980s, when early rule-based and neural network applications .. vascular medicine.
..
were limited by hardware and algorithmic constraints, DL has ..
accelerated supervised, unsupervised, and reinforcement learning. .. Selecting a deep learning software
..
In 2016, DeepMind’s AlphaGo Zero4,5 beat the world champion .. package and modelling strategy
in the ancient Chinese game of Go, and deep Q-learning6 proved .. The first practical step for DL is to choose an appropriate software
..
as accurate as a professional human player in 49 interactive video .. package to work with such as Keras, Tensorflow, or others. Keras
games. In 2018, DL applications by DeepMind rivalled humans in .. is often used as a starting point, as it can be used in a relatively
..
the 3D multiplayer videogame Quake III Arena.7 Figure 3 provides .. straightforward fashion with high-level programming languages,
a schematic of how neural networks, a common basic architec-
.. most commonly Python. Supplementary material online, Table S1
..
ture for DL, could be used to classify an ECG. . summarizes some platforms and related programming languages for
Table 1 A guide to approach of the deep learning applications in cardiovascular medicine research and clinical
practice

Step How to
....................................................................................................................................................................................................................
Identify research questions and A supervised learning problem (e.g. predict labelled outcomes) vs. unsupervised learning problem (e.g. identify or classify
outcomes new phenotypes of HFpEF or new genotypes of PAD)
• If supervised, is it a regression task (e.g. prediction of the cost of care for a PCSK9 inhibitor treatment) or classifi-
cation (e.g. predict if a given patient has a disease or not)
Data selection Public databases vs. EHR databases vs. Registry
• Identify limitations of databases and attempt to replace with rational variables (i.e. no lab values, no vital signs be-
tween admission and discharge, no medications, no specific ICD codes for MitraClip)
• Identify appropriate methodologies for database (e.g. do we realistically have enough data to attempt DL? If not,
are we better off selecting an alternate approach to our problem?)
Hardware selection Computer cluster vs. workstation with GPUs vs. cloud computing services
• Can we build or buy a computer cluster or workstation with GPUs? In the long run, this will be much cheaper
but with much greater expense upfront. Should we instead use cloud computing services (i.e. Google Cloud or
Amazon Web Services), which have little-to-no upfront cost but can cost more in the long run
Data preparation (1) De-identify data if needed
(2) Quality control of data—assessment of missingness and verification that our dataset contains what it should. Identify
mechanism of missing data and then data imputations for non-ignorable missing data
(3) Denoising of images/video/textures or variant calling in NGS data
(4) Exploratory analysis (summarization, visualization, identify structure of data/relationship between variables)
Feature selection In general, DL requires little a priori feature selection. If the input dataset is highly multidimensional, strategies such as vec-
tor embedding may be required first in order to pass features to other DL models
Data splitting Design and justify the proportion of training, validation and testing in the dataset (i.e. 70/10/20 or 80/10/10 or 60/20/20)
Modelling selection (1) One should always see if the task can be accomplished with a simple model or standard statistical approaches (can we
simply apply linear/logistic regression or polynomial regression to our dataset? If so, does it perform adequately?)
(2) If simpler algorithms do not work, more complex strategies such as DL may be needed. In general, reuse of pre-
trained models using transfer learning is preferred since many pre-trained models are well validated in a large database
and their performance characteristics and limitations are known
(3) If no pre-trained models are available for specific type of data or research questions, develop new algorithms based
on highest achievement model for specific type of data
Technical details for model Know some DL technical terms to communicate with data scientists or programmers and understand the process (learning
rate selection, tuning hyperparameter, batch dropout and normalization, regularization strategies, loss function selection,
and network optimization)
Evaluation of model discrimination Report ROC curve, C-statistics, NPV, PPV, sensitivity, and specificity
Evaluation of model calibration Compare with standard statistical approaches (i.e. multivariable regression), goodness-of-fit, calibration plots, or the deci-
sion-curve analysis
Ground truth Compare with human experts (cardiologists, electrophysiologist, primary care physicians)
Publication and transparency Share codes with journal (i.e. online supplements) or public space (i.e. Githut, bioRxiv). DL methodologies should be clear-
ly explained in details. Consider strategies for computational anonymization
Generalization and replication results Test with different datasets in a different population
Clinical trials To minimize risks or errors, testing prediction models in clinical trials is recommendated
Meta-analysis Meta-analysis of DL is needed to assess publication bias and heterogeneity
AI Guidelines With clinical trial and meta-analytic results, professional societies have to develop guidelines to regulate DL in clinical deci-
sion making or predictive analytics in clinical practice
Implement in clinical practice Start implementing in clinical practice and monitor the results closely
AI, artificial Intelligence; DL, deep learning; EHR, electronic health record; GPU, graphics processing unit; HFpEF, heart failure with preserved ejection fraction; NGS, next-gen-
eration sequencing; NPV, negative predictive value; PAD, peripheral arterial disease; PCSK9, proprotein convertase subtilisin/kexin type 9; PPV, positive predictive value; ROC
curve, receiver-operating characteristic curve.
DL. Second, one should choose a DL model appropriate for the .. tailored to investigators’ problems/datasets by modifying outcomes
..
problem under consideration, which may be pre-existing (pre- .. related layers [i.e. last few layers or final layer (softmax layer)] to
trained models) or require development of a custom model (novel
.. optimize their results. This concept is called transfer learning.24 For
..
models). In general, due to time restrictions, and computational re- .. example, if a pre-trained model was trained to accurately predict
quirement, pre-trained models are mostly used first, and then
.. diastolic dysfunction using global longitudinal strain in HFrEF
..
patients in a large database, we might be able to use resultant .. intrinsicoid deflection.26 Since a key advantage of DL is to learn com-
knowledge to conduct the prediction of diastolic dysfunction in .. plex features, overly complex preprocessing of data into features
..
HFpEF patients. There are several pre-trained models such as .. may also impede performance. One general goal of DL may be sum-
AlexNet,20 which recognizes visual patterns directly from pixel .. marized as letting the algorithm automatically perform feature selec-

..
images with minimal preprocessing, the VGG Network21 that per- .. tion for the user, instead of the user attempting to manually engineer
formed well in the ImageNet Large Scale Visual Recognition .. features.
..
Challenge in 2014, or other models (e.g. InceptionNet22). When ..
validated and pre-trained models are unavailable for specific appli- ..
.. Missing data imputation
cations, which may be common in cardiovascular medicine at this .. Similar to statistical tasks, the performance of DL can be highly sensi-
early stage, custom models may be required. Each will require de- ..
.. tive to missing data. The cut-off to discard features where entries are
sign (e.g. the number of layers, nodes, learning rate, and so on in .. missing remains debated. Decisions on how to treat missing data can
Figure 3), validation and tuning.4,25 Examples are provided in ..
.. be made by evaluating if the presence or absence of specific elements
Supplementary material online, Table S2. However, transfer learning .. correlate with desired outcomes or predictors. Those data that are
is challenging in medicine as there are differences in datasets, i.e. ..
.. correlated are ‘non-ignorable’, those that are not correlated may be
sources of data (non-medical videos vs. echocardiographic videos), .. ‘ignorable’.27 Ignorable missing data includes missing completely at
data quality, vendor or softwares. There is no standardized guide-
..
.. random (no relationship to any variables) and missing at random (a
line for when it is appropriate to use transfer learning between ..
datasets, and this often requires empirical trial-and-error
.. systematic relationship between missing values and observed data).
.. Identifying these types of missing data is a crucial before data imput-
approaches. ..
Learning rate is an important concept that determines how the
.. ation.28 Several strategies exist for imputation. One approach to ac-
.. count for systematic data omission is to insert a label for those
network adjusts weights during training based on correct or incor- ..
.. elements. For instance, since individuals too sick (e.g. with elevated
rect decisions. Figure 4 shows appropriate learning rates in a well- .. serum creatinine) to receive a cardiac computed tomography (CT)
trained network, and how inappropriate learning rates (too high) ..
.. may have worse outcome than those who do undergo cardiac CT, a
may impede network training. Figure 5 shows how complex cardio- .. label that indicates ‘too sick’ would likely still be prognostic for out-
vascular data that are not readily separated by a simple cut point or ..
.. come in a model missing CT data in sick patients. Here, CT data is
threshold (i.e. cannot be partitioned by a linear classifier) can be .. non-ignorable, but training can proceed with this label. In other cases,
transformed and now readily separated. ..
.. training can proceed with insertion of an imputed value, multiple
Finally, a major strength of DL is its ability to infer and classify data .. imputations (i.e. MICE or missForest R packages), or expected values
outside its original training set (i.e. generalization). To generalize well, ..
.. from the literature. One class of DL algorithms called autoencoders
DL strategies must avoid overfitting to the training data. Unlike statis- .. have been shown to produce best-in-class results for missing data
tical methods, for which strategies exist to avoid overfitting, in DL, ..
.. imputation.29
this is often empirical and trial-and-error. The first and best approach ..
is usually to increase power by adding data. Another common tech- ..
.. Training and expected results
nique in DL is data augmentation, i.e. increasing the number of train- ..
ing samples using the same raw data. In image analysis, for example, it .. Datasets should be separated into distinct partitions for training, test-
.. ing, and validation. One area of future improvement in DL is in con-
is common to rotate, invert, or skew an image if it does not distort ..
data or change its output class (e.g. rotation would not alter hypoki-
.. sidering data containing repeated measures or correlations (time
.. series, or where observations cluster by patient). However, this is an
nesis on echocardiography or the location of a space-occupying le- ..
sion), but may be unacceptable in traditional biostatistics. Other
.. active area of research within the DL community.30,31 More informa-
.. tion on sampling by patient for correlated data (i.e. repeated meas-
strategies include cross-validation of training data, and transformation ..
to reduce data complexity. In complex DL architectures (e.g. several
.. ures) should also be included. The specific training, testing, and
..
hidden layers in a neural network, Figure 3), reducing the numbers of .. validation proportions used in DL often vary depending upon the
neurons or layers may be effective. The ultimate test of generalization
.. task at hand, and in general selecting one strategy is more of an ‘art’
..
is prediction in a new dataset. .. than ‘science’. 70/10/20 or 80/10/10 or 60/20/20 splits are common,
.. but no standard methodologies exist to determine optimum propor-
..
.. tions although each should come from similar data distributions to
Data preparation .. prevent mismatched proportions between sites. Training strategies
Data preparation must be tailored to each specific DL architecture ..
.. are thus largely empirical with no standardized approaches.
to optimize performance and requires feature selection and imput- ..
ation of missing data.
..
.. Novel deep learning techniques
..
.. Anticipated advances in DL may mitigate some concerns of empiri-
Feature selection .. cism and lack of a theoretical framework. Capsule-based neural net-
..
This involves transforming raw data to features that reduce dimen- .. work,32,33 meta-learning (learning how the network learns),34–36
sionality of the data, but that still represent the relevant clinical or .. DeepMath (learn mathematical proofs),37 or self-play (two agents
..
physiological question, and may reduce the risk of overfitting. In .. learn by win and loss)34 are promising innovations. Studies to opti-
Figure 3, input data were digitized ECG voltage-time series, while in
.. mize learning rate, network architectures, and activation functions
..
Figure 2B features were used instead such as QRS duration and the . may improve the learning efficiency of DL. There are efforts to adapt

Figure 4 Impact of deep learning design on learning: effect of learning rate. (A) Efficient learning. Cost function (network error) gradually descends
(‘Gradient descent’) to achieve the optimal point (called local minimum) as a function of weight. (B) Learning rate is too high, so that the cost function
overshoots the minimum and oscillates. This network design may not be trained effectively for this problem. (C) Gradient descent examining two var-
iables on cost function simultaneously.
several non-neural network-based methods for DL (Supplementary

.. interpreting cardiac images rapidly and consistently, DL may circum-
..
material online, Table S2). Support vector machines, for instance, are .. vent clinical limitations of fatigue or distraction, variable inter- and
effective for high-dimensional data38,39 and could be useful in cardio-
.. intra-observer interpretation, and time-consuming interpretation of
..
vascular DL with diverse datasets.40,41 Enriching the theoretical .. large datasets.
framework for DL may improve our ability to interpret their
.. Zreik et al.55 applied DL to automatically identify significant coron-
..
conclusions. .. ary artery stenosis in rest coronary CT angiograms in 166 patients.
..
.. Compared with matched invasive fractional flow reserve (FFR) meas-
.. urements, the network produced a c-statistic of 0.74 ± 0.02 with spe-
..
Deep learning applications in .. cificities of 77%, 71%, and 59% at sensitivities of 60%, 70%, and 80%,
.. respectively, providing a possible alternative to invasive FFR.55
cardiovascular medicine ..
.. Betancur et al.56 applied DL to single-photon emission computed
.. tomography myocardial perfusion imaging in 2619 consecutive
Table 2 summarizes early DL applications in cardiovascular medi- ..
cine,42–51 while Supplementary material online, Table S3 lists DL stud- .. patients at exercise (38%) or pharmacological stress. After
.. 3.2 ± 0.6 years follow-up, DL better predicted major adverse cardiac
ies in other disciplines. ..
.. events (MACE; 9.1% of patients overall) for DL using imaging with
.. stress test data than imaging data alone (area under the receiver-
Imaging for ischaemic and structural ..
.. operating characteristic curve: 0.81 vs. 0.78; P < 0.01), and both were
heart disease .. superior to existing assessments. Motwani et al.57 used DL on CT
..
Deep learning has been used to classify images in many medical spe- .. angiography in 10 300 patients with suspected coronary disease to
cialties,52–54 recently extended to cardiac imaging (Table 2). By
.. improve prediction of 5 year all-cause mortality over existing CT or

Figure 5 Classifying complex data. (A) Transforming data to enable linear separation of non-linearly separable raw data. Raw non-linear data are
transformed by mapping functions that may include time, frequency, or other operations. This projects them into higher-dimensional parameters
space in which they are now linearly separable. One example is classifying patients with heart failure with preserved ejection fraction whose response
to beta-blockers may vary due to obesity, atrial fibrillation, left ventricular hypertrophy, diabetes, or other factors. Data transformation to a higher-di-
mensional space now enables a simple partitioning process. (B) Bias–variance tradeoff. Model with high bias (straight line), when a straight line could
not classify appropriately (here, between atrial fibrillation and normal sinus rhythm) in both training dataset (5.B.a) and testing dataset (5.B.b). This
leads to prediction errors on other datasets (low variance - frequent errors). In contrast, model with low bias (i.e. due to overtraining) when data is
fitted well in training set (5.B.c), but not in testing set (5.B.d), leading to reduced generalization (high variability due to difference between training and
validation sets).
..
clinical indices as supported by others.58 In a novel study applying DL .. echocardiographic views.44 DeepVentricle,71 a DL application for
to retinal fundus photographs trained on 284 335 patients, DL pre- .. cardiac MRI images, received FDA clearance for clinical use.72 Bai et
..
dicted MACE with c-statistic 0.70 in two validation populations .. al.,62 applied DL to automatically segment 93 500 labelled MR images
(12 026 and 999 patients).59 Several reports have used DL to define .. in 4875 subjects from the UK Biobank with similar accuracy to
..
structural heart disease from echocardiography and MR imaging .. experts for segmenting left and right ventricles on short-axis images
(Table 2). Deep learning can diagnose structural disease from limited .. and left and right atrium on long-axis images.
.
Table 2 Deep learning applications in cardiovascular medicine
First author Disease application Images/slides (N) Type of images Results Algorithms
................................................................................................................................................................................................................................................................................................
Arrhythmias
Bumgarner et al.60 AF detection 169 records A rhythm strip from Kardia 93% sensitivity and 84% specificity for AF detection Kardia Band automated
mobile phone algorithm
Pyakillya et al.50 ECG classification 8528 records ECG records 86% accuracy for classification results Deep CNN
Hannun et al.49 Arrhythmias detection (i.e. VT, 64 121 ECG records ECG records The model outperforms the average cardiologist performance on Deep CNN
Mobitz I, Mobitz II) most arrhythmias (measured by sequence level accuracy and set
level accuracy)
Tison et al.8 AF detection 9000 ECGs ECG from smartphone On recumbent ECG, c-statistic 0.97; on ambulatory ECG, c-statistic Deep CNN
0.72
Deep learning for cardiovascular medicine
Xia et al.48 AF detection 123 848 samples ECG signal (STFT and SWT) 98.29% accuracy for STFT and 98.63% accuracy for SWT Deep CNN
Cardiac MRI
Avendi et al.63 Image segmentation (LV shape 45 MRI datasets Cardiac MRI images 90% accuracy (in terms of the Dice metric) Deep CNN and stacked
detection) autoencoders
Bai et al.62 Image segmentation (LV and 93 500 images Cardiac MRI images Reported as Dice metric, mean surface distance, and hausdorff surface Deep CNN (a fully CNN)
RV) distance
Luo et al.64 Image segmentation (LV 1140 subjects Cardiac MRI images High accuracy in LV volumes prediction (measured accuracy by root Deep CNN
volumes) mean squared error)
Oktay et al.65 Image quality (i.e. LV segmenta- 1233 healthy adult Cardiac MRI images Directly compare image quality and segmentation results (i.e. LV cavity Deep CNN (super resolution
tion, motion tracking) subjects volume differences or surface-to-surface distances) approaches)
Oktay et al.47 Image segmentation (cardiac 1200 images Cardiac MRI images Reported as Dice metric, mean surface distance, and hausdorff surface Deep CNN
pathologies) distance
Echocardiography
Dong et al.66 LV segmentation 60 subjects Echocardiography images Reported as Dice metric, mean surface distance, and hausdorff surface Deep GAN (VoxelAtlasGAN)
distance
Gao et al.67 8 echo views classification 432 video images Echocardiography images 92.1% accuracy for classification results Deep CNN
Knackstedt et al.68 Assessment of LV volumes and 432 video images Echocardiography images 92.1% accuracy for classification results Deep CNN
EF
Luong et al.43 Assessment of echo image qual- 6916 images Echocardiography images The average absolute error of the model compared with manual Deep CNN
ity feedback scoring was 0.68 ± 0.58
Madani et al.44 15 echo views recognition 223 787 images Echocardiography images 91.7% accuracy for classification results (F-score 0.904 ± SD 0.058) Deep CNN
Heart failure
Nirschl et al.69 Clinical heart failure 209 patients H&E stained whole-slide images > 93% accuracy for both training and testing datasets vs. 75% accuracy Deep CNN
classification for pathologists
Seah et al.70 Prediction of CHF 103 489 images Chest X-rays Model achieved an AUC of 0.82 (At a cut-off BNP of 100 ng/L) Deep GAN
Islam et al.105 Pulmonary oedema detection 7284 images Chest X-rays The same architecture does not perform well across all abnormalities Deep CNN
Myocardial perfusion imaging
Betancur et al.56 Prediction of obstructive CAD 1638 patients Fast SPECT MPI AUC for disease prediction by deep learning was higher than for total Deep CNN
perfusion deficit
AF, atrial fibrillation; AUC, the area under the receiver operating characteristic curve; CAD, coronary artery disease; CHF, congestive heart failure; CNN, convolutional neural network; ECG, electrocardiogram; EF, ejection fraction; GAN,
generative adversarial network; HCM, hypertrophic cardiomyopathy; LV, left ventricle; MPI, myocardial perfusion imaging; MRI, magnetic resonance imaging; RV, right ventricle; SPECT, single-photon emission computerized tomography;
9
STFT, short-term Fourier transform; SWT, stationary wavelet transform.


Take home ﬁgure Deep learning process flow for cardiovascular medicine. Deep learning has the ability to produce actionable clinical in-
formation from diverse datasets. Such data may span (i) comprehensive, traditional clinical data; (ii) non-traditional ‘real-world’ data such as
near-continuous streams from wearable devices but also questionnaires or online forms. The deep learning process flow commences with
designing the most appropriate model. Deep learning is usually implemented by deep neural networks with convolutional layers defined by spe-
cific parameters (e.g. max pooling, activation function, and learning rate). Several algorithms traditionally applied in machine learning (i.e. SVM,
RF, KNN, RNN, AE, GAN) can be combined to address complex problems. Data is selected and pre-processed (curated), and missing elements
are imputed. Training proceeds until the deep learning machine converges at acceptable accuracy. The deep learning machine is then ready to
be applied to unseen test data.
..
Deep learning for outcomes prediction in .. Deep learning for arrhythmia detection
heart failure .. and phenotyping
..
Several studies have applied ML to predict outcomes in heart fail- .. Several studies have used DL to diagnose AF from the ECG. Tison et
.. 8
ure (HF).41,73,74 Choi et al.75 applied neural networks to detect .. al. trained a DL network on 9750 ambulatory smartwatch ECGs,
new onset HF from electronic health records in 3884 patients .. then applied it to 12-lead ECGs. The network performed well in 51
..
who developed incident HF and 28 903 who did not, linking time- .. recumbent patients before cardioversion (c-statistic 0.97 vs. 0.91 for
stamped events (disease diagnosis, medication and procedure .. current ECG algorithms), but less well in a cohort with ambulatory
..
orders). Networks provided a c-statistic for incident HF of 0.78 .. ECGs (c-statistic 0.72, sensitivity 67.7% and specificity 67.6%). A sep-
(12-month observation) and 0.88 (18-month observation), both .. arate study used AI to diagnose AF from electrical ECG sensors in a
..
significantly higher than the best baseline method. Medved et al.76 .. smartphone case (Kardia), or a watch-strap which communicates via
compared the International Heart Transplantation Survival .. Bluetooth to a smartphone (Kardia Band).60 In 100 patients with 169
..
Algorithm developed using DL training, with the Index for .. simultaneous wearable and traditional ECGs, 57 recordings were un-
Mortality Prediction After Cardiac Transplantation (IMPACT), in .. interpretable. For interpretable ECGs, the device identified AF with a
..
transplant recipients from 1997 to 2011 from the UNOS registry. .. K coefficient of 0.77 (sensitivity 93%, specificity 84%) although phys-
In 27 705 patients, using those before 2009 for training and .. ician interpretation improved results further.60 Thus, while these
..
those from 2009 for validation, DL provided a c-statistic for 1- .. data are promising, further advances in analytic algorithms and sensor
year survival of 0.654 for IHTSA, which reduced error by
.. technology are needed for automatic use. Emerging sensors beyond
..
1 compared with the IMPACT model (c-statistic 0.608). These .. optical sensors (photoplesthysmography) in the iWatch8 include
results, while modest, show promise for DL beyond current
.. changes in facial reflectance to visible light,78 bioimpedance in weigh-
..
clinical indices. Future studies may apply DL to multi-variable .. ing scales79 and others. The accuracy of each sensor needs validation
data (e.g. histopathology, echo, ECG, labs, multi-omics, wearable
.. since, in recent comparisons against gold standards, wearable sensors
..
technology) to study HF outcomes. The recent BIOSTAT-CHF .. had acceptable accuracy for resting heart rate yet not for ambulatory
trial, a large registry for risk prediction for HF in 11 European
.. exercise heart rates nor energy expenditure.80 Current ESC guide-
..
countries, has multi-level data that could be used to reclassify .. lines provide a Class I recommendation for the opportunistic screen-
..
HF patients.77 . ing for silent AF in patients >65 years of age by pulse or ECG rhythm
..
strip, based on evidence on cost effectiveness.81 Therefore, integrat- .. has several limitations. Problems are that the c-index is only a
ing DL into wearable technology for intermittent screening for silent .. measure of discrimination, ignores calibration indices and has no
..
AF may be cost effective by preventing sequelae such as stroke. .. single universally accepted c-statistic cut-off or range of acceptable c-
Finally, ML shows promise in identifying novel arrhythmia pheno- .. statistics.92,93 This and other limitations can be addressed by calibra-

..
types, using unsupervised learning (since labelled data do not exist, by .. tion: dividing underlying continuous variables representing the diag-
definition). For AF, Inohara et al.82 analysed the large clinical ORBIT-AF .. nostic into partitions (e.g. deciles) and testing diagnostic ability in
..
database to identify clusters labelled atherosclerotic-comorbid, tachy- .. each. It is important to use multiple metrics of accuracy since, for ex-
brady device, low comorbidity, and younger behavioural disorder. The .. ample, c-statistics discriminate true outcomes (e.g. high risk from low
..
cluster approach did not improve CHADS2VASc, ORBIT, and ATRIA .. risk patients), but are insensitive to systematic errors and do not iden-
scores for endpoints of stroke or bleeding; however, there was a slight .. tify whether the model is anchored at the right level of absolute risk
..
improvement in combination (c-statistics 0.67–0.72). In hypertrophic .. across the spectrum of observable risks. This should be applied to
cardiomyopathy, Lyon et al.26 identified four clusters of high risk of sud- .. machine learned models, but has yet to be done. In addition, receiver-
..
den cardiac arrest from ECG (primary T wave inversion) and echocar- .. operating characteristic curve is an ordinal technique which assumes
diographic (septal and apical hypertrophy) features. The clinical utility .. that the underlying biological process is monotonic, yet this is often
..
of novel machine learned phenotypes should be validated in independ- .. not true. For instance, blood pressure at extreme high or low levels
ent populations compared with traditional clinical classification.
.. are of disproportionate importance. Machine learning has the advan-
..
.. tage that it is not constrained by a monotonic assumption. Reporting
.. a comparison between DL with traditional statistical results (CNN vs.
..
Challenges for, and limitations of, ..
..
logistic regression), the Brier score, the goodness of fit,94 calibration
deep learning .. plots,95 standardized checklist/strobe diagrams of prediction mod-
.. els96 or the decision-curve analysis97 would be helpful.
..
To date, cardiovascular results of DL are promising but still modest, .. Sixth, there is likely a positive publication bias in medical studies of
and several challenges must be overcome. First, and most important- .. ML and DL. However, since the number of such studies is currently
..
ly, DL is often criticized in the clinical context as a black box which .. limited, a meta-analysis of such studies may not be worthwhile, so
cannot easily be explained. Interpretability may be enabled by capsule .. that funnel plots or other indices of publication bias or heterogeneity
..
based networks, or strategies that systematically censor inputs to de- .. (I2) would be difficult to quantify. As discussed in several of the specif-
fine those that most affect classification. Meta-analyses of several DL .. ic examples, negative results are indeed discussed. These studies
..
algorithms applied to the same data may increase confidence in .. should be used to refine DL methods, which then need to be tested
results. A number of techniques may enable ‘model-agnostic’ metrics .. via external independent multicentre validation.
..
for interpretability of complex models.18 Marblestone et al.83 .. Seventh, DL and standard statistical methods may often be used for
hypothesized analogies between DL and human cognitive functioning, .. the same problems, although in addition to parallels between the meth-
..
proposing that integrating heterogeneous ‘cost functions’ over time .. ods, divergences exist. Both are influenced by aberrations in sample
may simplify learning. Thus, speculatively, insights into human cogni-
.. data, and can suffer from overfitting. This may be more predictable for
..
tion may ultimately provide insights to interpret DL models. .. statistical methods than for ML which relies on empirical validation.
Second, all ML may suffer from overfitting (Figure 5) if data is lim-
.. Cardiovascular problems which can be stated as a clear hypothesis may
..
ited and/or algorithms complex. Indeed, in some clinical studies DL .. be equally addressed by traditional statistics or DL. Conversely, if simple
provided similar results to statistical models (e.g. logistic regres-
.. hypotheses are less readily formulated due to complex interactions, DL
..
sion).84 This may simply mean that different analyses are better suited .. may be advantageous. Thus, both techniques are complementary tools.
to different types of data. Future studies may integrate DL with statis-
.. Finally, a theoretical framework to guide big-data and DL designs is
..
tical classification. .. urgently needed. Rather than ask simply if data quantity is sufficient for
.. study, it is pertinent to ask if data quality and diversity are sufficient to
Third, DL faces recent ethical criticisms if biased or poor-quality ..
data lead to biased predictions or, worse, facilitate manipulation of .. span the parameter space necessary to address the question. Adding
.. data dimensions increases the chance of alpha error, i.e. finding
results. Adversarial examples, cases in which slight modifications to ..
input data cause a major change in DL classification, are a significant .. chance associations in traditional biostatistics, yet it may enrich train-
..
concern for DL with potentially serious medical sequelae.85–87 Some .. ing for DL. Are data reliable (‘garbage in, garbage out’)? In a large big-
methodologies have been proposed to prevent adversarial examples .. data arrhythmia study comparing genotype with phenotype in 2022
..
(i.e. reactive strategies), but they remain ineffective.88 .. patients with long QT and Brugada syndromes,98 variability in genetic
Fourth, DL studies must enable replication by other groups since .. testing compromised its results. A final consideration is how to or-
..
differences in algorithms, initiating conditions, or parameter tuning .. ganize data given that most data structures are generic and there are
may alter results.89 One replication study,90 for example, demon- .. few which are standardized for cardiovascular medicines.
..
strated different results from another91 using the same algorithms. ..
Thus, a standardized approach to perform and validate AI-related ..
..
clinical studies is needed. One initial step would be to require investi- .. Training programmes in deep
..
gators to deposit their data and a link to the code for their DL model. .. learning
Medicine lags behind computer science in this respect. ..
Fifth, DL in cardiovascular medicine have thus far compared c-sta-
.. Educational programmes should incorporate classes on DL, given its
..
tistics (area under the receiver-operating characteristic curve), which . already ubiquitous presence. Such classes should cover the rationale,
..
solutions, technical, and ethical challenges it poses in medicine. At the .. analysis may not only identify hidden information in complex, hetero-
undergraduate and graduate level, such training may focus on its com- .. geneous datasets, but also may bridge the gap between disease
..
plementary role to biostatistics, and on detailed software program- .. pathogenesis, genotypes, phenotypes to enable personalized medi-
ming and hardware aspects. In medical education, implementation of .. cine. However, to transform cardiovascular care, DL will have to ad-

..
a broad AI curriculum is likely to enrich understanding of many condi- .. dress challenges in obtaining extensive labelled data, in improving
tions in cardiovascular medicine with heterogeneous aetiologies .. interpretability and robustness, and in developing standardized
..
and/or phenotypes such as HFpEF, AF, and hypertension.99,100 .. approaches for validation and testing. Deep learning is one of the
The medical curriculum should also discuss ethical and legal chal- .. most exciting areas of innovation in cardiovascular medicine that
..
lenges, and their potential to shape medical practice. A significant bar- .. holds the possibility to provide more efficient care with improvement
rier to implement DL more broadly is the need for at least some .. in outcomes.
..
programming familiarity. This may be less problematic for newer gen- ..
erations of trainees. Deep learning is an excellent vehicle to foster ..
..
interdisciplinary teams of engineers, physicians, businessmen, legal, .. Supplementary material
and ethical teams.100 It may be helpful to borrow engineering ..
.. Supplementary material is available at European Heart Journal online.
approaches such as ‘hackathons’, e.g. the Data Science Bowl, ..
PhysioNet/Computing in Cardiology Challenges or Kaggle
..
.. Funding
Competition. Scientists including biomedical trainees could compete ..
at these or traditional medical conferences to analyse cardiovascular
.. S.M.N. is supported, in part, by grants from the National Institutes of
.. Health (NIH R01 HL 83359; K24 HL103800).
sample datasets (imaging, ECGs) using DL. ..
Funding opportunities for DL outside medicine are increasing, but
.. Conflict of interest: S.M.N. is author of patents owned by Stanford
..
funding from ESC/AHA/ACC/NIH are increasingly needed.101 .. University and the University of California Regents, and is a consultant to
.. Beyond Limits.ai. R.S.R. reports research grants to his institution from
Crowdfunding has been an alternative for DL funding outside medi- ..
cine and, although rare thus far in cardiovascular research, the poten- .. Akcea, Amgen, AstraZeneca, Medicines Company and Regneron; con-
.. sulting from Akcea, Amgen, C5, CVS Caremark; non-speaker bureau lec-
tial of crowdfunding for cardiovascular DL research is intriguing.102 .. tures from Amgen and Kowa; and stock ownership in MediMergent.
Guidelines must be developed to standardize broad applications of ..
.. J.K.M. discloses the following relationships - medical advisory board: GE
AI in medicine. This will require complex discussions between mul- .. Healthcare, Arineta; equity interest: Cleerly. All other authors declared
tiple stakeholders including regulatory agencies in Europe and Asia, ..
.. no conflict of interest.
the U.S. Food and Drug Administration (FDA) and others, patient- ..
advocacy and privacy groups, professional societies in medicine and ..
..
computer science, other healthcare organizations and technology .. References
companies. .. 1. Ponikowski P, Voors AA, Anker SD, Bueno H, Cleland JG, Coats AJ, Falk V,
.. Gonzalez-Juanatey JR, Harjola VP, Jankowska EA, Jessup M, Linde C,
.. Nihoyannopoulos P, Parissis JT, Pieske B, Riley JP, Rosano GM, Ruilope LM,
..
.. Ruschitzka F, Rutten FH, van der Meer P. 2016 ESC Guidelines for the diagnosis
Future direction of deep learning ..
..
and treatment of acute and chronic heart failure: the task force for the diagno-
sis and treatment of acute and chronic heart failure of the European Society of
in cardiovascular medicine .. Cardiology (ESC). Developed with the special contribution of the Heart Failure
..
.. 2. Association (HFA) of the ESC. Eur J Heart Fail 2016;18:891–975.
Regitz-Zagrosek V, Roos-Hesselink JW, Bauersachs J, Blomström-Lundqvist C,
Deep learning promises to better integrate medical data sources, ad- .. Cı́fková R, De Bonis M, Iung B, Johnson MR, Kintscher U, Kranke P, Lang IM,
dress the heterogeneity in patient disease types, bridge the gap be- ..
.. Morais J, Pieper PG, Presbitero P, Price S, Rosano GMC, Seeland U, Simoncini
tween omics research and bedside phenotypes103,104 and ultimately .. T, Swan L, Warnes CA;ESC Scientific Document Group. 2018 ESC guidelines
enable personalized medicine. This may require advances in the sci- .. for the management of cardiovascular diseases during pregnancy. Eur Heart J
.. 2018;39:3165–3241.
ence to overcome current limitations including a limited theoretical .. 3. Baumgartner H, Bonhoeffer P, De Groot NM, de Haan F, Deanfield JE, Galie N,
foundation for design and testing, limited interpretability, and strat- .. Gatzoulis MA, Gohlke-Baerwolf C, Kaemmerer H, Kilner P, Meijboom F,
.. Mulder BJ, Oechslin E, Oliver JM, Serraf A, Szatmari A, Thaulow E, Vouhe PR,
egies to resolve overfitting.61 Cardiovascular medicine, in particular, ..
is well suited to benefit from smart analysis of continuous and mas-
.. Walma E. ESC guidelines for the management of grown-up congenital heart dis-
.. ease (new version 2010). Eur Heart J 2010;31:2915–2957.
sive data streams in this new era of wearable sensors, to integrate .. 4. Silver D, Huang A, Maddison CJ, Guez A, Sifre L, van den Driessche G,
traditional health data with lifestyle indices (Take home figure). If data
.. Schrittwieser J, Antonoglou I, Panneershelvam V, Lanctot M, Dieleman S,
.. Grewe D, Nham J, Kalchbrenner N, Sutskever I, Lillicrap T, Leach M,
privacy and security concerns were satisfied, this integration would .. Kavukcuoglu K, Graepel T, Hassabis D. Mastering the game of go with deep
..
form the basis of a medical internet of things between medical devi- .. 5. neural networks and tree search. Nature 2016;529:484–489.
ces and analytic systems. Seamless integration of diverse sources of .. Kelleher K. Deepmind’s ai is now beating humans at quake because winning in
data could enable continuous disease monitoring, risk stratification

.. go wasn’t terrifying enough. https://www.businessinsider.com/deepmind-ai-beat-
.. humans-quake-2018-7 (5 February 2019).
and early warnings of potential decompensation. .. 6. Mnih V, Kavukcuoglu K, Silver D, Rusu AA, Veness J, Bellemare MG, Graves A,
.. Riedmiller M, Fidjeland AK, Ostrovski G, Petersen S, Beattie C, Sadik A,
.. Antonoglou I, King H, Kumaran D, Wierstra D, Legg S, Hassabis D. Human-
..
.. 7. level control through deep reinforcement learning. Nature 2015;518:529–533.
Conclusion .. Jaderberg M, Czarnecki WM, Dunning I, Marris L, Lever G, Garcia Castaneda A,
.. Beattie C, Rabinowitz NC, Morcos AS, Ruderman A, Sonnerat N, Green T,
Deep learning is a rapidly developing field in the computer sciences

.. Deason L, Leibo JZ, Silver D, Hassabis D, Kavukcuoglu K, Graepel T. Human-
.. level performance in first-person multiplayer games with population-based
with great promise for cardiovascular medicine. Using DL for big data . deep reinforcement learning. ArXiv e-prints. 2018.
8. Tison GH, Sanchez JM, Ballinger B, Singh A, Olgin JE, Pletcher MJ, Vittinghoff E,
.. environments. ICLR 2018: International Conference on Learning
..
Lee ES, Fan SM, Gladstone RA, Mikell C, Sohoni N, Hsieh J, Marcus GM. Passive .. Representations, Vancouver, BC, Canada.
detection of atrial fibrillation using a commercially available smartwatch. JAMA .. 35. Peng XB, Andrychowicz M, Zaremba W, Abbeel P. Sim-to-real transfer of ro-
Cardiol 2018;3:409–416. .. botic control with dynamics randomization. 2018. IEEE International
9. Bello GA, Dawes TJ, Duan J, Biffi C, de Marvao A, Howard LS, Gibbs JSR, .. Conference on Robotics and Automation (ICRA), Brisbane, Australia.

Wilkins MR, Cook SA, Rueckert D. Deep learning cardiac motion analysis for .. 36. Frans K, Ho J, Chen X, Abbeel P, Schulman J. Meta learning shared hierarchies.
human survival prediction. arXiv preprint arXiv:1810.03382. 2018. .. ICLR 2018: International Conference on Learning Representations, Vancouver,
10. Al’Aref SJ, Anchouche K, Singh G, Slomka PJ, Kolli KK, Kumar A, Pandey M,
.. BC, Canada. eprint arXiv:1710.09767. 2017:arXiv:1710.09767.
..
Maliakal G, van Rosendael AR, Beecy AN, Berman DS, Leipsic J, Nieman K, .. 37. Alemi AA, Chollet F, Een N, Irving G, Szegedy C, Urban J. Deepmath—deep se-
Andreini D, Pontone G, Schoepf UJ, Shaw LJ, Chang H-J, Narula J, Bax JJ, Guan .. quence models for premise selection. 2016 Advances in Neural Information
Y, Min JK. Clinical applications of machine learning in cardiovascular disease and .. Processing Systems (NIPS 2016), Barcelona, Spain. ArXiv e-prints.
its relevance to cardiac imaging. Eur Heart J 2018; doi:10.1093/eurheartj/ehy404. .. 38. Bisgin H, Bera T, Ding H, Semey HG, Wu L, Liu Z, Barnes AE, Langley DA,
11. Heden B, Ohlsson M, Rittner R, Pahlm O, Haisty WK Jr, Peterson C, .. Pava-Ripoll M, Vyas HJ, Tong W, Xu J. Comparing SVM and ANN based ma-
Edenbrandt L. Agreement between artificial neural networks and experienced .. chine learning methods for species identification of food contaminating beetles.
electrocardiographer on electrocardiographic diagnosis of healed myocardial in-
.. Sci Rep 2018;8:6532.
..
farction. J Am Coll Cardiol 1996;28:1012–1016. .. 39. Jodas DS, Marranghello N, Pereira AS, Guido RC. Comparing support vector
12. Vasquez C, Hernandez A, Mora F, Carrault G, Passariello G. Atrial activity en- .. machines and artificial neural networks in the recognition of steering angle for
hancement by wiener filtering using an artificial neural network. IEEE Trans .. driving of mobile robots through paths in plantations. Proc Comput Sci 2013;18:
Biomed Eng 2001;48:940–944. .. 240–249.
13. Vos JE, Scheepstra KA. Computer-simulated neural networks: an appropriate .. 40. Abdullah A, Veltkamp RC, Wiering MA, An ensemble of deep support vector
model for motor development? Early Hum Dev 1993;34:101–112. .. machines for image categorization. 2009 International Conference of Soft
14. Webber WR, Litt B, Wilson K, Lesser RP. Practical detection of epileptiform
.. Computing and Pattern Recognition (SoCPaR 2009), Malacca, Malaysia.
..
discharges (EDs) in the EEG using an artificial neural network: a comparison of .. 41. Li Y, Zhang T. Deep neural mapping support vector machines. Neural Netw
raw and parameterized EEG data. Electroencephalogr Clin Neurophysiol 1994;91: .. 2017;93:185–194.
194–204. .. 42. Rubin J, Abreu R, Ganguli A, Nelaturi S, Matei I, Sricharan K. Recognizing abnor-
15. Narayan SM. Restricted connectivities in neural networks. Masters of Science .. mal heart sounds using deep learning. The 25th International Joint Conference
Thesis. University of Birmingham, UK, Department of Computer Science 1990. .. on Artificial Intelligence IJCAI16, New York, NY. arXiv preprint
p1–153. .. arXiv:1707.04642. 2017.
16. Krittanawong C, Zhang H, Wang Z, Aydar M, Kitai T. Artificial intelligence in
.. 43. Luong C, Abdi A, Jue J, Gin K, Fleming S, Abolmaesumi P, Tsang T. Automatic
..
precision cardiovascular medicine. J Am Coll Cardiol 2017;69:2657–2664. .. quality assessment of echo apical 4-chamber images using computer deep learn-
17. Kaelbling LP, Littman ML, Moore AW. Reinforcement learning: a survey. J Artif .. ing. 2016. The American Heart Association’s Scientific Sessions 2017 in
Intell Res 1996;4:237–285. .. Anaheim, CA.
18. Ribeiro MT, Singh S, Guestrin C. Model-agnostic interpretability of machine .. 44. Madani A, Arnaout R, Mofrad M, Arnaout R. Fast and accurate view classifica-
learning. The Proceedings of the 2016 ICML Workshop on Human .. tion of echocardiograms using deep learning. NPJ Digit Med 2018;1:6.
Interpretability in Machine Learning (WHI 2016), New York, NY. .. 45. Bernard O, Bosch JG, Heyde B, Alessandrini M, Barbosa D, Camarasu-Pop S,
19. Johnson KW, Torres Soto J, Glicksberg BS, Shameer K, Miotto R, Ali M, Ashley .. Cervenansky F, Valette S, Mirea O, Bernier M, Jodoin P-M, Domingos JS,
E, Dudley JT. Artificial intelligence in cardiology. J Am Coll Cardiol 2018;71:
.. Stebbing RV, Keraudren K, Oktay O, Caballero J, Shi W, Rueckert D, Milletari
..
2668–2679. .. F, Ahmadi S-A, Smistad E, Lindseth F, van Stralen M, Wang C, Smedby O,
20. Krizhevsky A, Sutskever I, Hinton GE. Imagenet classification with deep convo- .. Donal E, Monaghan M, Papachristidis A, Geleijnse ML, Galli E, D’hooge J.
lutional neural networks. Advances in Neural Information Processing Systems .. Standardized evaluation system for left ventricular segmentation algorithms in
(NIPS) 2012, Harrah’s Lake Tahoe, Stateline, NV. .. 3D echocardiography. IEEE Trans Med Imaging 2016;35:967–977.
21. Simonyan K, Zisserman A. Very deep convolutional networks for large-scale .. 46. Keraudren K, Oktay O, Shi W, Hajnal JV, Rueckert D. Endocardial 3D ultra-
image recognition. arXiv preprint arXiv:1409.1556. 2014. .. sound segmentation using autocontext random forests. In: Proceedings MICCAI
22. Szegedy C, Ioffe S, Vanhoucke V, Alemi AA. Inception-v4, inception-resnet and
.. Challenge on Echocardiographic Three-Dimensional Ultrasound Segmentation
..
the impact of residual connections on learning. 2017. International Conference .. (CETUS), Boston, MA. 2014. p.41–48.
on Learning Representations 2016, San Juan, Puerto Rico. .. 47. Oktay O, Ferrante E, Kamnitsas K, Heinrich M, Bai W, Caballero J, Cook SA, de
23. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. .. Marvao A, Dawes T, O’Regan DP, Kainz B, Glocker B, Rueckert D.
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), .. Anatomically constrained neural networks (ACNNS): application to cardiac
Las Vegas, NV. .. image enhancement and segmentation. IEEE Trans Med Imaging 2018;37:
24. Pan SJ, Yang Q. A survey on transfer learning. IEEE Trans Knowl Data Eng 2010; .. 384–395.
22:1345–1359.
.. 48. Xia Y, Wulan N, Wang K, Zhang H. Detecting atrial fibrillation by deep convo-
..
25. Afshar P, Mohammadi A, Plataniotis KN. Brain tumor type classification via cap- .. lutional neural networks. Comput Biol Med 2018;93:84–92.
sule networks. ArXiv e-prints 2018. 25th IEEE International Conference on .. 49. Hannun AY, Rajpurkar P, Haghpanahi M, Tison GH, Bourn C, Turakhia MP, Ng
Image Processing (ICIP), Athens, Greece. .. AY. Cardiologist-level arrhythmia detection and classification in ambulatory
26. Lyon A, Ariga R, Minchole A, Mahmod M, Ormondroyd E, Laguna P, de Freitas .. electrocardiograms using a deep neural network. Nature Medicine 2019;25:
N, Neubauer S, Watkins H, Rodriguez B. Distinct ECG phenotypes identified in .. 65–69.
hypertrophic cardiomyopathy using machine learning associate with arrhythmic .. 50. Pyakillya B, Kazachenko N, Mikhailovsky N. Deep learning for ecg classification.
risk markers. Front Physiol 2018;9:213.
.. J Phys Conf Ser 2017;913:012004.
..
27. Rubin DB. Inference and missing data. Biometrika 1976;63:581–592. .. 51. Shashikumar SP, Shah AJ, Li Q, Clifford GD, Nemati S, A deep learning ap-
28. Siddique J, Harel O, Crespi CM. Addressing missing data mechanism uncertainty .. proach to monitoring and detecting atrial fibrillation using wearable technology.
using multiple-model multiple imputation: application to a longitudinal clinical .. In: 2017 IEEE EMBS International Conference on Biomedical & Health Informatics
trial. Ann Appl Stat 2012;6:1814–1837. .. (BHI). Orlando, FL. 2017. p.141–144.
29. Beaulieu-Jones BK, Lavage DR, Snyder JW, Moore JH, Pendergrass SA, Bauer .. 52. Ehteshami Bejnordi B, Veta M, Johannes van Diest P, van Ginneken B,
CR. Characterizing and managing missing structured data in electronic health .. Karssemeijer N, Litjens G, van der Laak JAWM, Hermsen M, Manson QF,
records: data analysis. JMIR Med Inform 2018;6:e11. .. Balkenhol M, Geessink O, Stathonikos N, van Dijk MC, Bult P, Beca F, Beck
30. Falissard L, Fagherazzi G, Howard N, Falissard B. Deep clustering of longitudinal
.. AH, Wang D, Khosla A, Gargeya R, Irshad H, Zhong A, Dou Q, Li Q, Chen H,
..
data. arXiv preprint arXiv:1802.03212. 2018. .. Lin H-J, Heng P-A, Haß C, Bruni E, Wong Q, Halici U, Öner MÜ, Cetin-Atalay
31. Karch J, A machine learning perspective on repeated measures. http://edoc.hu- .. R, Berseth M, Khvatkov V, Vylegzhanin A, Kraus O, Shaban M, Rajpoot N,
berlin.de/18452/18293 (5 February 2019). .. Awan R, Sirinukunwattana K, Qaiser T, Tsang Y-W, Tellez D, Annuscheit J,
32. Hinton GF. A parallel computation that assigns canonical object-based frames .. Hufnagl P, Valkonen M, Kartasalo K, Latonen L, Ruusuvuori P, Liimatainen K,
of reference. In: Proceedings of the 7th International Joint Conference on .. Albarqouni S, Mungal B, George A, Demirci S, Navab N, Watanabe S, Seno S,
Artificial Intelligence, Vancouver, BC, Canada. Vol. 2. 1981. p683–685. .. Takenaka Y, Matsuda H, Ahmady Phoulady H, Kovalev V, Kalinovsky A,
33. Sabour S, Frosst N, E Hinton G. Dynamic routing between capsules. Advances
.. Liauchuk V, Bueno G, Fernandez-Carrobles MM, Serrano I, Deniz O,
..
in Neural Information Processing Systems (NIPS 2017), Long Beach, CA. .. Racoceanu D, Venâncio R. Diagnostic assessment of deep learning algorithms
34. Al-Shedivat M, Bansal T, Burda Y, Sutskever I, Mordatch I, Abbeel P. .. for detection of lymph node metastases in women with breast cancer. JAMA
Continuous adaptation via meta-learning in nonstationary and competitive . 2017;318:2199–2210.
53. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S.
.. 74. Miller K, Hettinger C, Humpherys J, Jarvis T, Kartchner D. Forward thinking:
..
Dermatologist-level classification of skin cancer with deep neural networks. .. building deep random forests. arXiv preprint arXiv:1705.07366. 2017.
Nature 2017;542:115. .. 75. Choi E, Schuetz A, Stewart WF, Sun J. Using recurrent neural network models
54. Mundhra D, Cheluvaraju B, Rampure J, Rai Dastidar T, Analyzing microscopic .. for early detection of heart failure onset. J Am Med Inform Assoc 2017;24:
images of peripheral blood smear using deep learning. In: Jorge Cardoso (ed.), .. 361–370.

Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical .. 76. Medved D, Ohlsson M, Höglund P, Andersson B, Nugues P, Nilsson J.
Decision Support. Springer International Publishing, 2017. p178–185. .. Improving prediction of heart transplantation outcome using deep learning
55. Zreik M, Lessmann N, van Hamersvelt RW, Wolterink JM, Voskuil M, Viergever
.. techniques. Sci Rep 2018;8:3613.
..
MA, Leiner T, Isgum I. Deep learning analysis of the myocardium in coronary .. 77. Voors AA, Anker SD, Cleland JG, Dickstein K, Filippatos G, van der Harst P,
CT angiography for identification of patients with functionally significant coron- .. Hillege HL, Lang CC, Ter Maaten JM, Ng L, Ponikowski P, Samani NJ, van
ary artery stenosis. Med Image Anal 2018;44:72–85. .. Veldhuisen DJ, Zannad F, Zwinderman AH, Metra M. A systems biology study
56. Betancur J, Commandeur F, Motlagh M, Sharir T, Einstein AJ, Bokhari S, Fish .. to tailored treatment in chronic heart failure: rationale, design, and baseline
MB, Ruddy TD, Kaufmann P, Sinusas AJ, Miller EJ, Bateman TM, Dorbala S, Di .. characteristics of BIOSTAT-CHF. Eur J Heart Fail 2016;18:716–726.
Carli M, Germano G, Otaki Y, Tamarappoo BK, Dey D, Berman DS, Slomka PJ. .. 78. Couderc JP, Kyal S, Mestha LK, Xu B, Peterson DR, Xia X, Hall B. Detection of
Deep learning for prediction of obstructive disease from fast myocardial perfu-
.. atrial fibrillation using contactless facial video monitoring. Heart Rhythm 2015;
..
sion SPECT: a multicenter study. JACC Cardiovasc Imaging 2018;11:1654–1663. .. 12:195–201.
57. Motwani M, Dey D, Berman DS, Germano G, Achenbach S, Al-Mallah MH, .. 79. Dı́az DH, Ó C, Pallas-Areny R. Heart rate detection from single-foot plantar
Andreini D, Budoff MJ, Cademartiri F, Callister TQ, Chang HJ, Chinnaiyan K, .. bioimpedance measurements in a weighing scale. Conf Proc IEEE Eng Med Biol
Chow BJ, Cury RC, Delago A, Gomez M, Gransar H, Hadamitzky M, Hausleiter .. Soc 2010;2010:6489–6492.
J, Hindoyan N, Feuchtner G, Kaufmann PA, Kim YJ, Leipsic J, Lin FY, Maffei E, .. 80. Shcherbina A, Mattsson CM, Waggott D, Salisbury H, Christle JW, Hastie T,
Marques H, Pontone G, Raff G, Rubinshtein R, Shaw LJ, Stehli J, Villines TC, .. Wheeler MT, Ashley EA. Accuracy in wrist-worn, sensor-based measurements
Dunning A, Min JK, Slomka PJ. Machine learning for prediction of all-cause mor-
.. of heart rate and energy expenditure in a diverse cohort. J Pers Med 2017;7:2.
..
tality in patients with suspected coronary artery disease: a 5-year multicentre .. 81. Kirchhof P, Benussi S, Kotecha D, Ahlsson A, Atar D, Casadei B, Castella M,
prospective registry analysis. Eur Heart J 2017;38:500–507. .. Diener H-C, Heidbuchel H, Hendriks J, Hindricks G, Manolis AS, Oldgren J,
58. Li X, Liu H, Yang J, Xie G, Xu M, Yang Y. Using machine learning models to pre- .. Popescu BA, Schotten U, Van Putte B, Vardas P, Agewall S, Camm J, Baron
dict in-hospital mortality for ST-elevation myocardial infarction patients. Stud .. Esquivias G, Budts W, Carerj S, Casselman F, Coca A, De Caterina R,
Health Technol Inform 2017;245:476–480. .. Deftereos S, Dobrev D, Ferro JM, Filippatos G, Fitzsimons D, Gorenek B,
59. Poplin R, Varadarajan AV, Blumer K, Liu Y, McConnell MV, Corrado GS, Peng .. Guenoun M, Hohnloser SH, Kolh P, Lip GYH, Manolis A, McMurray J,
L, Webster DR. Prediction of cardiovascular risk factors from retinal fundus
.. Ponikowski P, Rosenhek R, Ruschitzka F, Savelieva I, Sharma S, Suwalski P,
..
photographs via deep learning. Nature Biomedical Engineering 2018;2:158–164. .. Tamargo JL, Taylor CJ, Van Gelder IC, Voors AA, Windecker S, Zamorano JL,
60. Bumgarner JM, Lambert CT, Hussein AA, Cantillon DJ, Baranowski B, Wolski K, .. Zeppenfeld K. 2016 ESC guidelines for the management of atrial fibrillation
Lindsay BD, Wazni OM, Tarakji KG. Smartwatch algorithm for automated de- .. developed in collaboration with EACTS. Eur Heart J 2016;37:2893–2962.
tection of atrial fibrillation. J Am Coll Cardiol 2018;71:2381–2388. .. 82. Inohara T, Shrader P, Pieper K, Blanco RG, Thomas L, Singer DE, Freeman JV,
61. Cherkassky V. The nature of statistical learning theory. IEEE Trans Neural Netw .. Allen LA, Fonarow GC, Gersh B, Ezekowitz MD, Kowey PR, Reiffel JA,
1997;8:1564. .. Naccarelli GV, Chan PS, Steinberg BA, Peterson ED, Piccini JP. Association of
62. Bai W, Sinclair M, Tarroni G, Oktay O, Rajchl M, Vaillant G, Lee AM, Aung N, .. atrial fibrillation clinical phenotypes with treatment patterns and outcomes: a
Lukaschuk E, Sanghvi MM, Zemrak F. Automated cardiovascular magnetic res-
.. multicenter registry study. JAMA Cardiol 2018;3:54–63.
..
onance image analysis with fully convolutional networks. J Cardiovasc Magn .. 83. Marblestone AH, Wayne G, Kording KP. Toward an integration of deep learn-
Reson 2018;20:6. .. ing and neuroscience. Front Comput Neurosci 2016;10:94.
63. Avendi M, Kheradvar A, Jafarkhani H. A combined deep-learning and .. 84. Jansche M, Maximum expected f-measure training of logistic regression models.
deformable-model approach to fully automatic segmentation of the left ven- .. In: Proceedings of the Conference on Human Language Technology and Empirical
tricle in cardiac MRI. Med Image Anal 2016;30:108–119. .. Methods in Natural Language Processing. Vancouver, British Columbia, Canada,
64. Luo G, Dong S, Wang K, Zuo W, Cao S, Zhang H. Multi-views fusion CNN for .. 2005. p.692–699.
left ventricular volumes estimation on cardiac MR images. IEEE Trans Biomed
.. 85. Yuan X, He P, Zhu Q, Li X. Adversarial examples: attacks and defenses for
..
Eng 2018;65:1924–1934. .. deep learning. IEEE Trans Neural Netw Learn Syst 2019;doi: 10.1109/
65. Oktay O, Bai W, Lee M, Guerrero R, Kamnitsas K, Caballero J, de Marvao A, .. TNNLS.2018.2886017.
Cook S, O’Regan D, Rueckert D. Multi-input cardiac image super-resolution .. 86. Elsayed GF, Shankar S, Cheung B, Papernot N, Kurakin A, Goodfellow I, Sohl-
using convolutional neural networks. MICCAI 2016, the 19th International .. Dickstein J. Adversarial examples that fool both computer vision and time-
Conference on Medical Image Computing and Computer Assisted Intervention, .. limited humans. 32nd Conference on Neural Information Processing Systems
Athens, Greece. .. (NeurIPS 2018), Montréal, Canada. ArXiv e-prints 2018.
66. Dong S, Luo G, Wang K, Cao S, Mercado A, Shmuilovich O, Zhang H, Li S.
.. 87. Papernot N, McDaniel P, Goodfellow I, Jha S, Berkay Celik Z, Swami A.
..
Voxelatlasgan: 3D left ventricle segmentation on echocardiography with atlas .. Practical black-box attacks against machine learning. ASIA CCS ’17 Proceedings
guided generation and voxel-to-voxel discrimination. MICCAI 2018, the 21st .. of the 2017 ACM on Asia Conference on Computer and Communications
International Conference on Medical Image Computing and Computer Assisted .. Security, Abu Dhabi, United Arab Emirates. ArXiv e-prints 2016.
Intervention, Granada, Spain. arXiv preprint arXiv:1806.03619. 2018. .. 88. Meng D, Chen H. Magnet: a two-pronged defense against adversarial examples.
67. Gao X, Li W, Loomes M, Wang L. A fused deep learning architecture for view- .. CCS ’17 Proceedings of the 2017 ACM SIGSAC Conference on Computer and
point classification of echocardiography. Information Fusion 2017;36:103–113. .. Communications Security, Dallas, Texas, USA. ArXiv e-prints 2017.
68. Knackstedt C, Bekkers SC, Schummers G, Schreckenberg M, Muraru D, Badano
.. 89. Olson RS, La Cava W, Mustahsan Z, Varik A, Moore JH. Data-driven advice for
..
LP, Franke A, Bavishi C, Omar AM, Sengupta PP. Fully automated versus stand- .. applying machine learning to bioinformatics problems. Pac Symp Biocomput
ard tracking of left ventricular ejection fraction and longitudinal strain: the .. 2018;23:192–203.
FAST-EFs multicenter study. J Am Coll Cardiol 2015;66:1456–1466. .. 90. Voets M. Deep learning: from data extraction to large-scale analysis. Master’s
69. Nirschl JJ, Janowczyk A, Peyster EG, Frank R, Margulies KB, Feldman MD, .. thesis. UiT Norges arktiske universitet, 2018.
Madabhushi A. A deep-learning classifier identifies patients with clinical heart .. 91. Gulshan V, Peng L, Coram M, Stumpe MC, Wu D, Narayanaswamy A,
failure using whole-slide images of H&E tissue. PLoS One 2018;13:e0192726. .. Venugopalan S, Widner K, Madams T, Cuadros J, Kim R, Raman R, Nelson PC,
70. Seah JCY, Tang JSN, Kitchen A, Gaillard F, Dixon AF. Chest radiographs in con- .. Mega JL, Webster DR. Development and validation of a deep learning algorithm
gestive heart failure: visualizing neural network learning. Radiology 2018;290:
.. for detection of diabetic retinopathy in retinal fundus photographs. JAMA 2016;
..
514–522. .. 316:2402–2410.
71. Lieman-Sifry J, Le M, Lau F, Sall S, Golden D, Fastventricle: cardiac segmentation .. 92. Steyerberg EW, Harrell FE Jr, Borsboom GJ, Eijkemans MJ, Vergouwe Y,
with ENET. In: International Conference on Functional Imaging and Modeling of the .. Habbema JD. Internal validation of predictive models: efficiency of some proce-
Heart. Toronto, ON, Canada, 2017. p127–138 .. dures for logistic regression analysis. J Clin Epidemiol 2001;54:774–781.
72. Pyrkov TV, Slipensky K, Barg M, Kondrashin A, Zhurov B, Zenin A, Pyatnitskiy .. 93. Hand DJ. Measuring classifier performance: a coherent alternative to the area
M, Menshikov L, Markov S, Fedichev PO. Extracting biological age from biomed- .. under the roc curve. Machine Learning 2009;77:103–123.
ical data via deep learning: too much of a good thing? Sci Rep 2018;8:5210.
.. 94. Gong G, Quante AS, Terry MB, Whittemore AS. Assessing the goodness of fit
..
73. Shah SJ, Katz DH, Selvaraj S, Burke MA, Yancy CW, Gheorghiade M, Bonow .. of personal risk models. Stat Med 2014;33:3179–3190.
RO, Huang CC, Deo RC. Phenomapping for novel classification of heart failure .. 95. Gerds TA, Andersen PK, Kattan MW. Calibration plots for risk prediction mod-
with preserved ejection fraction. Circulation 2015;131:269–279. . els in the presence of competing risks. Stat Med 2014;33:3191–3203.
96. Moons KG, Altman DG, Reitsma JB, Ioannidis JP, Macaskill P, Steyerberg EW, .. 99. Krittanawong C, Bomback AS, Baber U, Bangalore S, Messerli FH, Wilson Tang
Vickers AJ, Ransohoff DF, Collins GS. Transparent reporting of a multivariable
.. WH. Future direction for using artificial intelligence to predict and manage
..
prediction model for individual prognosis or diagnosis (tripod): explanation and .. hypertension. Curr Hypertens Rep 2018;20:75.
elaboration. Ann Intern Med 2015;162:W1–73. .. 100. Krittanawong C. Future physicians in the era of precision cardiovascular medi-
97. Steyerberg EW, Vergouwe Y. Towards better clinical prediction models: seven .. cine. Circulation 2017;136:1572–1574.

steps for development and an ABCD for validation. Eur Heart J 2014;35: .. 101. Lardinois F. Andrew ng officially launches his $175m ai fund. https://techcrunch.
1925–1931. .. com/2018/01/30/andrew-ng-officially-launches-his-175m-ai-fund (5 February 2019).
98. Van Driest SL, Wells QS, Stallings S, Bush WS, Gordon A, Nickerson DA, .. 102. Krittanawong C, Zhang HJ, Aydar M, Wang Z, Sun T. Crowdfunding for cardio-
Kim JH, Crosslin DR, Jarvik GP, Carrell DS, Ralston JD, Larson EB, Bielinski
.. vascular research. Int J Cardiol 2018;250:268–269.
..
SJ, Olson JE, Ye Z, Kullo IJ, Abul-Husn NS, Scott SA, Bottinger E, .. 103. Krittanawong C, Johnson KW, Hershman SG, Tang WHW. Big data, artificial in-
Almoguera B, Connolly J, Chiavacci R, Hakonarson H, Rasmussen-Torvik .. telligence, and cardiovascular precision medicine. Expert Rev Precis Med Drug
LJ, Pan V, Persell SD, Smith M, Chisholm RL, Kitchner TE, He MM, Brilliant .. Dev 2018;3:305–317.
MH, Wallace JR, Doheny KF, Shoemaker MB, Li R, Manolio TA, Callis TE, .. 104. Krittanawong C, Kitai T. Identifying genotypes and phenotypes of cardiovascular
Macaya D, Williams MS, Carey D, Kapplinger JD, Ackerman MJ, Ritchie .. diseases using big data analytics. JAMA Cardiol 2017;2:1169–1170.
MD, Denny JC, Roden DM. Association of arrhythmia-related genetic var- .. 105. Islam MT, Aowal MA, Minhaz AT, Ashraf K. Abnormality detection and localiza-
iants with phenotypes documented in electronic medical records. JAMA
.. tion in chest x-rays using deep convolutional neural networks. arXiv preprint
2016;315:47–57.
.. arXiv:1705.09850. 2017.

Deep Learning For Cardiovascular Medicine: A Practical Primer

Uploaded by

Copyright:

Available Formats

Deep Learning For Cardiovascular Medicine: A Practical Primer

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Deep Learning For Cardiovascular Medicine: A Practical Primer

Uploaded by

Copyright:

Available Formats

European Heart Journal (2019) 00, 1–15 CLINICAL REVIEW

doi:10.1093/eurheartj/ehz056 Frontiers in cardiovascular medicine

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

.. heterogeneous phenotypes.1–3 Deep learning (DL) is a branch of

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

There are several parallels, as well as divergences, between trad-

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

rules for computers to follow but, due to the complexity of human

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

several non-neural network-based methods for DL (Supplementary

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

STFT, short-term Fourier transform; SWT, stationary wavelet transform.

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

data could enable continuous disease monitoring, risk stratification

Deep learning is a rapidly developing field in the computer sciences

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

Downloaded from https://academic.oup.com/eurheartj/advance-article-abstract/doi/10.1093/eurheartj/ehz056/5366208 by Macquarie University user on 28 February 2019

You might also like