Artificial Intelligence, Machine Learning, and Big Data Applications in Cardiac Electrophysiology and Arrhythmia Management

© 2023 HMP Global. All Rights Reserved.

Any views and opinions expressed are those of the author(s) and/or participants and do not necessarily reflect the views, policy, or position of EP Lab Digest or HMP Global, their employees, and affiliates. 

EP LAB DIGEST. 2023;23(9):1,7-9.

The proliferation of artificial intelligence (AI) and machine learning (ML) applications in recent years can largely be attributed to the tripartite conglomeration of: (1) advances in ML algorithms and the advent of deep learning (DL); (2) increased computing power including scalable cloud computing; and (3) availability of large datasets. AI describes the ability of a computer to simulate human intelligence or behavior, and ML is a form of AI that enables a computer to independently learn and improve from experience, which in practical applications, comprises data-driven computational modeling. In the field of cardiac electrophysiology (EP), a wide range of enriched and large datasets (Table) can be harnessed by ML to predict a prespecified outcome.

An example of an outcome of interest is risk prediction of sudden cardiac death.¹ In this brief review, we will share key fundamental concepts in ML, focus on important areas of advancement in cardiac EP made possible through ML, and provide insights into future directions and challenges.

Key Concepts in ML and Neural Networks

Whereas traditional statistical approaches rely on modeling to identify relationships between variables, ML allows systems to independently optimize prediction accuracy for an outcome of interest through an iterative process. The algorithm will seek improvement in predictive accuracy until the prediction error (or “loss”) is minimized. In traditional statistical approaches, the relationship between predictors and outcome is assumed, while in ML approaches, it is learned. The theoretical advantage of a ML approach is that it is not hindered by human limitations in interpretation of data or preformed biases in our understanding of relationships between different data. Moreover, ML can tackle large, multidimensional datasets with complex behaviors which may be challenging to analyze using conventional statistical approaches.

ML can be applied to structured or unstructured data. Structured data are clearly defined and organized, such as standard clinical electrocardiogram (ECG) or echocardiogram variables. Alternatively, unstructured data could include raw echocardiogram images or free text files. ML can be supervised or unsupervised. Supervised ML is characterized by the use of labeled datasets. Supervised ML can be further subcategorized into those using classification algorithms such as decision trees with categorical end points (eg, the presence or absence of atrial fibrillation [AF]),² or linear regression algorithms which fit data to a continuous end point (eg, age).^3,4 Unsupervised ML uses algorithms to categorize unlabeled data into discreet subsets (or “clusters”) and can work independently to validate existing classifications⁵ or discover novel ways to categorize data.⁶ Supervised learning models are usually more accurate than unsupervised models, although they require upfront human intervention to appropriately label data. However, unsupervised learning can be useful to identify new disease subtypes or novel phenotypic disease expressions.

The recent advances in ML have come about in large part due to the advent of deep learning, which is ML based on artificial or deep neural networks (DNN, Figure 1).⁷ These networks consist of 3 or more interconnected layers of nodes, each node processing a specific input and each layer processing a specific dimension of the data and feeding forward to the next. DNNs can adapt themselves beyond their initial programming based on new data inputs, with the advantage that they are not, as compared to traditional ML, limited by a plateau in performance beyond a specific volume of data. However, DNNs require large amounts of data for adequate training. Additionally, DNNs can be further adapted toward specific functions. The most commonly used variation is the convolution neural network (CNN). This is used for image processing and incorporates convolutional layers to extract discriminatory features such as edges, textures, and objects.⁸ A recurrent neural network (RNN) is another variation of DNNs which processes time series data and has value in training with sequential repetitive data.

ML Applications in Cardiac EP

ML can be applied to a wide range of applications in cardiac EP depending on the outcome of interest, the data used for training, and the training approach. Broadly, applications can be useful for a number of functions: (1) automating repetitive tasks; (2) improving the performance of existing tools and technologies; (3) enhancing human performance; (4) reducing human bias; (5) repurposing existing tools with augmented or novel applications; (6) providing new insights into disease mechanisms; (7) deriving new phenotypic classifications of diseases; or (8) aiding in novel drug or gene therapies (Figure 2). Most ML research in EP has focused on utilizing large clinical datasets, raw ECG data, or diagnostic imaging. Emerging ML research has utilized data from wearable electronic devices, implantable electronic devices, or electroanatomic mapping and catheter ablation data.

ML Applications Using ECGs

ECGs are particularly amenable to ML-based research, as they are subject to standardized hardware, filtering, and acquisition methodology allowing for pooling of data within and between institutions. Because ECGs have existed in clinical practice over several decades, there is the opportunity for analyses of long-term outcomes. In addition, ECGs are biologically enriched with potential for complex feature extraction and novel predictions. The seminal work by Attia et al was the first to demonstrate the potential clinical utility of ML in identifying low left ventricular ejection fraction (LVEF) from standard 12-lead ECGs.⁹ A ML algorithm was trained against echocardiographic LVEF using 52,870 patient ECGs and yielding an area under the curve (AUC) of .93 for identifying LVEF ≤35% when tested against a further 44,959 patients. Primary care physicians armed with such an AI-enabled ECG were twice as likely to identify new left ventricular systolic dysfunction (LVSD) in patients.¹⁰ These findings were subsequently extended to the identification of LVSD using smartwatch-based single-lead ECGs in 2454 participants.¹¹ ML applied to 12-lead ECGs of 14,613 participants in the Atherosclerosis Risk in Communities (ARIC) study yielded an AUC of .76 in identifying individuals who developed heart failure (803 participants), with a similar performance to ARIC and Framingham heart failure risk calculators.¹² These and other studies have demonstrated the power of ML to repurpose the ECG for risk stratification and subclinical disease identification for a wide range of clinical conditions (Figure 2). In a study of phospholamban cardiomyopathy, ML models outperformed expert cardiologist interpretation of ECGs in differentiating cases from controls.¹³ Some studies have demonstrated that even when ML falsely identifies an individual to have an adverse phenotype, such individuals have increased risk of future development of that phenotype (eg, LVSD⁹ or aortic stenosis¹⁴), indicating that the ECG can harbor signals predicting the future development of specific end points. The future risk of developing AF can be predicted from a sinus rhythm ECG by an adequately trained ML algorithm with a C statistic of .69.¹⁵

ECGs have also been successfully used to predict age and sex. Training for sex in this study was performed using ECGs from 275,056 patients yielding a 90.4% classification accuracy with an AUC of .97.³ The ECG features determining sex prediction appear to be regulated by testosterone¹⁶ and sex-misclassified individuals have higher mortality compared with those correctly classified.¹⁷ Individuals whose ECG-predicted age is greater than chronological age have increased risk of myocardial infarction, AF, LVSD, and death.^18,19 Both modifiable risk factors such as obesity, hypertension, diabetes, smoking, alcohol intake, and physical exercise, as well as genetic markers of cardiomyopathy (eg, SCN5A) based on a large UK Biobank study (n = 34,432),²⁰ correlate with a large discrepancy between ECG-predicted and chronological age.

In most of these studies, the specific ECG characteristics identified remained “black box” or unexplained. However, some studies have provided insights using explainable AI, in which the algorithms can be disentangled in a way to reveal which of the extracted features provide the highest predictive accuracy. For example, one study found that the most important AI-ECG features for predicting age are T wave morphology indices in leads V4 and V5, and P wave amplitude in leads AVR and II;⁴ while in another study, prolonged PR interval, P wave duration, reduced P wave height, P wave axis, right bundle branch delay, and reduced QRS-T voltage were associated with life-threatening ventricular arrhythmia.²¹ Such explainable AI, though sacrificing predictive accuracy, may be more clinically applicable,²² allow for detection of bias or overfitted models, and improve trust in new models.

Future Challenges of AI and ML in EP

Good AI practice starts by asking a question that can be appropriately answered using a ML approach. The importance of accurate upstream data processing is critical. Additional challenges include correct clinical phenotyping, addressing complex or multiple data types, inter-institutional collaboration, data biases, overfitting, adequate testing, data drift, and compliance with ethical and legal standards or regulations. Importantly, setting study standards and approaches to standardize reporting of results is needed to allow for better appraisal of and comparison between studies.

Although there are hurdles to overcome, advances in AI and ML will bring a monumental shift in health care and transform the field of cardiac EP. They have the potential to answer some of the most important questions at hand: preventing arrhythmic death, optimizing health care quality, and developing novel therapeutics. It is crucial that the EP community continues to embrace these technologies and conceptualize novel ways to incorporate these tools into clinical practice. 

Disclosure: The authors have completed and returned the ICMJE Form for Disclosure of Potential Conflicts of Interest. They report no conflicts of interest regarding the content herein.

References

1. Barker J, Li X, Khavandi S, et al. Machine learning in sudden cardiac death risk prediction: a systematic review. Europace. 2022;24(11):1777-1787. doi:10.1093/europace/euac135

2. Noseworthy PA, Attia ZI, Behnken EM, et al. Artificial intelligence-guided screening for atrial fibrillation using electrocardiogram during sinus rhythm: a prospective non-randomised interventional trial. Lancet. 2022;400(10359):1206-1212. doi:10.1016/s0140-6736(22)01637-3

3. Attia ZI, Friedman PA, Noseworthy PA, et al. Age and sex estimation using artificial intelligence from standard 12-lead ECGs. Circ Arrhythm Electrophysiol. 2019;12(9):e007284. doi:10.1161/circep.119.007284

4. van der Wall HEC, Hassing GJ, Doll RJ, et al. Cardiac age detected by machine learning applied to the surface ECG of healthy subjects: creation of a benchmark. J Electrocardiol. 2022;72:49-55. doi:10.1016/j.jelectrocard.2022.03.001

5. Jang JH, Kim TY, Lim HS, Yoon D. Unsupervised feature learning for electrocardiogram data using the convolutional variational autoencoder. PLoS One. 2021;16(12):e0260612. doi:10.1371/journal.pone.0260612

6. Cui C, Qin H, Zhu X, et al. Unsupervised machine learning reveals epicardial adipose tissue subtypes with distinct atrial fibrosis profiles in patients with persistent atrial fibrillation: a prospective 2-center cohort study. Heart Rhythm. 2022;19(12):2033-2041. doi:10.1016/j.hrthm.2022.07.030

7. Kabra R, Israni S, Vijay B, et al. Emerging role of artificial intelligence in cardiac electrophysiology. Cardiovasc Digit Health J. 2022;3(6):263-275. doi:10.1016/j.cvdhj.2022.09.001

8. Chang AC. Chapter 6 - Other key concepts in artificial intelligence. In: Chang AC, ed. Intelligence-Based Medicine. Academic Press; 2020:141-180.

9. Attia ZI, Kapa S, Lopez-Jimenez F, et al. Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nat Med. 2019;25(1):70-74. doi:10.1038/s41591-018-0240-2

10. Rushlow DR, Croghan IT, Inselman JW, et al. Clinician adoption of an artificial intelligence algorithm to detect left ventricular systolic dysfunction in primary care. Mayo Clin Proc. 2022;97(11):2076-2085. doi:10.1016/j.mayocp.2022.04.008

11. Attia ZI, Harmon DM, Dugan J, et al. Prospective evaluation of smartwatch-enabled detection of left ventricular dysfunction. Nat Med. 2022;28(12):2497-2503. doi:10.1038/s41591-022-02053-1

12. Akbilgic O, Butler L, Karabayir I, et al. ECG-AI: electrocardiographic artificial intelligence model for prediction of heart failure. Eur Heart J Digit Health. 2021;2(4):626-634. doi:10.1093/ehjdh/ztab080

13. Bleijendaal H, Ramos LA, Lopes RR, et al. Computer versus cardiologist: is a machine learning algorithm able to outperform an expert in diagnosing a phospholamban p.Arg14del mutation on the electrocardiogram? Heart Rhythm. 2021;18(1):79-87. doi:10.1016/j.hrthm.2020.08.021

14. Cohen-Shelly M, Attia ZI, Friedman PA, et al. Electrocardiogram screening for aortic valve stenosis using artificial intelligence. Eur Heart J. 2021;42(30):2885-2896. doi:10.1093/eurheartj/ehab153

15. Christopoulos G, Graff-Radford J, Lopez CL, et al. Artificial intelligence-electrocardiography to predict incident atrial fibrillation: a population-based study. Circ Arrhythm Electrophysiol. 2020;13(12):e009355. doi:10.1161/circep.120.009355

16. Naser JA, Lopez-Jimenez F, Chang AY, et al. Artificial intelligence-augmented electrocardiogram in determining sex: correlation with sex hormone levels. Mayo Clin Proc. 2023;S0025-6196(22)00539-0. doi:10.1016/j.mayocp.2022.08.019

17. Siegersma KR, van de Leur RR, Onland-Moret NC, et al. Deep neural networks reveal novel sex-specific electrocardiographic features relevant for mortality risk. Eur Heart J Digit Health. 2022;3(2):245-254. doi:10.1093/ehjdh/ztac010

18. Chang CH, Lin CS, Luo YS, et al. Electrocardiogram-based heart age estimation by a deep learning model provides more information on the incidence of cardiovascular disorders. Front Cardiovasc Med. 2022;9:754909. doi:10.3389/fcvm.2022.754909

19. Lima EM, Ribeiro AH, Paixao GMM, et al. Deep neural network-estimated electrocardiographic age as a mortality predictor. Nat Commun. 2021;12(1):5117. doi:10.1038/s41467-021-25351-7

20. Libiseller-Egger J, Phelan JE, Attia ZI, et al. Deep learning-derived cardiovascular age shares a genetic basis with other cardiac phenotypes. Sci Rep. 2022;12(1):22625. doi:10.1038/s41598-022-27254-z

21. Sammani A, van de Leur RR, Henkens M, et al. Life-threatening ventricular arrhythmia prediction in patients with dilated cardiomyopathy using explainable electrocardiogram-based deep neural networks. Europace. 2022;24(10):1645-1654. doi:10.1093/europace/euac054

22. Simon ST, Trinkley KE, Malone DC, Rosenberg MA. Interpretable machine learning prediction of drug-induced QT prolongation: electronic health record analysis. J Med Internet Res. 2022;24(12):e42163. doi:10.2196/42163