Detection and Classification of Cardiac Arrhythmias by Machine Learning: A systematic review

Rafael Fernandes and João Salinet
Biomedical Engineering, Engineering, Modelling and Applied Social Sciences Centre


Introduction: Machine learning (ML) techniques can perform as better as humans at key healthcare tasks. Recent advances make it possible to perform, using ML, automatic high-level feature extraction and classification of cardiac arrhythmia. In this work, we aimed through a systematic literature review identify the principal methods, databases and contributions of ML on cardiac arrhythmias classification.

Methods: Electronic database including PubMed, Science Direct, IEEE, Scielo, Scopus and Web of Science were searched, from 2014 to 2019, by combining the following keywords "ECG", "heart signals", "arrhythmia", "classification" and "machine learning". Outcomes of interest included the strategies applied for the algorithms, the database and metrics used.

Results: A total of 524 articles were initially identified. 28 studies selected as eligible for this research and included in the final analysis. ECG signals were used by 26 articles and classification,were based on heart rate, morphological characteristics and temporal ECG metrics. The remaining 2 studies were based on body surface potential mapping signals. Classifications classes ranged from 2 to 17, with prevalence of 2 classes (71.4% of the studies). The most frequent applied methods were Artificial Neural Network (10 articles), followed by Support Vector Machines and Mixed techniques (5 articles respectively). MIT-BIH Arrhythmia Database was used in 16 studies (57%), whereas 6 (21.4%) utilized their own data. The approaches basis for evaluating the results is the confusion matrix, where up to 67.8% of the studies used precision, sensitivity and specificity, and 42.8% used accuracy or AUC-ROC.

Conclusion: Classification of cardiac arrhythmias through ECG is of increasing interest from the research groups, and ML classification is showing rising levels of performance. It would benefit both patients and clinicians. Public databases and scientific challenging competitions based on clinical problems (such PhysioNet - Computing in Cardiology Challenge) have contribute without precedents to the evolution of the field.