ECG synthesis for cardiac arrhythmias: Integrating self-supervised learning and generative adversarial networks.

Arrhythmia classifiers relying on supervised deep learning models usually require a substantial amount of labeled clinical data. The distribution of these labels is strictly related to the statistics of cardiovascular diseases among the population, which inherently narrows models' performance for classification tasks. Furthermore, during acquisition and data retrieval from electronic health records, concerns arise regarding patient anonymization due to stringent clinical policies. We introduce a conditional generative architecture for electrocardiography time series, which integrates self-supervision and generative adversarial principles. Empirical validation confirms the enhancement of morphological plausibility in synthetic data, showcasing its effectiveness in generating realistic signals. We propose a novel model (ECGAN), proving its capability of conditioning the probability distribution of ECG recordings. The proposed methodology is assessed upon various rhythm abnormalities including severe congestive heart failure, myocardial infarction, sinus rhythm, and premature ventricular contractions. Our proposed workflow for synthetic time series assessment demonstrates competitive performance compared to state-of-the-art models, achieving an average improvement of 2.4% in arrhythmia classification accuracy across MIT-BIH, BIDMC, and PTB datasets, while ensuring realistic synthetic data and improving training stability.