Harnessing Machine Learning for Predicting Atrial Fibrillation from Electrocardiogram Signals

Background: Deep learning (DL) models can discern electrocardiogram (ECG) reports to identify individuals at risk for atrial fibrillation (AF). The etiology of AF in India differs from the West, with a high prevalence of valvular heart disease (mainly rheumatic). This study evaluates a DL model of 12-lead ECG reports comparing normal sinus rhythm with intermittent AF to screen high-risk populations for AF. Methodology: Electronic medical records of 69,700 patients (200,650 ECGs) attending a tertiary cardiac center from December 2010 to October 2024 were searched. The 10s, 12-lead ECGs of patients with intermittent AF (sinus rhythm with documented AF) or normal sinus rhythm (without history of AF or risk factors of AF) were classified with a DL model (3304 records for training, 159 for testing) of one-dimensional convolutional neural network (CNN). At a 50% threshold, the model was adjusted for hyperparameters, evaluated for performance, and screened individuals with risk factors such as diabetes mellitus (DM), hypertension (HTN), valvular disease, and ischemic heart disease (IHD). Results: The model generation used 1956 records with normal sinus rhythm and 1507 with intermittent AF. The sensitivity and specificity were 85% and 69% (positive predictive value: 73%, negative predictive value: 82%; AUC: 0.84-0.88). The model assessed risk of AF was: 327/516 in post-IHD, 235/502 in HTN and DM, 137/387 in DM and 499/1040 in HTN, and 902/1398 in valvular disease. Conclusion: The CNN model developed based on the ECG signatures offers 85%–90% sensitivity in screening individuals at risk of developing AF. The prediction could facilitate early intervention to prevent the sudden onset of complications, including stroke.