Interpretable arrhythmia detection in ECG scans using deep learning ensembles: a genetic programming approach

Cardiovascular diseases remain the leading cause of death in developed countries. This study introduces deep learning ensembles for arrhythmia detection and atrial fibrillation (AF) recurrence prediction from electrocardiogram scans, supported by explainable artificial intelligence (XAI) methods. Validation used two datasets: Guangdong Provincial People's Hospital, China (Dataset G, 1172 patients, 71.4 ± 6.3 years, 66% women, 20.5% with arrhythmia) and Liverpool Heart and Chest Hospital, UK (L, 909 patients, 60.5 ± 10.71 years, 33% women, 29.7% with arrhythmia). Our ensembles outperformed individual and voting models with the area under the receiver operating characteristic curve (ROC-AUC): 0.980 (95%CI: 0.956-0.998, p = 0.03) for Dataset G, 0.799 (95%CI: 0.737-0.856, p = 0.07) for Dataset L. The models trained on combined training sets achieved ROC-AUC: 0.980 (95%CI: 0.952-1.0) and 0.800 (95% CI: 0.739-0.861) for the G and L test sets. Precision-recall AUC for AF recurrence was 0.765 (95%CI: 0.669-0.849) for ensembles vs. 0.737 (95%CI: 0.648-0.821) for individual models. XAI enhanced interpretability for clinical applications.