Quantitative Attention Validation and Continual Learning for ECG Arrhythmia Detection with Elastic Weight Consolidation

Abstract

Deep learning to predict ECG arrhythmia has a reputation of achieving accuracies of over 98% but this is usually due to intra-patient splits which inflate results and do not extrapolate to clinical locations. Our interpretable CNN-BiLSTM-Attention model was used to classify binary arrhythmia based on patient-level splitting and cross-dataset transfer learning between the Chinese Shaoxing database and MIT-BIH. Stage 2b Our model provided 88.1 percent AUC on test data (MIT-BIH). The statistical analysis showed that the arrhythmic and normal rhythms involve quite different attention mechanisms (Cohen d = 0.93, p < 0.001) as arrhythmic samples involve more broadband temporal integration and normal rhythms rely on narrowband feature detection. The learning experiments of stage 3 curriculum revealed the retention-plasticity paradox of continual learning. Elastic Weight Consolidation had two solutions, EWC-Highest, which was the best with the highest MIT-BIH performance, high-level attention discrimination, and EWC-Balanced with the best complex pattern learning and acceptable retention. This study can be used to show that methodologically sound 88% accuracy with interpretable attention is more clinically useful than standards set by less generalizable evaluation procedures.