ECG-SENNsation: Trustworthy 12-Lead ECG Classification by Self-Explaining Neural Networks for Clinical Decision Support and Knowledge Discovery

Background: Deep learning models show promise for detecting cardiovascular diseases in electrocardiograms (ECGs), but their black-box nature hinders clinical adoption. Current explainable AI (xAI) only provides approximate explanations for the models' decision-making. However, reliable explanatory heatmaps are essential for meaningful, AI-assisted ECG analysis, optimizing resources in clinical routines and opening opportunities for knowledge discovery. Methods: We developed a self-explaining neural network (SENN) to generate reliable heatmaps per class for 12-lead ECG classifications and utilized the PTB-XL dataset to detect five super diagnostic classes. Our SENN-based heatmaps were compared with 12 other xAI techniques for stability. The model focus was evaluated through the analysis of heatmaps across ECG segments. False classifications are reviewed by an experienced cardiologist and features revealed by the heatmaps are compared to clinical knowledge. We further identified differences in the model focus of patient heatmaps and tested for morphological group differences. Results: ECG-SENNsation achieved a ROC-AUC score of 93.1 % and provided class specific (p<0.0001) and the most stable (p<0.0001) explanations. Clustering by model focus in myocardial infarction (MI) heatmaps identified patient groups with distinct morphological characteristics. In the commonly neglected aVR lead, ECG-SENNsation identified a distinct morphological interrelationship between T wave and QRS complex. Derived from this, a novel marker has been developed to discriminate between specific MI subgroups with a large effect size (p<0.0001, r _rb =0.78). Conclusions: ECG-SENNsation provides high classification performance and robust explanations for trustworthy clinical decision support. It generates reliable heatmaps for knowledge discovery, revealing differences in morphology, currently not used in standard ECG diagnostics. Our methodology shows great potential for disease marker identification in the ECG and biosignals in general.