MeDeA: Decomposable attention for intrinsically interpretable ECG diagnosis

Abstract

Deep learning has achieved strong performance on electrocardiogram (ECG) classification, yet many high-performing models still provide limited class-specific interpretability. We present MeDeA (multi-disease decompositional attention), a decomposable attention framework for intrinsically interpretable ECG diagnosis. The central idea is to combine a shared 1D backbone with disease-specific attention heads so that each predicted condition is associated with a dedicated explanatory pathway instead of relying solely on post-hoc attribution. We further describe a multi-query formulation, termed multi-query attention head (MQAH), in which each disease head contains a small set of global learnable queries to capture complementary morphological patterns. Experiments on the PTB-XL dataset show that MeDeA achieves a macro-F1 score of 0.775, matching strong black-box baselines while improving attribution faithfulness. We also report efficiency metrics including parameter count, weight footprint, and forward latency. Finally, we discuss limitations related to retrospective public datasets, shifts across hospitals and devices, sampling-rate mismatch, and annotation quality.