A clinically-oriented XAI framework for arrhythmia triage via Hierarchically-Decomposed Neural Attribution: From dual-stream interpretation to risk stratification.

Deep learning models for biomedical signal analysis achieve high accuracy but suffer from interpretability limitations that hinder their deployment in critical applications. This paper presents the Hierarchically-Decomposed Neural Attribution (HDNA) framework, a novel explainable AI (XAI) methodology for biomedical data analysis with application to electrocardiogram (ECG) signal processing. HDNA employs a dual-stream architecture that decomposes the decision-making process into two sequential, interpretable components: a rhythmicity stream that identifies critical temporal events through contrastive learning-based deviation analysis, and a morphology stream that performs targeted gradient-based feature localization on these events. The framework's "Where-then-What" reasoning pipeline incorporates a structural gating mechanism that reduces computational overhead by 90% compared to full-sequence gradient methods, achieving real-time inference (2.4 ms latency) on standard CPU hardware. As a case study in computational biology, we evaluated HDNA on cardiac arrhythmia detection using the PTB-XL (Macro F1: 0.760 ± 0.008) and MIT-BIH (Weighted F1: 0.89 ± 0.01) datasets, demonstrating robust performance with high explanation faithfulness (93.0% pointing accuracy, 0.78 ± 0.04 IoU). The framework's lightweight design (1.4M parameters) and efficient explanation generation make it suitable for resource-constrained systems and real-time biomedical signal monitoring applications. This work contributes to the field of computational biology by providing a transparent, interpretable machine learning methodology that can be applied to diverse biomedical data analysis tasks.