Home | Publications | SLJ+26

Prior-Aligned Frequency-Domain Explanations for Heart Sound Classification: A Scale-Consistent Attribution Approach

MCML Authors

Björn Schuller

Prof. Dr.

Core PI

Health Informatics

Abstract

Deep learning models for phonocardiogram (PCG) classification often achieve high accuracy. However, they lack transparency regarding clinically meaningful features. Current explainable artificial intelligence (XAI) methods rarely verify whether model decisions align with established cardiac physiology. To address this, we propose Scale-Consistent Attribution (SCA). This training-time regulariser explicitly embeds domain knowledge into the model. SCA aligns spectral attention with a soft clinical prior, distinguishing low-frequency heart sounds from high-frequency murmurs. We validated SCA on the PhysioNet 2016 and CirCor DigiScope datasets. Results demonstrate that SCA remarkably improves the physiological plausibility of explanations. Specifically, it reduces the divergence between model attribution and clinical priors by an order of magnitude. Crucially, the model maintains competitive classification accuracy. Ablation studies with incorrect priors confirm that these benefits stem from accurate clinical medical knowledge. Qualitative and quantitive analysis further reveals that SCA shifts model focus towards clinically relevant frequency regions. Thus, SCA offers a robust framework for trustworthy and explainable PCG classifiers.

article SLJ+26