Prior-aligned frequency-domain explanations for heart sound classification: a scale-consistent attribution approach.

Objective: Deep learning models for phonocardiogram (PCG) classification often achieve high accuracy but lack transparency in relation to clinically meaningful features. Existing explainable artificial intelligence (XAI) methods rarely ensure alignment between model attributions and cardiac physiology. Methods: We propose Scale-Consistent Attribution (SCA), a training-time regularization method that incorporates domain knowledge into model learning. SCA aligns spectral attention with a soft clinical prior, distinguishing low-frequency heart sounds from high-frequency murmurs. The method is evaluated on the PhysioNet 2016 and CirCor DigiScope datasets. Results: 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. Conclusion: SCA offers a robust framework for trustworthy and explainable PCG classifiers, improving the clinical plausibility of explanations while maintaining strong predictive performance.