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

Area of Science:

• Artificial Intelligence
• Biomedical Signal Processing
• Cardiology

Background:

• Deep learning models for phonocardiogram (PCG) classification achieve high accuracy but often lack transparency.
• Existing explainable artificial intelligence (XAI) methods struggle to align model attributions with crucial cardiac physiology.
• This limits the clinical utility and trustworthiness of AI in cardiovascular diagnostics.

Purpose of the Study:

• To introduce Scale-Consistent Attribution (SCA), a novel training-time regularization technique for PCG classification models.
• To ensure that AI-driven explanations for PCG analysis are aligned with established clinical knowledge and cardiac physiology.
• To improve the interpretability and reliability of deep learning models in cardiology.

Main Methods:

• SCA incorporates domain knowledge by aligning spectral attention with a soft clinical prior during model training.
• This approach specifically distinguishes between low-frequency heart sounds and high-frequency murmurs.
• The method was evaluated on the PhysioNet 2016 and CirCor DigiScope datasets.

Main Results:

• SCA significantly enhances the physiological plausibility of model explanations, reducing attribution divergence from clinical priors by an order of magnitude.
• The model maintained competitive classification accuracy, demonstrating the effectiveness of SCA.
• Ablation studies confirmed that the benefits are derived from accurate clinical knowledge, and qualitative analysis showed a shift towards clinically relevant frequency regions.

Conclusions:

• SCA provides a robust framework for developing trustworthy and explainable PCG classifiers.
• The method improves the clinical relevance of AI explanations while preserving high predictive performance.
• This represents a significant advancement in applying AI to cardiovascular diagnostics.