Physiology-Aware Temporal Attention and Frequency Band Selection for Robust Respiratory Sound Analysis.

Abstract

We present a physiology-aware and efficient framework for respiratory sound classification. The framework combines a CNN backbone for local feature extraction with Phase-Informed Saliency Temporal Attention (PISTA), which integrates respiratory-phase cues, acoustic salience, and temporal locality to focus attention on diagnostically relevant time regions. An importance-guided Frequency Band Selection (FBS) module removes non-relevant spectral bands. Multi-axis Group Distributionally Robust Optimization (GroupDRO) improves robustness to class, demographic, and recording-condition imbalance. Evaluation is conducted on ICBHI 2017 and SPRSound 2022/2023, with real-time profiling on a Raspberry Pi 3. The method achieves state-of-the-art performance on SPRSound 2022/2023 and competitive results on ICBHI 2017, while FBS reduces spectral dimensionality and computation by up to 50%. Integrating the proposed modules into transformer baselines yields new state-of-the-art results on ICBHI. Real-time inference on Raspberry Pi 3 confirms edge feasibility. The proposed system enables accurate and efficient respiratory sound classification. The framework supports reliable, real-time auscultation in low-resource and point-of-care settings.