Structured auscultation evidence modeling for multi-protocol respiratory sound classification

Digitally recorded respiratory sounds provide a practical source of physiological information for computer-assisted auscultation. Their acoustic presentation varies across subjects, recording locations, devices, filtering conditions, and environmental noise. Most classification models map acoustic representations directly to disease labels, which provides limited access to the auscultation findings associated with each prediction. This study investigated whether respiratory acoustic representations could be organized into structured evidence for classification across different label settings and evaluation levels.

We developed Evi-CARD to organize respiratory sound representations into structured evidence for classification. A frozen Open Respiratory Acoustic Foundation Model (OPERA) encodes overlapping $8s$ segments, from which a concept-aligned pathway estimates wheeze, crackle, inspiration, and expiration together with learned evidence weights. An attention-constrained residual pathway preserves complementary acoustic information, although its attention ceiling does not strictly bound the residual pathway's contribution to the final logits. Separate disease-query layers adapt the shared evidence representation to each classification setting. We assessed BRACETS A2 and A3 by subject-independent five-fold validation at the recording level and KAUH-primary by subject-independent five-fold validation at the patient level for Healthy vs. Obstructive classification. Each assessment was conducted within its source cohort and was not an external-validation experiment.

On BRACETS, Evi-CARD achieved balanced accuracies of 77.8% for the A2 three-class task and 70.2% for the more granular A3 five-class task. On KAUH-primary, pooled patient-level out-of-fold evaluation yielded an area under the receiver operating characteristic curve of 95.1%, a balanced accuracy of 88.0%, a Macro F1 of 88.1%, and an expected calibration error of 5.1%. Across HF_Lung_V1, SPRSound, and ICBHI, the complete concept layer also provided more accurate and better-calibrated concept predictions than the evaluated acoustic-probe and concept-readout baselines.

These findings show that protocol-specific respiratory sound classification can be based on concept-associated latent evidence while retaining a constrained route for information outside the predefined concept set. However, Evi-CARD is not a strict scalar concept bottleneck, and no clinical meaning is assigned to its residual-memory slots. The evidence is limited to within-cohort cross-validation; independent-cohort testing, device- and site-stratified evaluation, and clinician review remain necessary before considering clinical deployment.