Comparative modeling of mixed cardiopulmonary sounds in a low-resource paired dataset: Discrimination, calibration, and operating-point behavior

Abstract

Background Mixed cardiopulmonary recordings are common in bedside auscultation, yet most automated systems have been developed for isolated heart sounds or isolated respiratory sounds. Methods We conducted a comparative methods study on HLS-CMDS, a low-resource paired dataset containing mixed recordings with matched isolated heart and lung source recordings. The task was dual binary classification from a single mixed recording. We compared feature-based references, a shared-backbone multitask CNN, a target-domain student model, teacher-guided variants pretrained on PhysioNet/CinC 2016 and ICBHI 2017, and lighter source-aware variants using paired HLS-CMDS source recordings. A nested grouped five-fold evaluation was performed at the triplet level; within each outer training fold, an inner validation split was used for checkpoint selection, temperature scaling, and task-specific threshold selection. Results Under the revised nested evaluation, the light source-aware model showed the strongest mean discrimination (macro AUROC 0.7107 ± 0.1659; macro AUPRC 0.9318 ± 0.0423). The prevalence-defined no-skill macro AUPRC baseline was 0.8586 ± 0.0225. After inner-validation temperature scaling and threshold selection, the calibrated student-only model achieved the highest mean macro balanced accuracy (0.6894 ± 0.0548). The observed differences were interpreted cautiously because fold-to-fold variability was substantial. Conclusions In this small paired mixed-sound setting, restrained source-aware guidance showed the strongest discrimination tendency, whereas a simpler target-domain model achieved the best threshold-dependent balanced accuracy after calibration.