Skip to content
Open access

Domain-Adaptive Pretraining and a Leakage-Corrected Evaluation Protocol for Respiratory Sound Classification

2026 · IEEE Access · Vol 14, pp. 151079-151100 · 0 citations · 45 references
Computer Science

Abstract

Automated respiratory sound classification could make pulmonary screening scalable, but progress on its main benchmark, the International Conference on Biomedical and Health Informatics (ICBHI) 2017 database, is hampered by inconsistent protocols and limited reproducibility. We contribute respiratory audio self-distillation (RASD), a domain-adaptive pretraining method pairing a decoder-side masking bottleneck with scheduled Gaussian corruption of visible token features. The adapted encoder drives a metadata-free classifier aggregating overlapping two-second instances by attention. Across five runs on the official test split it attained an ICBHI score of $64.30~\pm ~0.81$ %, specificity 84.34 % and four-class sensitivity 44.25 %. A controlled study over three pretraining draws isolated the position of the masking bottleneck as the one change costing more than a point at every draw, showed the pretraining seed to matter as much as any change tested, and showed that removing the background-noise mixing collapses the seed-to-seed spread from 2.52 points to 0.10; we therefore report the standard deviation to expect when both stages are reproduced, 1.41 points against 0.81 with the encoder fixed. We document and correct an unreported patient-level leakage in the official split: the aggregate score is unchanged (64.28 %), but the contamination falls almost entirely on the rarest class, whose recall drops 13.79 points. Re-applying the method to SPRSound, a pediatric corpus recorded with different equipment, surpasses every published entry to its IEEE BioCAS Grand Challenge ( $0.7621~\pm ~0.0084$ against 0.6932); what is conditional on matching the pretraining audio to the classified unit is the contribution of the domain-adaptive stage within that margin. Code, checkpoints, confusion matrices, and the corrected split are released.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.