Reliability-aware transfer learning with BEATs for variable-length respiratory sound classification: a...

Random or central placement of respiratory sounds within padded windows can cause length-based masks to exclude recorded audio. We developed an exact support-propagation interface for the pretrained bidirectional encoder representation from audio transformers (BEATs), mapping sample support through filterbank frames and patch tokens to attention and pooling masks. We evaluated this interface alongside adaptation strategies, computational costs, probability calibration and class-conditional coverage. The study included 15 987 events from the International Conference on Biomedical and Health Informatics (ICBHI) 2017 and SPRSound databases, using patient-disjoint training, model-selection, calibration and primary test partitions at a fixed decision threshold of 0.5. The three-seed full-fine-tuning ensemble achieved average scores (AS; mean sensitivity and specificity) of 0.6619 on ICBHI (patient-cluster bootstrap 95% confidence interval, 0.6002-0.7152) and 0.9112 on SPRSound (0.8799-0.9362). Exact mapping corrected observed support exclusion and maintained finite outputs throughout placement tests; the length-based interface produced non-finite outputs in these tests and one of three matched training runs. Classification ablations showed no consistent advantage for exact mapping across the two completed seed pairs and no incremental gain from Jensen-Shannon consistency. Under matched update counts and learning-rate schedules, full fine-tuning exceeded rank-8 low-rank adaptation (LoRA) by 3.94 percentage points in AS averaged equally across databases (paired patient-cluster bootstrap 95% confidence interval, 2.54-5.40). LoRA reduced trainable parameters and recorded training time but did not reduce inference latency in the unmerged implementation. Temperature scaling improved the primary ensemble's held-out negative log-likelihood and Brier scores without uniformly improving calibration error. Entropy-based abstention achieved low accepted-set error but failed the prespecified validation requirement for adventitious-event coverage and was not applied. The integrated evaluation quantifies performance-resource trade-offs and shows that improved probability scores do not ensure adequate class-specific coverage for selective prediction. Clinical utility requires prospective validation.