Answers are grounded in this paper and cite their sources.
Introduction Respiratory sounds reflect pathological respiratory events but are also influenced by physiological, anatomical, and acquisition-related factors. Distinguishing these sources of variability is essential for reliable computational interpretation across heterogeneous recording environments. This study investigated associations between respiratory acoustic variability and sound labels, demographic characteristics, and acquisition-related factors, while evaluating their implications for cross-dataset classification performance. Methods Using the ICBHI dataset, we examined marginal sound-label associations and fitted multivariable models incorporating wheeze status, age, sex, auscultation location, recording device, and disease category, with participant-clustered inference. Associations were quantified using marginal and partial eta-squared effect sizes. Distribution shifts were examined across four respiratory sound datasets: ICBHI, SPRSound, SNUBH, and SMART. Four deep learning backbones were trained on the ICBHI training set for binary wheeze versus non-wheeze classification and evaluated across all four datasets. Standard cross-entropy fine-tuning was compared with metadata-aware domain adversarial training (DAT) using age, sex, location, or device as the supervision target. Results Associations between lung sound labels and acoustic features varied across feature summaries and label definitions, with Fbank-mean η² values of 0.04792 across four sound categories, 0.00657 for normal versus wheeze-containing sounds, and 0.03428 for normal versus any adventitious sound. In covariate-adjusted analyses, recording device remained strongly associated with Fbank mean (partial η² = 0.6764; 95% participant-bootstrap interval: 0.5393–0.7825). Substantial distribution differences persisted across datasets within shared sound-label, age-group, and auscultation-location categories. Classification performance under standard fine-tuning varied across external datasets, whereas metadata-aware DAT improved performance across multiple backbones and datasets. For the Audio Spectrogram Transformer, device-guided DAT increased the ICBHI Score from 73.39 to 77.87 on ICBHI and from 79.05 to 85.30, 77.70 to 83.24, and 57.81 to 67.07 on SNUBH, SMART, and SPRSound, respectively. Discussion Respiratory acoustic variability was associated with population and acquisition characteristics, while sound-label associations depended on feature summaries and label definitions. These findings highlight the importance of accounting for physiological and acquisition-related variability when interpreting respiratory acoustic signals and developing computational models intended to generalize across heterogeneous populations, devices, and recording environments.
No takes yet. Share an insight, caveat, or question.
Kim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: