Abstract Background COPD imposes substantial global disease burden, yet spirometry, the gold standard for airflow limitation assessment, is underutilized due to barriers in cost, equipment, and expertise. Improving accessibility of airflow limitation screening remains critical. Auscultation, a simple and accessible method, combined with digital analysis and AI, holds high translational potential. This study aimed to develop a deep learning system to detect airflow limitation from respiratory sounds (recorded via clinical electronic stethoscopes), validate its real-world utility in predicting FEV1/FVC 70%, and evaluate its performance in potential COPD screening when combined with the CAPTURE questionnaire, offering a novel deployable tool for COPD screening and management. Methods A two-stage study with rigorous “temporal + spatial” dual-dimensional validation was conducted. Stage 1: Prospective collection of clinical data, spirometry, and respiratory sounds from 5 centers across China (A-E). Task 1: Model development using data from centers A/B, based on a Transformer-based Masked Autoencoder (MAE) pre-trained model. Lung sound spectrograms, after preprocessing (denoising, quality screening), were used as input to identify airflow limitation. Task 2: Spatial external validation using data from centers C/D/E, with evaluation of multi-position auscultation (axillary midline, inferior scapular margin, midclavicular line). Stage 2: Temporal prospective validation (data prospectively collected after model training, ongoing). A subset underwent screening via model + CAPTURE (risk determined by integrating airflow limitation output and CAPTURE scores), compared with clinical diagnoses. As of October 1, 2025, 20062 lung sound recordings (3269 of AB quality) and 1650 spirometry reports were included. Findings In Stage 1, internal validation (A/B) showed an AUC of 0.842 (95% CI 0.838-0.846); spatial validation (C/D/E) showed an AUC of 0.833 (95% CI 0.821-0.845), with a sensitivity of 91.6% and specificity of 86.1%. Axillary midline and inferior scapular margin positions outperformed the midclavicular line; patients with complete 6-position data had a significantly higher F1-score. In Stage 2 (ongoing prospective collection), the model maintained favorable accuracy in airflow limitation identification; additionally, the model + CAPTURE strategy achieved 76% positive and 92% negative identification rates against clinical COPD diagnoses. Interpretation This respiratory sound-based airflow limitation detection model serves as a viable alternative to spirometry for identifying potential cases, demonstrating robust potential for COPD assessment via auscultation. Rigorous dual-dimensional validation supports its reliability; when combined with CAPTURE, it enhances screening utility, offering an accessible alternative to spirometry to reduce missed COPD diagnoses. This abstract is funded by: None
Liao et al. (Fri,) studied this question.