Construction and validation of a lightweight lung sound classification model assisted by physical simulation

To improve the objectivity and generalization of lung sound analysis, this study proposes a physics-informed workflow for automatic lung sound recognition. We reconstructed a 3D airway model from clinical CT scans and conducted coupled fluid dynamics and aeroacoustic simulations to analyze the airflow distortion mechanisms responsible for adventitious lung sounds. Guided by these aeroacoustic priors, we developed a data generation algorithm that synthesizes pathologically plausible acoustic events via statistical sampling, mitigating the long-tail distribution typical of medical datasets. We also introduce the Temporal-Spectral Fusion (TSF) network, a classification architecture based on multi-task contrastive learning. The TSF model incorporates an asymmetric time–frequency pyramid, dual-statistic frequency-axis pooling, and a large-kernel Conformer module to capture the temporal inertia of long-wave fluid dynamics, integrated with a class-aware GroupMix strategy. Evaluations show that the TSF network achieves a score of 59.82 on the ICBHI 2017 benchmark and approximately 90% binary classification accuracy on an independent clinical cohort. With a parameter size of 2.98 M and low computational overhead, the model was deployed onto a wearable device for integrated auscultation and sputum clearance. Preliminary testing conducted at an elderly care facility confirmed the engineering feasibility, safety, and usability of the closed-loop system.