Smartphone-Based Detection of Abnormal Lung Sounds Using Domain-Adaptive Deep Learning

Auscultation is a fundamental clinical practice for detecting respiratory abnormalities, but accurate interpretation requires specialized medical expertise. With the increasing adoption of electronic stethoscopes, large-scale datasets have enabled active research on deep learning for the detection of abnormal lung sounds. However, electronic stethoscopes remain costly and impractical for widespread use in everyday care. Smartphones, by contrast, are ubiquitous and equipped with sensitive microphones, offering a promising alternative for accessible auscultation and remote patient monitoring (RPM). Nevertheless, the scarcity of labeled smartphone recordings and acoustic disparities between consumer-grade microphones and medical-grade stethoscopes pose significant challenges for developing robust deep learning models. To address this gap, we recruited 611 pediatric patients, of whom 147 provided immediate-sequence paired recordings using both an electronic stethoscope and a smartphone, enabling direct analysis of device-specific spectral shifts. An additional external cohort of 48 patients was recorded exclusively with smartphones at a separate site to support independent validation. Using these datasets, we developed a domain-adaptive deep learning framework that integrates a smartphone-specific preprocessing pipeline to mitigate acoustic mismatch with a directional MixStyle module embedded in an audio spectrogram transformer, thereby adapting stethoscope features toward the smartphone domain. The framework was rigorously evaluated, achieving AUROCs of 0.911 on the primary dataset and 0.862 on the external smartphone-only cohort, consistently outperforming models trained solely on smartphone recordings. These findings demonstrate that combining stethoscope and smartphone data through domain adaptation enables accurate smartphone-based detection of abnormal lung sounds, establishing a validated pathway for scalable and cost-effective in-home respiratory RPM.