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

Area of Science:

• Medical Technology
• Artificial Intelligence in Healthcare
• Respiratory Medicine

Background:

• Auscultation is key for respiratory diagnostics but requires expertise.
• Electronic stethoscopes enable deep learning for lung sound analysis but are costly.
• Smartphones offer accessible auscultation and remote patient monitoring (RPM) potential.

Purpose of the Study:

• To develop a robust deep learning framework for detecting abnormal lung sounds using smartphone recordings.
• To address challenges of limited labeled smartphone data and acoustic disparities.
• To validate the framework's performance on independent datasets.

Main Methods:

• Recruited 611 pediatric patients with paired electronic stethoscope and smartphone recordings.
• Developed a domain-adaptive deep learning framework with a smartphone-specific preprocessing pipeline.
• Integrated a directional MixStyle module within an audio spectrogram transformer for acoustic adaptation.

Main Results:

• Achieved Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.911 on the primary dataset.
• Attained an AUROC of 0.862 on an external smartphone-only cohort.
• Demonstrated superior performance compared to models trained solely on smartphone data.

Conclusions:

• Domain adaptation enables accurate smartphone-based detection of abnormal lung sounds.
• This approach establishes a scalable and cost-effective pathway for in-home respiratory RPM.
• Validated framework facilitates accessible respiratory health monitoring using ubiquitous smartphone technology.