On-device cough detection and respiratory disease classification enhanced by generative data augmentation.

Area of Science:

• Artificial Intelligence
• Biomedical Signal Processing
• Digital Health

Background:

• Cough sounds are valuable, non-invasive biomarkers for respiratory health.
• Current smartphone-based cough analysis faces limitations in deployability, privacy, and data scarcity.
• Existing methods often rely on server-based deep learning, hindering real-world application.

Purpose of the Study:

• To develop a smartphone-compatible AI framework for automated cough detection and respiratory disease classification.
• To introduce a novel generative augmentation strategy using Variational Autoencoder (VAE) variants for improved classification under data scarcity.
• To enhance disease classification accuracy while addressing limitations of conventional audio augmentation.

Main Methods:

• A three-module AI framework: Cough Detection Module (CDM) for real-time segmentation, Disease Analysis Module (DAM) for classification (asthma, COVID-19, healthy), and Generative Augmentation Module (GAM) using five VAEs.
• GAM operates in the time-frequency domain for feature optimization and reconstructs in the time-domain for acoustic verifiability.
• Lightweight models optimized for on-device execution and Support Vector Machine classifiers with probabilistic fusion.

Main Results:

• The CDM ensures reliable cough segmentation across diverse conditions.
• The DAM effectively discriminates between asthma, COVID-19, and healthy coughs using cepstral and spectral features.
• The GAM framework successfully alleviates class imbalance, with clinically verifiable synthetic biomarkers, all operating in real-time on Android devices.

Conclusions:

• The integrated framework overcomes limitations in dataset availability, model generalizability, and deployability.
• It presents an interpretable generative approach for scalable and privacy-preserving respiratory health monitoring using smartphones.
• Smartphone-based acoustic sensing is demonstrated as a feasible tool for remote health surveillance.