Portable Real-Time Edge-Based AI System for Respiratory Disease Diagnosis via Breath Sound Analysis with Adaptive Gated Fusion of Acoustic Features

Abstract

Respiratory diseases remain a leading global cause of morbidity and mortality, especially in low-resource settings with diagnostic delays. Early detection improves outcomes, but conventional auscultation is subjective and inconsistent. This study presents a portable real-time Edge-AI system for respiratory disease screening, operating fully on-device to ensure privacy, low latency, and offline use. Breath sounds (normal, asthma, bronchitis, and pneumonia) were classified using Mel-Frequency Cepstral Coefficients (MFCC) and formant features. An adaptive gated fusion (AGF) balances feature contributions, and a lightweight bidirectional long short-term memory (BiLSTM) captures temporal patterns efficiently. On the Kaggle lung sound dataset, the system achieved 80.6% accuracy, 80.7% precision, 80.6% recall, and 80.6% F1-score (≤1.5% variance), outperforming existing deep model baselines by +3.5% accuracy (p < 0.05). Deployment on Raspberry Pi 4 showed ~180 ms latency per 10 s sample, ~42% CPU usage, and 7.5 h battery life. Field tests with 50 volunteers confirmed noise robustness and usability. These results highlight Edge-AI breath sound analysis as a scalable, privacypreserving tool for respiratory screening in resource-limited settings.