Embedded machine learning framework for respiratory disease detection

Respiratory diseases such as chronic obstructive pulmonary disease (COPD), pneumonia, asthma, and bronchiectasis remain leading global causes of morbidity and mortality. Traditional diagnostic methods like auscultation are subjective and dependent on the clinician’s expertise, often causing delays or errors. This study proposes a machine learning (ML)-based framework for respiratory disease detection via lung sound analysis, integrated into a real-time embedded system. Two open-source datasets, the International Conference on Biomedical and Health Informatics (ICBHI) dataset and the King Abdullah University Hospital (KAUH) dataset, were used. Lung audio signals were processed with noise reduction, truncation, and normalization, followed by data augmentation. From the processed signals, 30 statistical, temporal, and spectral features were extracted. To address class imbalance, SMOTE combined with Tomek links was applied. Various ensemble ML algorithms and deep learning architectures were trained. The ExtraTrees (ET) classifier achieved 99.87% accuracy with 10-fold cross-validation on ICBHI dataset and 93.68% accuracy on the external KAUH dataset, demonstrating its strong generalizability. An AutoGluon-based WeightedEnsemble L3 achieved 99.97% accuracy on the ICBHI test set and 95.65% on KAUH dataset, further proving its strong generalizability on new unseen data. To ensure interpretability, explainable AI techniques, including feature importance, SHAP, LIME, and perturbation analysis, were employed. Among DL architectures, LSTM achieved 99.58% accuracy with minimal loss. Finally, the ET model was deployed on a Raspberry Pi 3-based embedded system, which records and processes 10-second lung sounds, predicts disease, and displays results in real time. This research delivers an effective, interpretable, and portable respiratory disease screening system, bridging AI with real-world healthcare applications.