Deep Learning for Audio-Based Respiratory Disease Classification Using Clinical Sound Data

Respiratory diseases, including asthma, Chronic Obstructive Pulmonary Disease (COPD), and pneumonia, are common global health concerns that contribute to morbidity and mortality worldwide. These conditions often present with an initial symptom such as a cough. Early and accurate diagnosis is important for effective management and improved patient outcomes. Traditional diagnostic methods, however, are often limited by subjectivity and external noise interference, emphasizing the need for innovative approaches. This research investigates the use of hybrid deep learning models for classifying respiratory diseases using clinical sound data. There were 8 classes included in the data including healthy with seven diseases; COPD, URTI, Bronchiectasis, Bronchiolitis, Pneumonia, LRTI, and Asthma. Data was simplified into only 2 classes; COPD and non-COPD (which includes healthy and the other 6 diseases). Utilizing audio recordings from the Respiratory Sound Database, this study applies data preprocessing, feature extraction, and a classification model for 2 classes (COPD and non-COPD). The study compares three hybrid architectures-CNN + RNN, CNN + LSTM, and CNN + SVM-to determine the optimal model. Metrics like F1-score, recall, accuracy, and precision are used in evaluating models. The results showed an overall great performance in CNN + LSTM with an accuracy of ${8 8 \%}$. CNN + SVM also showed good capability in classifying the data with an accuracy of 85 %. However, CNN + RNN performed quite poorly due to the incapability in capturing and remembering the patterns of the audio data. This research contributes to the development of AI-driven diagnostic tools as an early warning to respiratory diseases.