Respiratory diseases pose a significant threat to global health, and a timely and accurate diagnosis of lung conditions is crucial for effective treatment. Traditional auscultation relies heavily on the experience of clinicians, which can result in misdiagnosis, particularly in settings with limited resources. Therefore, developing an automated lung sound classification system can improve diagnostic efficiency, support disease monitoring, and promote intelligent healthcare development.

The methodology follows a systematic workflow. First, a lung sound database is constructed. Patients of various ages and health statuses are recruited. Their lung sounds are collected using professional devices. The sounds are labeled by medical experts as either normal or adventitious (e.g., crackles or wheezes). Then, the sounds are split into training and test sets. Next, the lung sounds undergo feature extraction, which includes time-domain waveforms, frequency-domain waveforms, spectrograms, and Mel-frequency cepstral coefficients, to capture discriminative acoustic information. Then, various algorithms, such as SVM, CNN, XGBoost, RNN, and Transformer, are employed for model learning. Finally, the model’s accuracy and robustness in classifying lung sounds are assessed using metrics such as the confusion matrix, sensitivity, specificity, F1 score, and the AUC-ROC curve.