Abstract:

Asthma, COPD, pneumonia, lung fibrosis, and other respiratory conditions constitute a substantial worldwide health burden, and prompt identification is essential to successful treatment. Despite their clinical value, traditional diagnostic techniques including spirometry, chest X-rays, CT scans, and arterial blood gas analysis have drawbacks such invasiveness, radiation exposure, reliance on subjective interpretation, and restricted accessibility. This study investigates the automatic classification of respiratory sounds using deep learning models in order to overcome these difficulties. Preprocessing involved resampling, normalization, and segmentation into one-second audio segments for a dataset of 112 participants, comprising both healthy and ill instances. Mel-spectrograms for sequential models and spectrograms (STFT/CQT) for CNN-based image models were used as complementing feature representations. Four architectures—MobileNetV2, ResNet50, a hybrid CNN+BiLSTM, and a baseline CNN—were assessed. Adam or RMSProp were used to optimize the models after they had been trained and verified with augmented data. Accuracy, precision, recall, F1-score, and confusion matrices were among the evaluation criteria. According to the results, the CNN model outperformed CNN+BiLSTM (95.8%), ResNet50 (89.5%), and MobileNet (87.5%) in terms of total accuracy, achieving 97.9%. CNN consistently performed well across almost all disease categories, according to class-wise accuracy analysis. These results show that lung sound analysis using deep learning offers a viable, non-invasive, and precise diagnostic tool that can supplement or even replace conventional techniques in clinical practice.