Lung sound classification using YAMNet, neural network, and augmentation

Abstract

Globally, lung disease occupies a significant position as one of the main contributors to mortality rates. The characteristics of human respiratory sound signals can show a wide spectrum, ranging from normal patterns to indications of lung abnormalities. The proposed lung sound classification system is based on YAMNet as a pre-trained neural network model for medical audio recognition, which is then refined using artificial neural networks (ANN). This study presents the integration of multiple datasets and advanced pre-processing approaches. A total of 1,363 lung sound recordings from Kaggle, ICBHI, and Mendeley. This reflects the variety of clinical conditions, and differences in recording devices are combined. In order to increase the diversity of lung sound signal input, the pre-processing process is carried out through several stages, including adjusting the sampling frequency to 4 kHz, segmenting for 6 seconds, signal filtering with wavelet, min–max normalization, and data augmentation using window warping, jittering, cropping, and padding. A fold cross-validation scheme is employed to comprehensively evaluate the model's effectiveness. The evaluation results indicate that the model achieves an accuracy of 93.64%, a precision of 93.60%, a recall of 93.64%, and an F1-score of 93.52%, collectively reflecting outstanding classification performance. This work may incorporate deep learning technology into clinical practice, ultimately improving diagnosis accuracy and efficiency in the hospital setting.