SOUNDNet: Leveraging Deep Learning for the Severity Classification of Chronic Obstructive Pulmonary Disease Based on Lung Sound Analysis

Chronic obstructive pulmonary disease, commonly known as COPD, is a significant and prevalent health concern impacting individuals globally. The suggested study presents a unique framework based on lung sound recordings that use deep learning to analyze and classify COPD severity. Current methods to classify COPD severity, such as subjective analysis, expert input, spirometry, and lung function tests, are invasive and time-consuming. To address this, a novel framework using statistical analysis of lung sound samples is proposed for early detection and severity classification of COPD. The proposed framework extracts features from the Open Respiratory Sound Database and Respiratory@TR dataset, including Chroma Short-Term Fourier Transform (CSTFT), Mel-Frequency Cepstral Coefficients (MFCCs), Mel-Spectrogram (mSpec), Melspectrogram-snippet, and a combined format of these all features. Both two-class and multiclass classification approaches were proposed. For the two-class classification, a deep Convolutional Neural Network (CNN) architecture was employed, while a pretrained YAMNet model was utilized for the multiclass classification. We achieved 95.76% and 93% accuracy for two-class and multiclass classification respectively. Lastly, We compared the framework’s performance against current methods and traditional benchmarks for early COPD diagnosis.