Utilizing Deep Learning Algorithms for the Prompt Identification of Chronic Obstructive Pulmonary Disease

This study presents a Deep Learning (DL)-based approach for the early detection of Chronic Obstructive Pulmonary Disease (COPD) using a novel dual-branch Convolutional Neural Network (CNN) architecture.DL techniques are leveraged to recognize complex, early-stage patterns of the disease that may be overlooked by conventional medical assessments or traditional machine learning models, which are prone to misclassifying COPD as other lung conditions. To ensure robust model training, a pre-filtered dataset of lung sound recordings was used. These recordings, each 20 s in duration, were cleaned, standardized, and converted into two-dimensional representations using Mel spectrograms and Mel Frequency Cepstral Coefficients (MFCCs). These image-like features served as the input for the CNN model, enhancing its ability to distinguish COPD-specific acoustic patterns.To address the issue of class imbalance in the dataset, two data augmentation techniques, pitch-shifted noise injection and time-frequency masking, were applied, contributing to improved model generalization. The proposed CNN model achieved promising results, with a precision of 97.75%, an accuracy of 96.0%, a sensitivity of 97.96%, and an F1-score of 96.97% during validation. These performance metrics outperform those obtained from widely used CNN architectures, such as InceptionV3 and ResNet, highlighting the effectiveness of the proposed model. Overall, the proposed approach demonstrates significant potential as a reliable diagnostic support tool for early COPD detection.