Deep Learning for Respiratory Sound-Based Disease Classification

Conditions of the lungs, including pneumonia, COPD, asthma are the main sources of mortality, affecting a significant portion of the population and requiring early and accurate diagnosis. Traditional diagnostic procedures that are expensive and time-consuming include spirometry, blood tests, and imaging techniques such as chest radiography. This research is about deep learning-based methods for differentiating respiratory diseases using the Respiratory Sound database. Feature Extraction used are: Chroma STFT, Mel spectrogram, and MFCCs, thus expanding the training set and test set size. These precomputed features fed as input to multi-input model. A multi-input model of CNN, a combination of Resnet and Attention mechanisms is proposed. The model uses different CNN branches to precomputed features such as MFCCs, Chroma STFT, Mel Spectrogram separately. It then concatenates to obtain the feature vector from three feature models for final classification. Tested on a set of respiratory sound recordings, the model achieves an accuracy of 95% in categorizing six lung disorders. Results indicate models such as RESNET and Attention mechanisms can be used to diagnose respiratory diseases in an efficient and noninvasive manner. This approach offers a strong foundation for incorporating into clinical applications to raise the precision of diagnosis and the effectiveness of healthcare.