Multi-task Learning for Lung Sound and Lung Disease Classification

Recent advances in deep learning techniques have significantly increased the accuracy and efficacy of medical diagnosis. In this work, we propose a novel multitask learning (MTL) approach for concurrently classifying lung sounds and diseases. Our approach integrates MTL with four distinct deep learning architectures: 2D CNN, ResNet50, MobileNet, and DenseNet, to extract relevant features from lung sound recordings. The effectiveness of our proposed method is evaluated in this study using the ICBHI 2017 Respiratory Sound Database. The MTL for the MobileNet model outperformed the other models under consideration, achieving 74% accuracy for lung sound analysis and 91% accuracy for lung disease classification. The experiment's findings show how well our method works for simultaneously categorizing lung noises and lung illnesses. This study also computes the risk level for Chronic Obstructive Pulmonary Disease using the patient demographics from the database. Three machine learning algorithms as Random Forest classifiers, SVM, and logistic regressionwere used for this calculation. The Random Forest classifier achieved the greatest accuracy of 92% among three machine learning techniques.This activity significantly lessens the doctor's workload by assisting in both pathology diagnosis and good patient communication regarding potential causes or consequences.