Respiratory Disorder Detection using Deep Learning

Abstract

This thesis investigates the use of deep learning techniques to detect respiratory disorders using audio recordings, aiming to improve early diagnosis. Respiratory disorders, such as Asthma, Chronic Obstructive Pulmonary Disease, pneumonia, Lung fibrosis, Pleural effusion, Bronchiectasis and heart failure, often go undiagnosed due to the lack of accessible diagnostic tools, leading to severe health complications. This study utilizes a dataset of 336 audio recordings categorized into eight classes. Techniques such as data segmentation, augmentation, and feature extraction using Mel Frequency Cepstral Coefficients and spectrograms are employed. Several deep learning models, including Convolutional neural network, long short-term memory, recurrent neural network, Time-delay neural network, Autoencoder, Attention and wavenet even combined deep learning models are applied and compared based on their performance metrics. These deep learning models demonstrated high accuracy in classifying respiratory disorders, with Convolutional neural network, Time-delay neural network and wavenet models achieving above 99% in terms of accuracy, F1 score, recall, and precision.