Repository logo
 

Asthma and COPD Risk Prediction From Respiratory Sounds Using Deep Learning

Files

Thesis is embargoed until 4 July 2027(3.3 MB)

Date

2024

Authors

Supervisor

Al-Jumaily, Ahmed
Sabit, Hakilo

Item type

Thesis

Degree name

Doctor of Philosophy

Journal Title

Journal ISSN

Volume Title

Publisher

Auckland University of Technology

Abstract

In recent years, Deep Learning (DL) models have offered many promising improvements for the early detection of respiratory diseases like asthma and Chronic Obstructive Pulmonary Disease (COPD), using techniques such as audio analysis and imaging. Chronic respiratory diseases, as documented in global health statistics, have emerged as the third most prevalent cause of mortality worldwide. In the year 2019 alone, these diseases were responsible for an estimated 4 million deaths internationally. This substantial figure serves to illuminate the extensive and profound impact that respiratory diseases exert on the global health landscape. Statistical analyses indicate that a significant majority of chronic respiratory disease cases, exceeding two-thirds, are primarily attributed to either asthma or COPD. However, the high degree of diagnostic overlap between chronic pulmonary conditions of asthma and COPD, owing to their similar symptomatology, significantly contributes to diagnostic inaccuracies, leading to considerable mortality annually. To confront the pressing challenge presented by the high prevalence of respiratory diseases, particularly asthma and COPD, this study introduces an enhanced DL framework, employing a sophisticated multilayer Convolutional Neural Network (CNN) as the basis for a classifier model. This model is meticulously designed to identify and distinguish between asthma and COPD. Through the application of advanced CNN methodologies, the proposed model aims to augment the accuracy and efficiency in the diagnosis and differentiation of these two prevalent respiratory conditions, thereby contributing significantly to the field of medical diagnostics and treatment optimization. It entailed a comprehensive analysis involving the extraction of features and the development of a DL model aimed at differentiating between these two predominant respiratory diseases. The study utilized a sample of 200 respiratory sound recordings, of which 120 were allocated for training purposes and 80 for testing. The methodology encompassed extracting log energy (LE) and Mel energy from the Mel Frequency Cepstral Coefficients (MFCC) features of lung sounds as the foundational elements for classification. A variety of classifiers, including Support Vector Machine (SVM), Random Forest, k-Nearest Neighbor (KNN), and Multilayer Perceptron (MLP), were utilized. These classifiers primarily focused on discriminating between asthma and COPD based on lung sound features. Notably, the enhanced DL model achieved a significant classification accuracy of 86.25% when employing the SVM classifier. Moreover, the model demonstrates notable enhancements across key performance metrics, achieving a Sensitivity of 87.81%, Specificity of 55.31%, and an F1 Score of 77.51%. Tailored to precisely predict asthma and COPD, this model stands out for its capability to accurately identify these conditions. The adoption of such an advanced diagnostic tool in clinical settings can enhance the efficiency of medical practitioners, enabling them to make quicker, data-driven decisions.

Description

Keywords

Source

DOI

Publisher's version

Rights statement

Permanent link

http://hdl.handle.net/10292/17742

Collections

Doctoral Theses