Artificial Intelligence and Signal Analysis for COPD Classification: Detecting and Understanding of Respiratory Disease Severities.

Abstract

Chronic Obstructive Pulmonary Disease (COPD) is a pressing global health issue that demands precise and prompt diagnostic methods. This thesis marks a significant advancement in the field, as it explores the utilisation of artificial intelligence (AI) and advanced signal processing techniques to enhance the diagnostic potential of pulmonary auscultation audio for automatic COPD identification and severity assessment. The research comprises three main parts: 1) developing algorithms for compression and reconstruction of auscultation signals whilst the diagnostic integrity is preserved, 2) detecting COPD from healthy individuals and other respiratory conditions using processed audio signals, and 3) creating an AI model to accurately classify distinct COPD severities based on auscultatory features from a diverse dataset. A novel compression algorithm was developed, reducing audio file sizes by up to 97% whilst enabling high-fidelity reconstruction without compromising diagnostic information content. Using Random Forest Ensemble, a machine learning model achieved 70-80% accuracy in distinguishing COPD from non-COPD cases using the compressed framework. Furthermore, an AI classification system using Support Vector classifiers demonstrated 75% accuracy in the identification of COPD severity using projection methods and machine learning algorithms, a significant improvement over a non- projection approach. Projection methods, which transform data by learning from a breathing dataset to apply to the COPD severity dataset, were crucial in reducing noise and preserving diagnostic information in a form of transfer learning. The study utilised two datasets, one comprising 124 subjects and another with 42 individuals, providing a diverse range of auscultatory data for analysis. The findings validate the hypotheses that advanced signal processing can mitigate noise impacts whilst AI models trained on diverse data can discern auscultatory patterns indicative of COPD presence and severity. This research contributes to the field by developing a compression framework for pulmonary audio, an AI model for COPD identification and severity assessment, and insights into the strategic handling of heart sounds in compressed sensing and reconstruction. The automatic identification of COPD severity has significant implications for personalised care, enabling tailored treatment plans and more effective management of the condition. Moreover, the potential for utilisation in remote and community care settings brings advanced medical knowledge closer to patients. This approach could significantly reduce diagnostic delays and improve early intervention strategies, particularly in areas with limited access to specialist care. The promising findings underscore the need for larger, more diverse datasets to aid advanced feature engineering techniques in accurately capturing COPD severity nuances. The integration of AI into clinical practice represents a significant development with the potential to enhance diagnostic accuracy, remote care, and personalised patient care.