Novel AI-assisted respiratory sound analysis for clinical and home-based health monitoring

This thesis investigates the classification of respiratory anomalies using advanced audio feature extraction and deep learning techniques within the framework of Classification of Respiratory Audio Anomalies (CRAA). The study examines how different time–frequency representations and deep learning methods can enhance the detection and classification of abnormal respiratory sounds. By utilizing multiple spectrogram and scalogram representations, along with deep learning approaches such as joint learning models, ensemble systems, attention-based architectures, Inception–Residual networks, and an embedding-based alignment framework for cross-domain CRAA, the research demonstrates that respiratory anomalies can be detected with high accuracy and robustness, even in challenging clinical and home-based environments. Overall, the findings highlight the potential of audio-based deep learning systems as non-invasive, scalable, and cost-effective tools for the early detection and monitoring of respiratory diseases. The results support the development of practical respiratory screening solutions that can be integrated into telehealth platforms and mobile health applications, particularly in resource-limited settings. Ultimately, this work aims to enhance diagnostic support and help reduce the global burden of respiratory diseases, especially in regions with limited access to skilled healthcare professionals.