Explainable AI for Respiratory Disease Detection: Leveraging Deep Learning on Patient Audio Data

Respiratory diseases affect millions of people around the world, making it necessary for reliable and interpretable diagnosis. Lung sound analysis is a non- invasive and cost-effective approach for detecting respiratory abnormalities such as wheezes and crackles, which are critical indicators of respiratory conditions like Chronic Obstructive Pulmonary Disease (COPD). This study uses machine learn- ing techniques to detect crackles and wheezes from lung sounds automatically. Leveraging the respiratory sound database, 13 Mel-Frequency Cepstral Coeffi- cients (MFCCs) were extracted from the audio recordings to classify respiratory abnormalities. While deep learning models achieve high accuracy, their black-box nature limits transparency. This study proposes an explainable AI (XAI) solu- tion for respiratory disease classification using audio signals. This study ensures interpretability by identifying critical features influencing predictions by train- ing models on publicly available datasets and incorporating Local Interpretable Model Agnostic Explanations (LIME). Explainability analysis revealed criti- cal features influencing predictions, ensuring model transparency. This research advances the development of trustworthy AI-driven diagnostic tools, contributing to enhanced transparency in healthcare.