Artificial intelligence for accurate classification of respiratory abnormality levels using image-based features and interpretable insights

Abstract

Accurate classification of respiratory abnormality levels is crucial for early detection and diagnosis of respiratory diseases, making it a pivotal area in the field of medical diagnostics. This study proposes a novel artificial intelligence approach for accurate classification of respiratory abnormality levels. By transforming respiratory sound time-series data into image representations using recurrent plot, Markov transition field, and Gramian angular field, we capture intricate temporal patterns and spatial relationships. A deep neural network autonomously extracts discriminative features from these representations, subsequently integrated into machine learning classifiers. Leveraging the International Conference on Biomedical and Health Informatics (ICBHI) database, our methodology achieves remarkable classification accuracy of 100% for both binary and four-class scenarios, accurately distinguishing normal from abnormal sounds, and discriminating between crackles, wheezes, and their combinations. The SHapley Additive exPlanations (SHAP) method enhances interpretability, providing insights into feature importance and decision-making processes. This interpretable and high-performing approach offers significant promise for enhancing the accuracy and reliability of respiratory disorder diagnosis and treatment planning in clinical settings, potentially improving patient outcomes and healthcare efficiency.