CADDet: A Machine Learning Framework for Coronary Artery Disease Prediction Using Heart Sound Signals

Abstract

Coronary artery disease (CAD) continues at the forefront of mortality sources across the globe. Its early detection using heart sound signals seems promising for integration into Wearable Body Area Networks (WBANs). On the other hand, WBAN-based CAD detection systems face challenges such as noise, motion artefacts, and poor signal quality, which in turn reduce diagnostic performance. Literature surveys indicate that most current models struggle due to insufficient feature extraction, fragile classification, and poor generalisation, leading to the outlined dilemma. We propose a robust classification algorithm that combines MFCC feature extraction with Random Forests to achieve high detection accuracy, addressing these problems and filling the research gap. For our research, we used the Heartbeat Sounds datasets from Kaggle, which encompass recordings from both clinical and non-clinical environments (Sets A and B). We derived 13 MFCC features per recording and employed an 80-20 stratified train-test split to balance the evaluation. The Random Forest classifier, powered by 100 decision trees, has achieved astonishing effectiveness, with 95% overall accuracy, 0.97 F1 Score for healthy cases, and 0.86 F1 Score for pathological cases. Our results exceed those of five recent baseline papers by a wide margin in precision, recall, and overall classification accuracy. Thus, they support the validity of the method we proposed for CAD detection using real heart sound data.