Enhanced Coronary Artery Disease Classification Through Feature Engineering and One-Dimensional Convolutional Neural Network

Coronary artery disease (CAD) diagnosis remains a significant contributor to global mortality rates, highlighting the need for novel approaches. Existing CAD diagnostic tools rely on costly and complex biomarkers and scanners. In this paper, using only electrocardiogram (ECG) signals, we propose a novel learning-based model for CAD diagnosis. The proposed method works based on a one-dimensional convolutional neural network (1D-CNN), offering a cost-effective alternative for sophisticated cardiac health monitoring. Furthermore, we introduce the concept of feature engineering to improve the quality of the model training process and mitigate the challenge of ill-conditioned ECG data. Unlike existing approaches, which often overlook signal quality, our model applies a smart feature engineering, ensuring that only diagnostically reliable signals are used. This design improves robustness, generalisability, and suitability for real-world clinical settings. Utilising one of the most complex publicly available datasets, i.e., MIMIC III, sourced from Physionet, the performance of the proposed model, along with existing ones in classifying potential cases of CAD and non-CAD is investigated. Our findings confirm that the proposed model exhibits outstanding performance, highlighting the effectiveness of our integrated feature engineering approach with the CNN model.