Automated intra-patient and inter-patient coronary artery disease and congestive heart failure detection using EFAP-Net

Abstract

Coronary artery disease (CAD) and congestive heart failure (CHF) occur worldwide, putting patients at risk of death. Researchers have developed many automatic methods for CAD and CHF classification. However, most have neglected evaluating the performance of these methods in inter-patient experiments that can guarantee their generalization in practical applications. Furthermore, the applicability of these methods to noisy and extremely unbalanced data has also not been validated. To address these issues, we propose a novel CAD and CHF classification method based on ECG fragment alignment (EFA)-principal component analysis (PCA) convolutional network (EFAP-Net). EFA is a method for eliminating heart rate differences that can ensure the component consistency of heartbeats between individuals. To deeply mine discriminative features, a network containing two convolutional layers and an output layer is used to extract high-dimensional abstract features from heartbeats. In this network, we employed PCA with a certain noise robustness to extract the suitable convolutional kernels for each layer, providing a network that is quickly trained and yielding remarkable performance in processing unbalanced data. Finally, a linear support vector machine specializing in classifying high-dimensional features is adopted as the classifier. In the intra-patient experiment, 99.84%, 99.92%, and 99.80% accuracies were achieved on balanced datasets A (normal+CAD), B (normal+CHF), and C (normal+CAD+CHF), respectively. In the inter-patient experiment, 94.64%, 98.94% and 86.86% accuracies were achieved on datasets G (normal+CAD), H (normal+CHF), and I (normal+CAD+CHF), respectively. Additionally, multi-level noisy and unbalanced ECG data were classified well. Hence, this method can be implemented to diagnose CAD and CHF effectively.