A novel method for automated congestive heart failure and coronary artery disease recognition using THC-Net

Abstract

Coronary artery disease (CAD) and congestive heart failure (CHF) lead to many deaths worldwide. Generally, an electrocardiogram (ECG) is employed as the diagnostic tool for CAD/CHF recognition. However, since ECG changes are sometimes subtle, visually distinguishing long-term ECG abnormalities is time consuming and laborious. To address these issues, we proposed a novel two-channel hybrid convolutional network (THC-Net) for automatic ECG recognition. THC-Net contains a canonical correlation analysis (CCA)-principal component analysis (PCA) convolutional network, an independent component analysis (ICA)-PCA convolutional network, and a Dempster-Shafer (D-S) theory-based linear support vector machine (SVM). The CCA-PCA and ICA-PCA convolutional networks are developed to extract deep features containing the lead-correlation and lead-specific information, respectively, from ECGs. Compared to common convolutional neural networks (CNNs), their kernels can be directly extracted by CCA, ICA, and PCA with a faster training time. Then, the D-S theory-based linear SVM, which can process multi-channel uncertainty information, is employed as the classification model. In this work, an accuracy of 95.54% was obtained for classifying normal, CHF and CAD patients based on leave-one-out cross-validation. Additionally, experiments on multi-level noisy and imbalanced data yielded remarkable results. Hence, the proposed method has the potential to diagnose CAD and CHF in clinical settings.