A Modified Convolutional Neural Network for ECG Beat Classification

Abstract

Most convolutional neural networks (CNNs) used to classify electrocardiogram (ECG) beats tend to focus only on the beat, ignoring its relationships with its surrounding beats. This study aimed to propose a hybrid convolutional neural network (HCNN) structure, which classified ECG beats based on the beat's morphology and relationship such as RR intervals. The difference between the HCNN and the traditional CNN lies in the fact that the relationship can be added to any layer in the former. The HCNN was fed with RR intervals at 3 different positions, trained using data from 2170 patients. It was then evaluated with labeled clinical data from 2102 patients to classify ECG beats into premature ventricular contraction beat, atrial premature contraction beat (APC), left bundle branch block beat, right bundle branch block beat, and normal sinus beat. The results showed that the performance of the proposed HCNN method (with an average score of 86.61% on 12 leads) was 3.31% higher than that of the traditional CNN (83.30%) on the test set. In particular, the APC improved most significantly from 57.67% to 76.92% in terms of sensitivity and from 58.80% to 78.46% in terms of the positive predictive value in lead V1.