A knowledge-based deep learning method for ECG signal delineation

Abstract

The delineation of electrocardiograms (ECG) is a crucial step designed to extract signal characteristics and assist cardiologists in diagnosing certain diseases. It refers to the delineation of both the onset and offset of the P wave, the QRS complex, and the T wave. However, to date previous studies have neither investigated the role of medical knowledge nor investigated the interference of signal noise. In this paper, we propose a knowledge-based deep learning method for ECG signal delineation, which adds domain knowledge and individual feature knowledge to improve delineation performance. The proposed method is novel in three distinct ways. First, the method aligns the encoded knowledge with ECG signals with reference to the R wave peak position. Second, the method incorporates domain knowledge to delineate ECG signals under the encoder–decoder framework. Third, the method introduces individual feature knowledge to adjust the delineation results adaptively. An evaluation conducted on the QT database demonstrates that the proposed method can obtain, on average, high performance with sensitivity of 99.62% and positive predictivity of 99.81%. The evaluation also shows that the method can obtain a sensitivity and a positive predictivity above 90% in most noisy cases.