A novel time representation input based on deep learning for ECG classification

Abstract

Electrocardiogram (ECG) is an important tool used to analyze abnormal heart activity and assess heart health, especially in remote cardiac health monitoring. Although deep learning has achieved significant results in automatic ECG classification, how to combine the characteristics of ECG physiological signals to construct inputs or features with differentiation is still a key point of classification. To this end, a novel representation input method with temporal characteristics was proposed in this paper. At first, the temporal characteristic of ECG signals was extracted and transformed into a time representation input with the original input. Subsequently, the deep learning network combining Convolutional Neural Network and Long Short-Term Memory was employed for feature extraction. Simultaneous attention mechanism was used to focus on feature differences. The proposed method was validated in the classification of five classes of heartbeats (Normal heartbeat, Left bundle branch block heartbeat, Right bundle branch block heartbeat, Atrial Premature Contraction, Premature ventricular contraction), achieving a higher average accuracy, precision, sensitivity, and specificity of 98.95%, 97.07%, 96.54%, and 99.33% respectively in the MIT-BIH arrhythmia database. The results show that our method is able to combine the periodic characteristics of ECG to construct a better temporal representation input than traditional feature fusion. This method can provide a new way to classify similar physiological signals with periodic characteristics.