Classification of cardiac electrical signals between patients with myocardial infarction and healthy controls by using time-frequency features and 3D convolutional neural networks

Electrocardiogram (ECG) signal classification plays an important role in myocardial infarction (MI) detection and screening. Despite that much progress has been made, the interpretation of ECG signals is still extremely time-consuming, and heavily relies expertise of clinical cardiologists. In this paper, an automated classification method is developed based on cardiac time-frequency features and 3D convolutional neural networks for MI detection. First, an ECG feature representation scheme based on time-frequency spectrograms without complicated signal segmentation and morphological analysis, is proposed to elaborate the dynamical characteristics underlying time-varying ECG signals. Second, a new 3D convolutional neural networks (C3D) is adopted for in-depth feature learning underlying the extracted time-frequency features. The proposed 3D deep network can take advantage of the encoded spatial characteristics extracted from convolutional neural network and the full use of cardiac characteristics underlying all twelve leads. For the goal of two-class classification (MI or HC), a classification accuracy of 94.20%, 96.20% and 97.32% are achieved on the public PTB database under two-fold, five-fold and ten-fold cross-validation, respectively.