Contactless ECG prediction via FMCW radar by a multi-task conv-trans net

Abstract

The electrocardiogram (ECG) records a series of electrical signal sequences produced by the polarization and repolarization of the various structures of the heart. The pattern of the ECG diagram, especially the relationship between each wave or wave group, can convey much information. Moreover, this can be used in diagnosing myocardial ischemia, myocardial infarction, malignant arrhythmia, coronary heart disease, and other cardiovascular diseases. Currently, the primary way to obtain ECG is to acquire a series of electrical impulse signals in a heartbeat through electrodes attached to the patient's chest and limbs. This dissertation describes an ECG monitoring system based on FMCW radar, which does not require any contact with the patient's body surface, thus benefiting the monitoring of cardiac activity in groups such as skin disease patients or burn patients. The reconstruction of ECG signals is based on the features extracted by the hybrid model of CNN and Transformer. The multi-task learning method is used to improve the system's stability. The experimental results show that the system accurately captures R and T waves, and the average absolute prediction errors at the time nodes are 8.67 ms and 10.35 ms, respectively. This result indicates that the system can reproduce ECG signal morphology without contact.