A deep learning method of noncontact time domain heartbeat signal extraction from millimeter-wave radar data

Abstract Heart rate is an important physiological signal, and monitoring heart rate is crucial for medical treatment, rescue and health management. Traditional heart rate monitoring methods require the use of contact devices, which can cause inconvenience and discomfort during long-term monitoring. Non-contact vital sign monitoring based on radar technology has emerged as a promising alternative and triggered widespread research interest. Achieving precise separation of heartbeat signals and high-precision monitoring of heart rate from radar signals remains challenging due to the weak amplitude of heartbeat signals, interference from breathing, body movements, and various types of noise. This paper proposes a new method that uses filtered pulse waves to approximate the heartbeat component of the radar reconstructed chest displacement time sequences, and applies a Conv-TasNet model for heartbeat separation. The proposed method was evaluated usingthe synchronized electrocardiogram signal from the public data set as the gold standard. We report that the root mean square error of the inter-beat interval is 8.32 ms on average, which outperforms state-of-the-art techniques.