Impact of prediction delay and machine learning model complexity on the performance of near-real-time CW radar heartbeat detection

Contactless vital-sign measurement has gained increasing attention across a variety of sectors, including medical and automotive applications as well as sleep monitoring. Radar-based systems enable the detection of heartbeat-induced microvibrations on the skin surface. This work investigates two key aspects of machine-learning-based near-real-Time heartbeat detection in radar signals: The influence of a prediction delay relative to an electrocardiography (ECG) reference signal and the impact of model complexity. For a systematic analysis, 153 gated recurrent unit (GRU) models have been trained and evaluated. At a delay of 200 ms, F1 scores above 92.7% and interbeat interval (IBI) mean absolute errors (MAEs) of less than 40 ms can be achieved. With increasing delay, the error reduces further, reaching an F1 score of 95.9% and IBI MAE of 23.4 ms at 1300-ms delay. In model complexity analysis, the preferrable combination of model depth, width, and prediction delay strongly depends on the metric of interest.

Subjects

Contactless vital-sign measurement

continuous-wave (CW) radar

gated recurrent unit (GRU)

heartbeat

machine learning

model complexity

prediction delay

DDC Class

600: Technology

Options

Impact of prediction delay and machine learning model complexity on the performance of near-real-time CW radar heartbeat detection