Embedded Cardiorespiratory Monitoring Using a 77 GHz FMCW Radar: On-Chip Multi-Timescale Tiny 1D CNN and Whole-Node Power Evaluation

A 77-GHz radar can detect small chest movements induced by cardiac activity and respiration; however, typical systems transmit the acquired signals to an external computer for processing. This limits the development of autonomous and energy-efficient monitoring devices. This study presents a system in which radar signal processing, signal quality assessment, and heart rate (HR) and respiratory rate (RR) estimation are performed directly on the AWR1843 radar device. A Tiny 1D CNN used a 10-s observation window for HR and a 30-s window for RR, producing updated estimates every 1 s. The system was validated with 30 participants aged 18–72 years. The mean absolute error was 2.09 beats/min for HR and 1.05 breaths/min for RR. Compared with the implemented short-window FFT-based spectral estimator, the errors were reduced by 45.3% and 37.5%, respectively. INT8 quantization reduced the model weight size from 19.2 to 4.8 kB, the inference time from 4.6 to 1.9 ms, and the incremental energy per network execution from 2.7 to 0.9 mJ. However, this inference-level improvement did not produce a proportional reduction in whole-node power, which remained approximately 2.38 W because the continuously active radar front end and frame-level signal processing dominated the total power consumption. The system transmits only the estimated parameters and their associated reliability indicators. Consequently, the radar can operate as a standalone measurement node rather than continuously streaming raw data to an external computer.