Radar can pick up the tiny chest movements associated with a heartbeat, but an acquired signal can still be too ambiguous to produce a useful estimate. Yuxuan Hu and colleagues propose HEAR, a model that assesses heartbeat readability before deciding whether to report a heart rate.
Their HEAR study tests a practical distinction: a sensor being pointed toward someone does not establish that this particular recording contains a readable heartbeat component. The authors evaluate a learned selection score on two public real-world datasets, rather than assuming that a favorable viewing angle guarantees a reliable measurement.
Why similar positions can yield different signals
Echoes from multiple points on the body combine within the same radar range bin. Small differences between the phases of those echoes can reinforce or suppress the heartbeat component. The paper derives this effect while holding motion projections fixed, explaining why macroscopic geometry alone leaves part of the problem unresolved.
The team builds a controllable multi-scatterer FMCW simulator to generate training examples. It combines respiratory and heartbeat waveforms with body models, varies scattering and observation conditions, and uses the known simulated heart rate to label whether the dominant heartbeat-band peak agrees with the target. This gives the model supervision about spectral readability without requiring a person to label every simulated window.
The simulation library includes 2,240 paired respiratory and heartbeat waveforms and 48 body models spanning 12 postures and four body types. These are controlled training conditions, not a count of real patients or a clinical validation cohort.
A score alongside the heart-rate estimate
HEAR is a compact dual-task Transformer that predicts an observability score and heart rate. Its input includes spectral magnitudes and frequencies relative to the respiration fundamental, giving the network context about respiratory harmonics that might resemble heartbeat peaks.
At inference, HEAR processes 25-second windows extracted with a 10-second hop. A threshold decides which estimates to retain. The learned score can also select measurements for existing estimators, so the quality assessment is not tied only to the model's own prediction head.
The authors train HEAR on simulated observations and evaluate transfer without real-data adaptation. The two public datasets cover 134 subjects and radar recordings at 60 and 120 GHz. This zero-shot result concerns those datasets and acquisition conditions; it does not establish equivalent performance across homes, devices or medical use cases.
Lower error comes with lower coverage
On the 120 GHz dataset, selection reduces the heart-rate head's mean absolute error from 17.9 BPM at full coverage to 1.6 BPM at 50% coverage. The second number describes retained measurements after half the windows are excluded. Reporting it without the coverage change would conceal a central part of the method.
The complete pipeline has a reported edge-device processing latency of 50.8 milliseconds. That computation time is separate from the 25-second observation window and should not be read as the time required to acquire a fresh measurement.
The study supports evaluating error and coverage together. A monitoring application would also need to decide what happens during rejected windows and whether missing estimates occur at important moments. The paper presents evidence for measurement selection, not approval to use the system as a clinical monitor.
Comments