Contactless ECG from Tiny Chest Motion

System overview: from heart activity to contactless ECG

Figure 1. Use this as the story map: one heart activity produces both mechanical chest motion and electrical ECG; radar measures motion and AI translates it into an ECG-like signal.

Radar evidence: chest motion follows the cardiac cycle

Figure 2. This figure demonstrates the key evidence: the radar micro-motion waveform repeats with the ECG rhythm, even though radar is measuring motion rather than electricity.

How the radar signal is cleaned

Figure 3. Raw radar reflections are separated in 3D, heartbeat micro-motions are amplified, cardiac-related voxels are focused, and nearby signals are spatially filtered.

Finding repeatable heartbeat motion patterns

Figure 4. When radar micro-motions are aligned by ECG R-peaks, the grey traces share a similar repeated pattern; the red line is the average trend.

Figure 5. The algorithm looks for repeated radar micro-motion patterns and checks whether they align with cardiac cycles.

AI translation: mechanical domain → electrical domain

Figure 6. The encoder extracts spatial and temporal features from radar motion; the decoder reconstructs the ECG waveform.

Training detail: make small ECG features count

Figure 7. The µ-law transformation makes small ECG features more visible during training so the model does not only learn the large R peak.

Experimental setup

Figure 8. The radar is placed above the chest about 0.4–0.5 m away while wired ECG is recorded at the same time as ground truth.

Overall performance: good when still, weak under movement

Figure 9. The system performs best when the subject is still; body movement can corrupt radar measurements and produce failed ECG reconstruction.

Timing accuracy: strongest for R-peaks

Figure 10. The system compares Q, R, S, and T timing against wired ECG. R-peaks are reconstructed most accurately.

Morphology accuracy: reconstructed ECG vs wired ECG

Figure 11. The orange contactless ECG is compared with the blue ground-truth ECG. The examples show high shape similarity.

Does it generalize across people?

Figure 12. This checks whether the model generalizes to unseen people, showing timing error, RMS error, and correlation across 35 participants.

Rhythm-monitoring potential

Figure 14. The radar-reconstructed ECG tracks bradycardia, tachycardia, and irregular R-R intervals, suggesting rhythm-monitoring potential.

Daily-life limits: interference and distance

Figure 15. Performance is affected by nearby movement and sensing distance. The results show the best result around 0.5 m in a clean environment.