Contactless ECG via Millimeter-Wave Radar: How It Rebuilds Your Heart’s Electrical Signal

Basel, Switzerland, Source :

Imagine a monitor that watches your heart’s electrical rhythm without a single sticky electrode. That is exactly where millimeter-wave radar research is heading. A recent study shows how radar can reconstruct the ECG waveform from tiny movements of the chest, turning an ordinary radio reflection into a meaningful cardiac signal.

Wouldn’t it be better if long term heart monitoring felt like nothing at all? Respiration and heart rate are powerful indicators of health. If someone can track them continuously, doctors can spot subtle changes earlier, support diagnoses, and personalize care. Standard ECG electrodes and respiratory belts do this well, and they also rely on skin contact, cables, and careful placement. That makes them less comfortable and harder to use for long stretches. Radar sensing changes the equation. It uses reflected continuous waves to detect chest wall motion from a short distance, all without capturing a single image or sound. Privacy remains intact.

A deeper challenge stands between radar and ECG. An ECG records electrical activity. Radar records mechanical motion. A heartbeat causes an electrical signal, and then the chest moves slightly after that activation. The relationship between those two events is complex and nonlinear. Simply measuring the chest movement is not enough. To recover the ECG shape, the system has to learn exactly how tiny vibrations correspond to electrical changes.

Earlier approaches used wavelets, autocorrelation, and physiological models to estimate rates. More recent deep learning models tried to generate ECG waveforms from radar directly. Architectures like CNN-LSTM, autoencoders, and attention based systems showed promise. They also ran into a practical problem. Radar and ECG devices often start on separate clocks, sample at different speeds, and begin recording at different times. If the signals aren’t aligned, supervised learning loses accuracy. Radar also picks up clutter, posture changes, and motion artifacts that complicate the mapping.

The new study builds a complete pipeline that handles these issues head on. The researchers used a 77–81 GHz FMCW radar with one transmit and four receive antennas. A BIOPAC MP150 system recorded reference respiration and three-lead ECG at 2000 Hz. Volunteers sat about half a meter from the radar and stayed still during 120 second sessions. The result was 35 valid recordings from four subjects, which gave them more than one hour of paired radar and ECG data.

Preprocessing started with a fast Fourier transform along the radar’s fast time axis to produce range profiles. The strongest range bin represented the chest wall. The phase of the complex signal at that bin was unwrapped to recover continuous displacement. Before any ECG reconstruction could happen, the radar and BIOPAC clocks had to be synchronized. The team extracted the respiratory component from each radar channel with a 0.1 to 0.6 Hz band-pass filter. Then they used normalized cross-correlation across a delay window of three seconds. That search selected the best channel and estimated the time offset. This step may sound technical. It is the quiet hero of the entire process.

For the ECG model, the radar chest signal was band-pass filtered from 8 to 30 Hz to emphasise cardiac-related vibrations. The reference ECG was resampled to 250 Hz, filtered with a 50 Hz notch and a 0.5 to 40 Hz band-pass. Both signals were cut into 2048 sample windows with 50 percent overlap. That created 1015 paired samples from the self-built dataset. A random split assigned 70 percent to training, 15 percent to validation, and 15 percent to testing.

The reconstruction network combines a convolutional encoder-decoder with a bidirectional LSTM. The encoder compresses the radar signal into compact features at multiple scales. The decoder restores the time dimension. Then the BiLSTM layers read those features in both directions, so the model understands the context before and after each heartbeat. Dense layers turn every 64 dimensional feature into a single ECG amplitude. The network uses a joint time-frequency loss that combines mean squared error in the time domain with an STFT based spectral loss. This keeps the reconstructed ECG sharp in both amplitude and frequency, reducing the oversmoothing that can hide important waveform details.

The results make the case for radar-based ECG. On the self-built dataset, the reconstructed ECG reached a mean absolute error of 0.0555, a correlation coefficient of 0.5631, and an RRMSE of 0.3162. The average heart rate error was just one beat per minute, and the R-R interval error was around 10 milliseconds. When tested on a public dataset with resting, apnea, and Valsalva scenarios, the model achieved correlation coefficients of 0.8923, 0.6919, and 0.8387. Those numbers are noticeably higher than the published MultiRes-LinkNet reference values under the same conditions. The MAE values were also lower. This goes beyond an algorithm exercise. It shows that mechanical chest motion carries enough heartbeat timing information to reconstruct a clinically relevant electrical rhythm.

Detecting R peaks from the reconstructed ECG allowed reliable estimates of average heart rate and mean R-R interval. Those parameters are useful in rhythm monitoring, especially for people who need long term cardiac surveillance. The researchers acknowledge that some waveform details still vary from sample to sample, and the current framework operates offline. Motion and posture changes remain areas for future exploration. Future work will bring more participants, subject-disjoint evaluation, better local waveform fidelity, and motion-robust strategies.

There is something quietly inspiring here. A radar module that looks like a small box can detect the heart’s electrical signature by observing the body’s surface motion. This type of sensing fits naturally into smart healthcare. It could support sleep monitoring, sedentary health assessment, and home-based cardiac watch without asking the person to wear anything. For anyone who values ease and comfort, that is a future worth watching.