Contactless Vital Signs from mmWave Echo Phase

A radar can measure how fast someone is breathing and how fast their heart is beating without touching them and without a camera. This release presents a commercial 79 GHz radar, a processing chain that turns its echoes into breathing and heart rate, and a set of measurements that shows where the method works and where it should stay silent.

01How a radar sees a breath

The radar sends short chirps, signals whose frequency rises steadily, and listens to the echoes. The delay of each echo tells it how far away the reflecting surface is, so the person can be separated from the walls and furniture around them.

Distance alone is far too coarse to see a breath: the radar can only tell apart surfaces that are about 4 cm apart. The fine detail is in the phase of the echo, the position of the returning wave within its cycle. At 79 GHz one wavelength is 3.8 mm, and because the signal travels to the chest and back, a chest movement of 1 mm shifts the echo by about half a cycle. Breathing moves the chest by a few millimetres; the heartbeat adds a much smaller movement on top.

02The experiment

The radar, a Texas Instruments IWR1443BOOST evaluation board, was mounted on a tripod at chest height, between 0.55 and 1 m in front of a seated volunteer. Recordings were made in an anechoic chamber, a room lined with absorbers that suppress reflections, and in an ordinary office. Each one lasted 60 or 120 seconds while the person breathed normally.

Two independent references were taken for comparison, and neither is used by the algorithm. For heart rate, a finger pulse oximeter whose screen was filmed and read every 5 seconds. For breathing rate, a manual count of breaths over the whole recording.

A volunteer standing in a blue anechoic chamber in front of a radar on a tripod, with a second radar and the acquisition computer nearby — A standing measurement in the anechoic chamber (left) and the finger pulse oximeter used as the heart-rate reference (right).

A seated volunteer wearing a blue finger pulse oximeter that shows a pulse of 66 — A standing measurement in the anechoic chamber (left) and the finger pulse oximeter used as the heart-rate reference (right).

03From echo to vital signs

The processing has four steps. Each is shown here on one 120 s recording of a seated volunteer at about 1 m, called M8.

Find the person. The radar first sorts its echoes by distance. Averaged over the recording, the strongest echo comes from 1.10 m: the chest. Only that region is followed from here on.

Step 1Locating the chest

Echo strength at each distance from the radar. The shaded band is the region selected as the chest.

Recover the motion. The phase of the chest echo is converted into millimetres of movement towards or away from the radar. The radar has four receiving antennas, and each gives its own view of the same movement. Sudden body movements cause phase jumps far larger than any breath; those moments are marked invalid and left as gaps rather than filled in.

Step 2Chest motion recovered from the echo

Breathing appears as a regular oscillation of a few millimetres, seen by two of the receivers. Switch to Standing for the same volunteer on their feet: the body sways and the breathing pattern is much harder to isolate.

Count the breaths. Breathing rate is estimated on 30 s windows that move forward 2 s at a time. Instead of counting peaks, the chain compares the signal with shifted copies of itself: when the shift equals one breath, the copy lines up with the original. In this window the best match comes at 3.88 s, which is 15.5 breaths per minute.

Step 3Breathing rate from self-similarity

The 30 s window highlighted in step 2, and how closely it matches itself when shifted in time. The highest similarity marks the length of one breath.

Find the heartbeat. This is the hard part. The heartbeat moves the chest far less than breathing does, and breathing is not a perfect wave: it leaves traces at whole multiples of its own rate, called harmonics, which can land exactly where a heartbeat would be. In this window the fourth breathing harmonic falls at 61.9 per minute, right next to the heartbeat at 63.3.

So for each candidate peak the chain asks a precise question: does adding a heartbeat at this rate explain the signal better than breathing and its harmonics alone? The comparison uses the Bayesian information criterion, which rewards a better fit but penalises every extra parameter. A candidate that does not improve the explanation is discarded.

Step 4Heartbeat among the breathing harmonics

Spectrum of the same window between 50 and 150 per minute. Dashed lines mark multiples of the breathing rate; the accepted heartbeat is the peak that the breathing pattern alone cannot explain.

Agree, or say nothing. A window only produces a value when at least two of the four receivers agree to within 2 per minute and the window contains at least 15 s of clean signal. Otherwise it reports nothing. The result for a recording is the median of its accepted windows.

04What the measurements show

Eight seated recordings had both references available, at distances from 0.55 to 1 m.

Breathing rate1.19

Mean error in breaths per minute. All 8 recordings produced an estimate.

Heart rate2.43

Mean error in beats per minute, over the 7 recordings that produced an estimate.

Eight recordingsRadar against reference

Select a recording for its distance, duration and values. In M4 no heart-rate window passed the checks, so the chain returned no value instead of a guess. M7 has the largest errors in both rates.

Window by window, M8 stays close to both references. All 46 breathing windows and 41 of 46 heart windows were accepted, giving 14.2 breaths per minute against a reference of 14.5, and 64.3 beats per minute against 64.5.

Recording M8Every accepted window

Each dot is one accepted 30 s window. Gaps in the heart rate are windows the checks rejected. The references are single averages over the whole recording.

Standing is harder. In a 120 s standing recording the breathing rate still matched its reference, 9.6 against 9.5 breaths per minute, but only 4 of 46 windows were accepted and no heart-rate window was. A close final number is not enough on its own: how much of the recording supports it matters just as much.

One recording pointed the radar at the neck instead of the chest. The heart rate came out at 53.3 beats per minute against a reference of 55.3.

05When breathing stops

In two short tests the volunteer held their breath partway through a recording, or started breathing again after holding it. The change is immediate in the recovered chest motion.

Breath-hold testsBreathing on and off

In the first test, regular breathing gives way to an almost flat line at around 20 s. In the second, the flat line breaks into breathing after about 8 s.

These were exploratory tests rather than an apnea detector, but they show that the signal the chain works with responds at once to a pause in breathing.

06Towards contactless blood pressure estimation

Each heartbeat sends a pressure wave along the arteries. The time it takes to travel from the chest to the neck, the pulse transit time, shortens as blood pressure rises, which is why it is studied as a route to blood pressure monitoring without a cuff. Measuring it without contact means seeing the same heartbeat in two places at once, with a clock precise enough to time the difference.

As a first step, the setup was extended to two radars observing the same person at the same time, one aimed at the neck and one at the chest. The radars run independently, and every piece of data from both is time-stamped by the same receiving computer, which places the two recordings on one common clock.

A volunteer standing in an anechoic chamber between two radars on tripods, one at chest height and one higher, aimed at the neck — Two radars aimed at the chest (left) and the neck (right) of the same volunteer.

When the volunteer moves, both radars register the movement at the same moment on that clock, a check that the two recordings are aligned.

This is the experimental base, not yet a blood pressure measurement. The next steps are tighter timing between the two sensors, then the pulse transit time itself, and only then its relation to blood pressure.

07Scope and next steps

The results come from a small set of recordings, mostly of one seated person at rest, processed offline. They characterise the method; they are not a clinical validation, and the system is not a medical device. The references also have limits: the pulse oximeter measures the pulse at the finger rather than chest motion, and the manual count gives one average per recording.

The next steps are more volunteers, references synchronised with the radar, a more robust heartbeat stage and real-time operation.