Google Research has unveiled a system that uses the front-facing camera on a smartphone to estimate heart rate and resting heart rate during ordinary phone use, without requiring a deliberate scan or a wearable device. The work, published in Nature, aims to make heart monitoring more accessible by turning a device many people already carry into a health-sensing tool.
The research system, called PHRM, looks for brief face videos collected after a phone unlocks. It then applies deep learning to infer heart rate from subtle color changes in the face associated with blood flow. According to Google, the method reached a mean absolute percentage error below 10% compared with ECG-based measurements, which the company says meets industry accuracy standards across skin tones. For daily resting heart rate, the system posted an error below 5 beats per minute compared with a wearable tracker.
Google says the approach differs from earlier camera-based heart-rate methods because it is designed to work passively in daily life rather than in tightly controlled settings. The company used on-device software to process 8-second clips and combine estimates throughout the day into a resting heart rate reading. Confidence scores and Kalman filtering were used to smooth the results and improve reliability.
The training data behind the system was broad by research standards. Google says PHRM was developed using more than 350,000 video clips from nearly 700 consented participants in laboratory and real-world settings. The company also emphasized skin-tone representation, saying its dataset included at least 25% each from lighter and medium skin-tone groups and at least 33% from darker skin-tone participants, using the Monk Skin Tone scale.
That emphasis matters because remote photoplethysmography, the general technique behind camera-based pulse detection, has historically struggled to perform equally well across different skin tones. Google’s blog post notes that earlier studies often used smaller samples and underrepresented darker-skinned participants, which can make the pulse signal harder for cameras to detect.
To evaluate the system, Google trained PHRM on a set of laboratory recordings taken under varied lighting and activity conditions, then tested it on a separate group of participants. In that test, the company says the model was the only one among 15 leading published rPPG systems to achieve less than 10% error across all skin-tone groups.
Google also ran a free-living study in which 231 participants installed a custom app on their personal phones and used them normally for eight days while also wearing an ECG chest strap and a Fitbit tracker. The app recorded face videos after each unlock event. Participants reviewed the clips before uploading them, according to Google, to exclude sensitive content and other people.
On a held-out subset from that study, PHRM achieved an overall error rate of 6.09% for heart rate after confidence filtering. Google reported similar performance across light, medium and dark skin-tone groups, with the darkest group showing a higher error rate than the others but still staying below the company’s target threshold.
For resting heart rate, Google said the system produced estimates on 73.6% of participant-days among users with enough heart-rate readings to support the calculation. Its average error was 4.39 beats per minute versus the Fitbit reference.
Along with the paper, Google is releasing a large smartphone-video dataset and a pre-trained model for qualified researchers. The company says the work could help expand access to heart-health insights in places where wearables are less common.
Still, Google acknowledged limitations. The heart-rate success rate was lower for some skin-tone groups, especially the darkest group, and the company suggested that future improvements could include better camera exposure handling or extra sampling attempts. It also noted that outlier errors could come from motion and talking.
Even with those caveats, the research represents one of the most ambitious demonstrations yet of passive health monitoring from a consumer smartphone camera, and a sign that phones may eventually do more than display health data. They may also help collect it.