Skip to main content
Back

Google's PHRM Turns Your Phone Camera Into a Fitbit-Accurate Heart Monitor

Google's PHRM system uses your phone's selfie camera to passively measure heart rate after face unlock, matching wearables across all skin tones.

Repo, Full Access
  • Google Research unveiled PHRM, a passive heart rate system using the front-facing smartphone camera after face unlock.
  • Achieves HR MAPE of 6.09% and daily RHR MAE of 4.39 bpm versus Fitbit Charge 6.
  • Only rPPG model among 15 tested to hit under 10% MAPE across all skin tones.
  • Trained on 350,000+ clips from ~700 participants, balanced via the Monk Skin Tone scale.
  • Uses 8-second clips processed on-device by temporal shift CNNs plus Kalman filter aggregation.
  • Dataset and PHRM-mini model released on GitHub for qualified researchers.

Google Research just published a Nature paper describing a system that turns your phone's front-facing camera into a passive heart monitor. Called PHRM (Passive Heart Rate Monitoring), it captures short clips of your face after routine face unlock events, runs a small neural network on-device, and aggregates the results into a daily resting heart rate estimate that matches the accuracy of a Fitbit, no wearable required.

The release includes a GitHub repository with a pre-trained PHRM-mini model and what Google describes as the largest, most diverse smartphone-video rPPG dataset publicly available for research. Access is restricted to qualified researchers with IRB approval and non-commercial intent.

The unlock-and-measure loop

PHRM leverages the front-facing camera to capture video of the user's face in the seconds after face unlock events. The signal it extracts is photoplethysmography, the same technique pulse oximeters use, which detects tiny color fluctuations in skin caused by each pulse of blood. The difference is that the camera does it remotely, with no physical contact.

PHRM pipeline diagram showing screen unlock through video stabilization, face crop, neural network, confidence gating, and Kalman filter aggregation

The on-device pipeline processes 8-second facial video clips and uses computationally-efficient temporal shift convolutional neural networks to predict HR along with a confidence score. Temporal shift networks are a trick borrowed from efficient video understanding: instead of running expensive 3D convolutions across time, they shift a portion of feature channels along the time axis inside an otherwise standard 2D CNN, which gives you temporal reasoning at roughly the cost of a single-frame model.

From there, predictions get filtered by their confidence scores and fed through a Kalman filter that smooths a day's worth of measurements into one resting heart rate (RHR) number. The app recorded 8-second video clips and ECG data immediately after each face unlock, capturing an average of 231 clips per day. That is a lot of opportunistic sampling.

The skin tone problem this actually solves

Camera-based PPG has been a research curiosity for two decades, but it has a well-documented failure mode. Previous studies vastly underrepresented people with dark skin, in whom melanin makes the PPG signal more challenging for cameras to detect. Only recently have researchers investigated rPPG model performance on dark-skinned study participants more thoroughly, finding significantly lower accuracy. Pulse oximeters have the same problem, which is why the FDA has drafted guidance requiring diverse skin tone validation.

Google built PHRM specifically to clear that bar. They developed PHRM using over 350,000 video clips from nearly 700 diverse consented research participants in both laboratory and real-world settings, and devoted more model training to the most challenging cases. They leveraged colorimetric methods and the Monk Skin Tone scale to ensure participants with light, medium, and dark skin each met sampling thresholds, with the darkest-skin group making up at least 33% of the data.

They also imposed a non-inferiority criterion during training: PHRM's MAPE for HR for each group must differ from that of the others by less than 5 percentage points. In other words, the model is not allowed to be substantially worse on dark skin than on light skin and still ship.

What the numbers look like

Two studies anchor the validation. A controlled lab study with 365 participants under varied lighting and activity, and a first-of-its-kind "free-living" study where 231 diverse study participants installed a custom data collection app on their personal phones and used them as normal for eight days while wearing an ECG chest strap and a Fitbit Charge 6 fitness tracker.

On the held-out free-living validation set:

  • Overall HR MAPE of 6.09% after confidence gating, with MAPEs of 5.04%, 5.12%, and 7.84% for skin tone Groups 1, 2, and 3 respectively.
  • Each group met the non-inferiority target, and PHRM outperformed 15 leading rPPG models by a wide margin, remaining the only model to achieve MAPE below 10% across all skin tones.
  • Bland-Altman analysis showed PHRM underestimated HR by only 0.64 bpm on average, with 95% limits of agreement between -11.3 and 10.3 bpm.
  • Daily RHR matched the Fitbit Charge 6 with an MAE of 4.39 bpm, below the pre-specified 5-bpm target.
Bar charts comparing MAPE across skin tone groups for PHRM versus competing rPPG methods, with and without confidence gating

One nuance worth flagging: for the darkest skin group, RHR accuracy only crossed the target threshold once the Kalman filter had a few days to converge. MAE for all groups decreased over time as the RHR algorithm's Kalman filter converged, with Group 3's MAE significantly below 5 bpm from day three onwards.

Where it falls short

Even after the dataset balancing work, the success rate of capturing a usable measurement was uneven. PHRM's HR measurement success rate was lower for Group 2 and lowest for Group 3, likely due to the difficulty of detecting the PPG signal in darker skin. The model is accurate when it does fire, but it fires less often on darker skin tones.

Talking, head motion, and bad timing also produced outlier errors. The team suggests future work could use accelerometer gating to wait for at-rest moments, better video stabilization, and adaptive camera exposure.

Why this matters beyond a cool demo

The framing here is access. Roughly five billion people own a smartphone with a front-facing camera. A much smaller subset owns a wearable, and adoption is lowest in exactly the populations most at risk for cardiovascular disease. Higher PHRM-derived RHRs were associated with high body mass index and poor cardiovascular fitness after controlling for covariates, indicating that PHRM correctly captured the directionality of cardiovascular risk. That is the actual clinical signal you would want a population-scale health tool to surface.

This is still a research system, not a product. There is no consumer app launch attached, no Pixel feature announcement, and no pricing. What is shipping is the paper, the dataset, and the PHRM-mini model weights for researchers who can clear the access requirements. For anyone working on contactless vitals, on-device health ML, or fairness benchmarks in physiological sensing, this is now the reference baseline to beat.