Google PHRM Achieves 6.09% MAPE in Passive Heart Rate Tracking

Google Research has introduced a background tracking system that measures cardiovascular metrics through everyday smartphone use. Detailed in a recent research release, the Passive Heart Rate Monitoring (PHRM) system captures minute blood flow changes during routine device unlocks. The technology turns a standard front-facing camera into a continuous health monitor without requiring explicit user interaction.

Remote Photoplethysmography Pipeline

The PHRM system relies on remote photoplethysmography (rPPG) to detect invisible color variations in skin caused by blood flow in facial capillaries. The system triggers passively, capturing a brief 8-second video sequence immediately following a face unlock event or during other routine front-facing camera activations. The rPPG technique isolates specific color channels from the RGB video feed, measuring light absorption changes that correspond with each heartbeat to construct a pulse waveform.

To handle environmental noise, the data processing pipeline utilizes a Kalman filter. This component aggregates multiple intermittent readings over a 24-hour period while discarding outlier data caused by motion artifacts or sudden lighting changes. The aggregated output provides a stable daily resting heart rate estimate.

Validation and Benchmarks

Google partnered with DeepMind and the University of Washington to conduct the largest rPPG validation study published to date. The development phase processed 225,773 videos from 495 participants. The validation phase tested the model on 185,970 videos from 205 participants across uncontrolled real-world environments.

Environment	Mean Absolute Percentage Error (MAPE)
Laboratory Setting	5.65%
Real-World Setting	6.09%

The system’s resting heart rate estimates match dedicated hardware, falling within 5 beats per minute of a Fitbit Charge 6. This precision meets the Consumer Technology Association standards for heart rate monitoring, which require a MAPE below 10%. Similar to strict demographic requirements in other DeepMind medical AI validation efforts, the researchers balanced the training dataset using the Monk Skin Tone (MST) scale to ensure parity. The data pool included at least 25% light skin, 25% medium skin, and 33% dark skin representations.

Clinical and Privacy Implications

A commentary published alongside the research in the Annals of Internal Medicine highlighted concerns regarding clinical maintenance. Medical professionals cautioned that automating basic diagnostic measurements through consumer devices could degrade traditional clinical observation skills, drawing parallels to automation reliance in commercial aviation.

Capturing background facial video also introduces strict architectural requirements. Implementing this technology commercially mandates explicit user authorization and local execution. To avoid transmitting sensitive biometric video to cloud servers, the rPPG feature extraction and Kalman filtering must occur entirely through secure on-device processing.

Public health experts note this passive approach scales cardiovascular monitoring to billions of existing smartphones globally, bypassing the hardware costs of wearables. If you design mobile health applications, you must prepare for a fundamental shift toward passive, intermittent sensor reading pipelines over continuous hardware data streams.