Google Research has developed a system that monitors heart rate passively through smartphone cameras during regular device use, without requiring users to place their finger on the sensor.

The Passive Heart Rate Monitoring (PHRM) system captures 8-second video clips of users' faces through the front camera after face unlock events. It then applies deep learning to estimate heart rate with a mean absolute percentage error below 10% compared to electrocardiogram readings.

The technology works by detecting photoplethysmography signals — fluctuations in how light interacts with skin as blood pulses through it. Google's system uses temporal shift convolutional neural networks to process facial video and predict heart rate alongside confidence scores.

Addressing skin tone bias in health tech

Google trained the model on over 350,000 video clips from nearly 700 diverse participants. The company deliberately ensured participants with light skin comprised at least 25% of datasets, medium skin tones 25%, and dark skin tones at least 33%.

This approach addresses longstanding accuracy issues in photoplethysmography-based devices for people with darker skin. Previous remote PPG studies significantly underrepresented dark-skinned participants, leading to lower accuracy rates similar to problems seen with pulse oximeters.

The system aggregates heart rate measurements throughout the day to estimate daily resting heart rate with a mean absolute error below 5 beats per minute compared to wearable trackers. This matches the accuracy of dedicated fitness devices.

Google has released the largest publicly available dataset of smartphone videos for research alongside a pre-trained "PHRM-mini" model. Qualified researchers can apply for access to the dataset.

The research was published in Nature and represents a step toward making heart health monitoring accessible to the roughly five billion smartphone users worldwide. The technology could particularly benefit populations in low-resource environments where wearable devices have limited adoption.