Google researchers have developed a passive heart rate monitoring system (PHRM) that leverages smartphone front-facing cameras to track cardiovascular health during daily use. By analyzing 8-second facial video clips captured after phone unlock events, the system utilizes deep learning and photoplethysmography (PPG)—a technique measuring skin light fluctuation to track blood pulses—to estimate heart rate (HR) and daily resting heart rate (RHR). In laboratory settings using 365 participants, PHRM achieved a mean absolute percentage error (MAPE) of less than 10% across all skin tones, outperforming 15 existing remote PPG methods.

A "free-living" study of 231 participants tracked over eight days demonstrated that PHRM maintains wearable-level accuracy in real-world conditions. The system achieved an overall MAPE of 6.09% for HR measurements, with specific results of 5.04% for light skin, 5.12% for medium skin, and 7.84% for dark skin tones. Regarding daily RHR, the system reached a mean absolute error (MAE) of 4.39 beats per minute (bpm), significantly lower than the pre-specified 5-bpm target. The researchers utilized a diverse dataset of over 350,000 video clips from nearly 700 participants to ensure performance inclusivity across the Monk Skin Tone scale, with specific representation mandates requiring at least 33% of participants to have dark skin.

The PHRM pipeline employs computationally-efficient temporal shift convolutional neural networks to process video clips and Kalman filtering to aggregate RHR estimates over time. While the system shows robust cardiovascular risk assessment capabilities—aligning with body mass index and fitness metrics—the authors noted that HR measurement success rates remained lower for darker skin groups, suggesting a need for future optimizations like adjusted camera exposure. To foster continued progress, Google has publicly released the largest available rPPG dataset and a pre-trained "PHRM-mini" model for qualified researchers with Institutional Review Board (IRB) approval, prohibiting any commercial use or attempts to re-identify participants.

Google researchers have developed a passive heart rate monitoring system (PHRM) that leverages smartphone front-facing cameras to track cardiovascular health during daily use. By analyzing 8-second facial video clips captured after phone unlock events, the system utilizes deep learning and photoplethysmography (PPG)—a technique measuring skin light fluctuation to track blood pulses—to estimate heart rate (HR) and daily resting heart rate (RHR). In laboratory settings using 365 participants, PHRM achieved a mean absolute percentage error (MAPE) of less than 10% across all skin tones, outperforming 15 existing remote PPG methods.

A "free-living" study of 231 participants tracked over eight days demonstrated that PHRM maintains wearable-level accuracy in real-world conditions. The system achieved an overall MAPE of 6.09% for HR measurements, with specific results of 5.04% for light skin, 5.12% for medium skin, and 7.84% for dark skin tones. Regarding daily RHR, the system reached a mean absolute error (MAE) of 4.39 beats per minute (bpm), significantly lower than the pre-specified 5-bpm target. The researchers utilized a diverse dataset of over 350,000 video clips from nearly 700 participants to ensure performance inclusivity across the Monk Skin Tone scale, with specific representation mandates requiring at least 33% of participants to have dark skin.

The PHRM pipeline employs computationally-efficient temporal shift convolutional neural networks to process video clips and Kalman filtering to aggregate RHR estimates over time. While the system shows robust cardiovascular risk assessment capabilities—aligning with body mass index and fitness metrics—the authors noted that HR measurement success rates remained lower for darker skin groups, suggesting a need for future optimizations like adjusted camera exposure. To foster continued progress, Google has publicly released the largest available rPPG dataset and a pre-trained "PHRM-mini" model for qualified researchers with Institutional Review Board (IRB) approval, prohibiting any commercial use or attempts to re-identify participants.