AI research has reached the point where your phone can stare at your face for eight seconds and make a decent guess at your heart rate, which is either medical progress or the world’s most nervous selfie.

The Nature item highlighted by Benjamin Thompson points to a new study by Liao and colleagues in Nature about passive heart-rate monitoring during everyday smartphone use. The idea is wonderfully odd in that way modern machine learning often is: use tiny color changes in your skin, too faint for your eyes to notice, and let a deep-learning system estimate your pulse from ordinary front-camera video [1]. Back in my day, if you wanted your heart rate, you used two fingers, a clock, and the quiet dignity of being wrong by several beats per minute.

The Old Trick With New Shoes

The science underneath is called remote photoplethysmography, or rPPG, because apparently “camera pulse reading” was too easy to pronounce. Traditional PPG is what your smartwatch or fingertip pulse oximeter does: shine light into skin, measure how blood volume changes as your heart beats, and turn that light signal into a pulse wave. rPPG tries to do the same thing without touching you. It watches the face, where each heartbeat makes skin reflect light just a little differently.

The hard part is that life is rude. People move. Lighting changes. Phone cameras compress video. Skin tones vary. Someone unlocks their phone in a kitchen lit like a haunted refrigerator. Old-school signal-processing methods could work nicely in polite lab conditions, the way a napkin sketch works nicely until you try to build a bridge from it.

Liao’s team went bigger. They built a passive heart-rate monitoring system, or PHRM, trained and tested on a very large smartphone video dataset. The model used short eight-second clips from the user-facing camera, stabilized and preprocessed the face video, then predicted heart rate with a confidence score. Instead of forcing the model to cough up one lonely number like a contestant under pressure, they treated heart rate as a classification problem over plausible beats per minute, from 40 to 180 [1].

That uncertainty bit matters. If the video is shaky, dim, or full of nonsense, the model can act less certain. A wise habit. We could all use more machines that know when they are squinting.

What They Actually Found

The study was developed using 192,353 videos from 485 participants and validated on 162,546 videos from 211 participants across lab and free-living settings [1]. That “free-living” phrase means people were using their own phones in ordinary life, not sitting under perfect lights like a tomato in a grocery-store display.

Compared with ECG reference measurements, PHRM achieved heart-rate error below the 10 percent mean absolute percentage error threshold across light, medium, and dark skin-tone groups. The authors also report that its performance was non-inferior across those skin-tone groups, which matters because optical health sensors have a long and not especially charming history of working better for some skin tones than others [1].

For daily resting heart rate, the phone-based estimates had a mean absolute error under five beats per minute compared with a wearable tracker. That is not a cardiologist in your pocket. It is more like a careful neighbor who checks the mailbox often enough to notice patterns.

Why This Is More Than A Party Trick

Resting heart rate is not just trivia for gym people and smartwatch dashboards. Long-term changes can signal shifts in fitness, illness, stress, sleep, or cardiovascular risk. Wearables already do passive tracking, but not everyone owns one, charges one, or wants another wrist gadget nagging them like a tiny coach with Bluetooth.

Phones, though? Phones are everywhere. The paper notes smartphone ownership is broad, and regular phone use creates many chances for opportunistic measurement [1]. That could make heart monitoring more available, especially if the system runs privately on-device. The authors discuss protected execution and face authentication as ways to keep video secure and avoid measuring the wrong person [1]. Good, because nobody wants their phone accidentally logging Uncle Pete’s pulse during a family FaceTime and deciding you had a stressful afternoon.

The Fine Print By The Hearth

This field is lively, and the neighbors have been busy. A 2026 rPPG roadmap lays out the remaining obstacles: motion, lighting, compression, frame-rate variation, datasets, skin-region selection, filtering, and evaluation metrics [2]. Another 2026 study proposed physiology-informed correction to clean up bad rPPG heart-rate estimates under noisy conditions [3]. Debnath and Kim combined wavelet denoising with adaptive Kalman filtering, reporting low errors on benchmark and custom datasets, including varied skin tones [4]. And the 2024 PhysFlow paper attacked a particularly thorny problem: many rPPG datasets still underrepresent darker skin tones, so the authors used skin-tone transfer augmentation to reduce biased errors [5].

So, is your phone ready to replace medical equipment? No. Put that notion back on the shelf before it breaks something. The best version of this future is not “diagnosis by selfie.” It is quiet, low-friction trend tracking that may help people notice when something changes, then seek proper care when needed.

Still, there is something charming here. We taught machines to read faint color pulses from a face, not because the face is magical, but because biology leaves breadcrumbs everywhere. Back when I was training my first neural net, we had two layers and we were grateful. These kids today hand a phone camera some micro-blushes, a pile of videos, and a neural network, and the thing comes back with your pulse like it has been eavesdropping on your arteries.

References

1. Liao, S. et al. “Passive heart-rate monitoring during smartphone use in everyday life.” Nature (2026). https://doi.org/10.1038/s41586-026-10507-6
2. Elgendi, M. et al. “Roadmap of remote photoplethysmography from heart rate measurement toward clinical translation.” npj Digital Medicine (2026). https://doi.org/10.1038/s41746-026-02715-1
3. Tian, Y. et al. “Adaptive physiology-informed correction for reliable remote photoplethysmography heart-rate monitoring.” npj Digital Medicine (2026). https://doi.org/10.1038/s41746-026-02386-y
4. Debnath, U. and Kim, S. “Advanced signal-processing framework for remote photoplethysmography-based heart rate measurement.” PLOS ONE 21, e0340097 (2026). https://doi.org/10.1371/journal.pone.0340097
5. Comas, J., Alomar, A., Ruiz, A. and Sukno, F. “PhysFlow: Skin tone transfer for remote heart rate estimation through conditional normalizing flows.” arXiv:2407.21519 (2024). https://doi.org/10.48550/arXiv.2407.21519
6. Thompson, B. “Your phone can use tiny skin-colour changes to measure your heart rate.” Nature Podcast (2026). https://doi.org/10.1038/d41586-026-01793-1

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.