Problem

A phone camera can estimate pulse when a dedicated sensor is unavailable. Motion or weak contact can distort the signal and still produce a plausible number. PulseCheck explores when the camera signal supports showing an estimate and when a repeat is warranted.

What we built

PulseCheck combines adaptive peak detection with learned scoring of waveform timing, spectral evidence, and signal slope. A quality gate can withhold a pulse estimate and request another capture. The latest version checks whether the spectrum supports the rate it displays. The web demo replays synchronized fingertip PPG and ECG and shows both an improved case and a remaining error.

Evaluation

We developed and locked the pipeline using 1,540 windows from 24 people in BUT PPG v2.0.0. We evaluated 386 ten-second recordings from six other people, with same-window ECG QRS intervals as the reference. We counted displayed estimates with more than 5 bpm absolute error as incorrect.

Incorrect displayed estimates fell from 28.50 to 28.24 per 100 attempts. Repeat requests for estimates within tolerance stayed at 5.18 per 100, and all repeats stayed at 15.54 per 100. One held-out reading moved into tolerance. The latest version still missed tolerance on 109 of 326 displayed estimates. The repository includes the split, frozen settings, outputs, and reproduction instructions.

Who benefits, and who could be harmed

People who want a quick pulse check and lack a dedicated sensor are the intended future users. A repeat prompt could help them avoid relying on a weak capture. A wrong reading could falsely reassure someone, cause unnecessary concern, or delay seeking help. Frequent repeats could also make the feature hard to use.

Limitations and what we learned

Six people with repeated recordings provide limited evidence about new devices and populations. The benchmark used recorded signals and simulated repeat decisions. Live browser capture uses a separate experimental estimator that still needs testing against ECG on real phones.

What we would do with another month

Collect consented phone and ECG recordings across devices and conditions. Predefine the scoring and participant split, report results by device and condition, and test whether prompts help people capture a better signal.

Credits

Andrea Nemcova and colleagues provided BUT PPG v2.0.0 through PhysioNet under CC BY 4.0. OpenAI Codex assisted with analysis, code, and writing. OpenAI ImageGen generated the setup illustration, and OpenAI’s Speech API generated the demo narration.

Live demo · Source and README · Evaluation details · Full credits

Share this project:

Updates

Submission history