Vital signs without touching the patient

Crowdpulse turns a phone on a stand into a contact-free vitals station. It measures heart rate and breathing from the camera alone, and moves the patients whose numbers are changing to the top of the nurse’s list.

Create an account See how it works

2.5bpm
mean error against a fingertip oximeter, on the seconds it gave a reading: 89%
of readings within 5 bpm of the oximeter: 60%
of seconds gave a reading

People seated facing a camera. UBFC-rPPG public dataset, held-out set: 21 subjects, 804 one-second readings, evaluated 14 November 2026. Read the evaluation

No cuff, no clip, no contact

Vitals are taken once at the desk, and then the waiting room goes quiet. Crowdpulse makes every re-check a minute in a chair, with nothing to attach and nothing to clean.

Check in
The patient sits down facing a phone on a stand. The operator takes their name, a photo and what they have come in for.
Sit for a minute
The camera reads the colour of the forehead and cheeks and the rise of the chest. Pulse and breathing rate settle on screen as the operator watches.
Wait, and stay seen
Every sitting joins the worklist, and a re-check is another minute at the station. When a patient’s numbers move away from where they were at check-in, they rise to the top and the nurse’s phone buzzes.

Video is shown live to staff and is not stored. One profile photo is kept with the patient record.

The list reorders itself around who is changing

Nurses see one ranked worklist on the dashboard and on their phone. It shows each patient’s face, name and reason for the visit beside what was measured and when.

Trend flags: A pulse or breathing rate that has drifted since check-in flags the patient for a person to check.
Overdue rechecks: Anyone who has waited too long since their last sitting is marked, so nobody is forgotten in a corner.
A handoff in one sentence: The nurse’s phone gets what changed and over how long, ready to act on.

Checked against a contact sensor, second by second

We ran the same engine that ships on the phone over public recordings of real people, each filmed while wearing a fingertip oximeter, and compared every one-second reading with the oximeter.

Seated, facing a webcam

The condition the station is built for. UBFC-rPPG: 21 people seated at a webcam, each wearing a fingertip oximeter. The dataset's 42 subjects were divided by a fixed seed, and these 21 are the held-out set.

UBFC-rPPG public dataset, held-out set: 21 subjects, 804 one-second readings, evaluated 14 November 2026.

Second by second Every one-second reading against the oximeter at that second. Error is measured on the seconds when the engine gave a reading, which was 60% of seconds.

2.5bpm
Mean absolute error: 89%
Readings within 5 bpm: −1.0bpm
Bias against the oximeter

When it is not sure, it says No reading

A reading appears only after three consecutive windows agree and the signal clears a quality threshold. Otherwise the dashboard prints the words No reading and the reason, not a dash, a zero or the last good number.

How often it gave a reading: A reading in 60% of seconds, and at least one in 21 of 21 recordings. First reading after a median 8 s. UBFC-rPPG public dataset, held-out set: 21 subjects, 804 one-second readings, evaluated 14 November 2026.
How that share is counted: In each recording, the share of one-second ticks, counting the warm-up, in which the engine gave a heart rate. The shares are averaged over the 21 recordings, each counting equally.

How the test set was kept apart

Subjects were split into a tuning set and a held-out set by a fixed seed before the new recordings were scored. Every figure here comes from the held-out set.

Four of the 21 were not new to the engine: Four of the 21 held-out subjects had been used to check the engine during earlier development. They stay in the figures above. Without them, the mean absolute error is 2.3 bpm, with 90% of readings within 5 bpm (17 subjects, 681 one-second readings, evaluated 14 November 2026).
The reference: The fingertip oximeter heart rate recorded with the dataset. For 6 of the 21 subjects the reference file was checked against the dataset authors' copy; for the other 15, against two other public copies of the dataset.

The figures as a table

Dataset and set	Subjects	Recordings with a reading	Paired readings	MAE, bpm	Bias, bpm	Within 5 bpm	Seconds with a reading	Median time to first reading, s
UBFC-rPPG, held-out set, evaluated 14 November 2026	21	21 of 21	804	2.5	−1.0	89%	60%	8
The same, without the 4 subjects used in earlier engine checks	17	17 of 17	681	2.3	−0.7	90%	60%	8

Method: one heart rate per second from the engine, paired with the reference within 1.5 s. Each video goes through the same path as a sitting on the phone: frames, face regions, signal engine at its default settings. Footage: uncompressed, 640 × 480, 30 fps. Reference: fingertip oximeter heart rate recorded with each video. Dataset: UBFC-rPPG, 42 subjects, divided by a fixed seed into 21 for tuning the engine and 21 held out; every figure here is from the held-out set. Evaluated 14 November 2026. Not a clinical validation.

The camera pulse is established science with an FDA device class of its own

Each heartbeat pushes a little more blood into the skin, and blood absorbs green light. A camera records that as a colour change far too faint to see, repeating at the heart rate. In 2008 researchers recovered it with a consumer camera in ordinary room light¹, and the method has a name: remote photoplethysmography, or rPPG.

One regulation, a growing list

In 2021 the FDA granted a De Novo request for Oxehealth Vital Signs and created a device class for software that measures pulse and breathing rate with an optical camera, 21 CFR 870.2785². Contact-free products from five more companies have been cleared under it since. Each company’s first is listed here.

202126 MarOxehealth Vital SignsOxehealthDe Novo grantedDEN200019²
20231 SepFaceHeart Vitals SDKFaceHeart510(k) clearedK223622³
202418 NovInformed Vital CoreMindset Medical510(k) clearedK241633⁴
202423 DecPanopticAI Vital SignsPanopticAI510(k) clearedK240890⁵
202611 JunAnura Mobile Core SDKNuraLogix510(k) clearedK253650⁶
202618 JunSmartSpectra Vital Signs MonitorPresage Technologies510(k) clearedK254169⁷

Investors are paying for it

Binah.aiRaised a $13.5 million Series B in 2020, led by Maverick Ventures, for vital signs measured with a smartphone camera⁸.
Mindset MedicalAnnounced a Series A in June 2024 led by Grayhawk Capital and Questa Capital, with Amgen Ventures taking part, and was cleared five months later⁹.

Crowdpulse is not one of the cleared products. It applies the same measurement principle to one job: re-checking the waiting room between check-in and triage.

Sources

Set up a station with the phone you already have

Create an account, open the app on a phone in camera mode, and run your first sitting in a few minutes.

Create an account Sign in