Cameras can read your pulse from a video. The rest of the vital-sign claims are shakier

An ordinary camera can measure a heart rate. Skin flushes and fades by a tiny amount with each pulse as blood moves through it, and a webcam or phone camera can pick up that colour change in a face and turn it into a beat-to-beat signal — a technique called contactless, or remote, photoplethysmography (rPPG) [s1]. On the one thing it is asked to do most often, reading pulse rate, the pooled evidence says it works: a meta-analysis of clinical studies found camera-derived heart rate almost indistinguishable from an electrocardiogram [s1]. The trouble starts when the same cameras are sold as monitors of breathing, blood oxygen, and stress, where the evidence is thin, low quality, or absent.

This matters because the technology has quietly moved from the lab into consumer and telehealth software. Apps now offer a "scan" that reads vital signs from a selfie video during a virtual visit or a wellness check, and the marketing tends to list a fuller set of numbers than the science supports.

Where the evidence is solid: heart rate

A systematic review and meta-analysis in the Journal of Clinical and Translational Science, published in May 2023, evaluated contactless photoplethysmography in adults in clinical settings [s1]. It included 12 studies with 654 individuals; heart rate was by far the most-studied vital sign [s1]. Pooling the four studies that compared camera-derived heart rate against ECG, the mean bias was −0.13 beats per minute (95% CI, −1.22 to 0.96) — effectively zero, with a confidence interval well under two beats either way [s1]. For pulse rate in a cooperative, adequately lit subject, the camera is a genuinely accurate instrument.

A larger meta-analysis in Anaesthesia in 2022 reached a compatible conclusion by a harsher route [s2]. It gathered 84 studies of novel wearable and contactless devices measuring heart rate, respiratory rate and oxygen saturation in clinical settings, of which 29 assessed contactless devices [s2]. For contactless devices, the pooled heart-rate bias was 2.18 beats per minute, with 95% limits of agreement from −6.71 to 10.88 [s2]. The small bias is reassuring; the width of the limits of agreement is the warning, because limits of agreement — not the average — describe how far a single reading can stray.

Where it gets thin: breathing, oxygen, and everything else

Respiratory rate looked reasonable on average: the Anaesthesia analysis put the contactless bias at 0.30 breaths per minute (limits of agreement −3.94 to 4.29) [s2]. But the authors' overall judgement is the part worth quoting in spirit: they rated the studies of contactless devices as more experimental, more prone to patient-selection and rater bias, and generally of low quality, with small datasets (typically fewer than 100 people), often unblinded and frequently analysed with inappropriate statistics [s2]. Blood-oxygen evidence was thinner still — the pooled review found only a small number of oxygen-saturation studies even among wearables, and none robust enough to anchor a claim [s2].

Then there is the layer vendors add on top: "stress," "wellness," or "recovery" scores, usually derived from heart-rate variability — the beat-to-beat timing differences the pulse signal contains. A 2026 scoping review in Frontiers in Digital Health looked specifically at whether rPPG is physiologically valid for stress and mental-workload monitoring [s3]. Its finding splits cleanly: camera-based heart-rate estimation is robust, but heart-rate-variability reliability degrades under head motion, changing light, and the short analysis windows those stress apps typically use [s3]. Worse, the review noted that many studies validate their stress classifiers against protocol labels or against another wearable, rather than against an ECG or high-quality contact sensor — so the physiological claim underneath the product is frequently untested [s3].

How to read a contactless vital-sign number

The device-agnostic rule that falls out of these three reviews is simple. A camera-derived pulse rate, taken from a still, well-lit subject, is trustworthy to within a couple of beats [s1][s2]. A respiratory rate is plausible but rests on weaker evidence [s2]. A blood-oxygen reading, a blood-pressure figure, or a "stress level" from the same selfie is, on current published evidence, closer to a demonstration than a measurement — the conditions that make rPPG work (no motion, steady illumination, long enough windows) are exactly the conditions a real-world app rarely has, and the metrics that degrade fastest are the ones being marketed hardest [s3].

That pattern — consumer sensing outrunning its validation — is the recurring story of this category. It shaped the Apple Watch sleep-apnea notification, which catches severe cases well and misses moderate ones; it runs through skin-tone differences in wrist optical heart-rate sensors; and it is why cardiologists still decline to endorse cuffless blood-pressure devices despite a run of passing validations. For the regulatory line between a wellness claim and a medical one, see our coverage of general-wellness guidance for wearables.

What to watch

The open question is not whether cameras can sense physiology — they can — but whether the harder measurements survive contact with a moving, differently lit, real population, and whether independent validation catches up before the marketing does. Larger, blinded, pre-registered studies against gold-standard references, especially for respiration, oxygenation and any blood-pressure claim, are what would move camera-based vitals from a convincing demo to a clinical tool. Until then, the honest reading is narrow: trust the pulse, doubt the rest.

Sources

• Clinical applications of contactless photoplethysmography for monitoring in adults: A systematic review and meta-analysis — Journal of Clinical and Translational Science, 2023-05-15
• Novel wearable and contactless heart rate, respiratory rate, and oxygen saturation monitoring devices: a systematic review and meta-analysis — Anaesthesia, 2022-08-10
• When is rPPG physiologically valid for stress and workload monitoring? A PRISMA-ScR scoping review — Frontiers in Digital Health, 2026-08-19