How accurate is HRV from a phone camera?

Short answer

Accurate enough for time-domain metrics when you sit still. In a peer-reviewed study, Welltory's camera measurement showed no differences from a simultaneous electrocardiogram in 26 professional cyclists measured lying down and seated, with correlations of r = 0.77 to 0.94 and a standard error of measurement below 6%. The same study found no difference between one-minute and five-minute readings. What a camera cannot do is measure while you move, run overnight, or reliably produce frequency-domain metrics from a short recording — and the validation literature as a whole was built on healthy, mostly athletic adults.

Note: this is a wellness measurement, not a medical test. It is not an electrocardiogram and does not detect arrhythmias.

Accuracy is about where the sensor sits, not what it costs

If you are deciding whether you need a device at all, start with measuring HRV without a wearable. This article is about how much to trust the reading once you have it.

Every consumer method of measuring heart rate variability does the same arithmetic on the same raw material: the intervals between heartbeats. What differs is how cleanly those intervals are detected — and that depends on the physics of the sensor and on whether the body part it is attached to is moving.

Two things follow from this table that are worth stating plainly.

A chest strap reads electricity; everything else reads light. Optical methods — photoplethysmography — infer the beat from changes in blood volume, which means the signal degrades when the tissue moves relative to the sensor. That is a property of the technique, not a flaw in any one product.

A still fingertip is an easy place to read a pulse. The finger is densely perfused, and during a seated reading there is almost no motion artefact. A controlled one-minute fingertip recording is, for beat detection, a favourable situation — which is why short camera readings hold up well.

What the research actually found

Our own validation

Twenty-six professional cyclists were measured at rest, both supine and seated, with an electrocardiogram, a chest-strap app and Welltory's camera measurement running at the same time. Against ECG, Welltory showed no differences in either position, correlations of r = 0.77 to 0.94, reliability rated good to excellent, and a standard error of measurement below 6%. The study also compared five-minute and one-minute recordings and found no difference between them.

What this does not cover. The participants were 26 elite endurance athletes. That is the population the authors studied, and it is the population their conclusion describes. We are not going to stretch it into a claim about everyone.

How short a recording can be

This has been studied separately and at scale. In 3,387 adults, time-domain HRV computed from 120-second recordings agreed with the four-to-five-minute reference standard at r = 0.986 for RMSSD and r = 0.956 for SDNN. Even ten-second recordings produced valid RMSSD (r = 0.853–0.862), though SDNN needed at least thirty seconds. The authors' conclusion was unambiguous: it is unnecessary to use recordings longer than 120 seconds to obtain accurate measures of RMSSD and SDNN in the time domain.

Two caveats they state themselves, and we repeat because they matter:

• The recordings were beat-to-beat finger pressure, not a phone camera. That study validates the duration, not the sensor. The sensor was tested separately.

• The conclusions apply to time-domain metrics only. Frequency-domain analysis is a different problem — recordings under 60 seconds cannot support high-frequency analysis at all, and low-frequency components need at least 120 seconds.

Why frequency metrics are treated differently

Total Power, HF, LF and VLF are computed from a frequency spectrum, and a spectrum needs enough cycles to exist. In practice that means at least 300 R-R intervals — roughly five minutes at a typical resting heart rate. Agreement between optical smartphone recordings and ECG is good for the time domain and weaker for LF and HF.

This is why Welltory keeps frequency analysis in a separate longer measurement mode rather than showing it on every reading. If a long reading comes back with fewer than about 280 usable intervals, the app says so instead of producing a spectrum anyway.

The gap nobody advertises

Here is the part that rarely appears in accuracy comparisons, including favourable ones.

Validation studies for consumer HRV devices have overwhelmingly been run on healthy, often athletic adults. That is not a conspiracy — athletes are easy to recruit, compliant with protocols, and produce clean signals. But it means the published accuracy figures describe performance in close to the best possible conditions.

The people most likely to care about their autonomic state are frequently not in those conditions. Dysautonomia, tremor, Raynaud's and poor peripheral circulation all degrade optical signal quality. Lower absolute HRV values — common in energy-limiting conditions — mean a smaller dynamic range, so the same absolute noise becomes a larger relative error.

This matters most in energy-limiting conditions, where a reading can also mislead for physiological reasons — that is a separate problem, covered in when a good HRV reading is not good news.

So the honest position is this: treat any published accuracy figure, ours included, as an upper bound measured in favourable conditions. If you live with a chronic condition, the number that matters is whether your own readings behave consistently across a week. That you can check yourself, and it is more informative than anyone's correlation coefficient.

What "accurate" actually means here

Accuracy claims in this field are usually reported as a correlation coefficient, and a correlation on its own can hide a real problem. It is worth understanding the three numbers that matter, because they answer different questions.

Correlation (r) tells you whether two methods move together. If your HRV rises on Tuesday, does the other method also show a rise? A high correlation means the methods agree about direction and relative magnitude. It does not mean they produce the same number.

Bias tells you whether one method reads consistently higher or lower than the other. Two methods can correlate almost perfectly and still differ by a fixed offset. For tracking your own trend, a constant bias is harmless. For comparing your number against someone else's, or against a published norm, it is not.

Limits of agreement tell you how far apart two methods can be on any individual measurement. This is the number that matters if you plan to act on a single reading rather than a trend.

The researchers who studied recording length make this point explicitly: for one metric in their data, the correlation stayed the same between two recording lengths while the agreement statistic improved substantially. Reading only the correlation would have led to the wrong conclusion. It is a useful reminder that when a company quotes a single r value at you — ours included — you are seeing one third of the picture.

This is also why the practical advice in this article is to compare methods on your own body across a week. You are effectively measuring your own limits of agreement, in the conditions you actually use.

How measurement conditions change the number

Before blaming hardware, it is worth knowing how much of the variation is behavioural. These effects are large — often larger than the difference between devices.

Posture. Lying down and sitting produce meaningfully different values from the same person minutes apart. This is real autonomic response, not error. In some conditions the postural change is large enough to be used diagnostically in clinical settings.

Breathing rate. Slow or paced breathing raises heart rate variability substantially and immediately. A reading taken while breathing deliberately is not comparable to one taken while breathing normally, regardless of which device took it.

Time of day. Autonomic balance shifts across the day — sympathetic activity tends to be higher in the morning, parasympathetic in the evening. Comparing a morning reading to an evening one tells you about the time of day as much as about you.

Recent activity. Heart rate takes time to settle. Measuring while it is still falling produces a drifting signal within the recording itself, which lowers quality and skews the result.

Recording length. Shorter recordings are noisier, and the effect is not the same for every metric: RMSSD tolerates short windows well, SDNN needs longer, and frequency-domain metrics need longer still.

Put together, these mean that a careful reading from a phone camera can easily be more informative than a careless one from an expensive device. Consistency of conditions beats hardware in most real-world use.

Sensor by sensor, in more detail

Chest strap. Detects the electrical depolarisation of the heart, the same physical event an ECG records. Beat detection is therefore the cleanest available outside a clinic, and it holds up during movement. The trade-off is comfort: few people wear one continuously, which rules out overnight and all-day data for most users.

Fingertip with a phone camera. The finger is densely perfused and, during a seated reading, essentially motionless. Light from the flash or screen passes through the tissue and the camera records the resulting changes in absorption. The limits are structural: you have to be still, you have to hold the position, and it cannot run continuously.

Finger ring. Optical, but at a site with good perfusion and relatively little movement during sleep — which is why rings do well at overnight measurement specifically. The trade-off is that they cannot be worn for some kinds of training, and their strength is the night rather than the moment.

Wrist device. The most convenient position and the hardest one optically. The wrist has less perfusion than a fingertip, the sensor sits against a moving joint, and fit and skin tone both affect signal quality. Wrist devices compensate with continuous sampling and heavy filtering, which works well for trends and less well for a single precise reading.

What actually ruins a reading

In practice, accuracy is lost to behaviour far more often than to hardware. Welltory scores every measurement on two components — behaviour quality and technical quality — and these are the things that move the score.

Movement. Any movement changes the signal. For a camera reading, that includes shifting your finger or changing how hard you press.

Controlled breathing. Deliberate slow or even breathing raises HRV substantially. It is a real physiological effect, not an artefact — which is exactly why it makes readings incomparable. Breathe normally.

Talking. Changes your breathing rhythm, which changes your heart rhythm.

An unsettled heart rate. If you measure while your heart rate is still coming down from stairs or a walk, the reading drifts during the recording and accuracy drops. Wait 10–15 minutes.

Cold hands. Poor peripheral perfusion is one of the most common causes of a weak optical signal.

Room light. Too bright or too dark both interfere with a camera reading.

Arrhythmia. If you take a reading during an arrhythmia episode, accuracy will be low — and our algorithm cannot yet tell arrhythmia apart from a poor signal. It may take several attempts.

How to compare your own methods

If you have both a phone and a wearable, you can settle the question for your own body rather than trusting anyone's table.

• Take a camera reading and a wearable reading back to back, seated, without moving between them.

• Repeat on five or six different days, at the same time of day.

• Compare the trend, not single values. Different sensors and different processing will not produce identical numbers, and they do not need to. What matters is whether they move together.

If they track each other across a week, either is usable for you. If one is erratic while the other is stable, you have learned something specific about your own signal that no published study could have told you.

When two devices disagree

This happens often enough to be worth a plan, because the instinct — deciding which device is lying — is usually the wrong question.

Different absolute numbers are expected. Sensors detect beats differently, and every app applies its own filtering, artefact handling and scoring on top. Two products can be measuring the same heart correctly and report different values. Absolute agreement across brands was never the goal and is not a realistic standard.

Different directions are the signal worth investigating. If one method says your variability rose across the week and the other says it fell, something is wrong — and it is usually measurement conditions rather than hardware. Check whether the two were taken in the same position, at the same point in your day, and with the same breathing.

A device that is erratic for you specifically is telling you something useful. Optical signal quality varies between people. Cold hands, tattoos over the sensor site, tremor and circulation all affect it. If one method produces wildly scattered values for you while another is stable, that is not a product defect in the abstract — it is a fit problem between that sensor and your body, and the practical answer is to use the one that works.

Do not average them. Combining values from methods with different biases produces a number that describes neither. Pick the one you will use consistently and keep the other as context.

How artefacts are handled, and why it matters

A raw optical recording always contains imperfections: a missed beat, an extra peak from a movement, a stretch where the signal drops out. What an app does with those has a large effect on the final number, and it is rarely disclosed.

Filtering too little leaves artefacts in, which inflates variability metrics — a single missed beat looks like an enormous interval. Filtering too aggressively smooths away real variation and flatters the result. Welltory's approach is to detect low-quality segments, exclude beats that fail quality checks, and then report how much of the recording survived, through the accuracy score and the total beats count. That is why a reading can come back marked low quality rather than simply producing a confident-looking number from bad data.

The practical consequence for you: check the accuracy score before you interpret a reading, and treat the total beats count as part of the result rather than a technical detail.

What none of these methods can do

• Diagnose anything. HRV is a wellness signal. It is not an electrocardiogram and it does not detect arrhythmias.

• Replace continuous monitoring. A spot reading describes the moment it covered.

• Give reliable frequency-domain metrics from a short optical recording.

• Work equally well for everyone. Readings may be unreliable if you are under 18 or pregnant, and optical signals are weaker for some bodies than others.

When to see a clinician rather than an app: fainting or near-fainting, chest pain, a racing or irregular heartbeat that does not settle, breathlessness at rest, or a resting heart rate that changes sharply and stays changed.

Discounts for blog readers: up to 36% off

See what affects your energy, stress, sleep, and daily state with Welltory

Explore your data