Which health tracker is accurate, and is it safe to wear one all day?

The short answer

Two questions, and most buying guides answer neither.

Which numbers on it are real. Some are excellent and some are invented, and the difference does not run between brands. It runs between measurements, on the same device, on the same wrist, in the same study.

And whether it is doing anything to you. A tracker sits against one patch of skin for sixteen hours a day for years, shining a light into it and talking to your phone. That is a fair thing to ask about, and the answer turns out to be about glue and metal rather than about anything the sensor emits.

How accurate are trackers overall?

That is also why the sections below name measurements. The findings per measurement hold across brands and generations, because they follow from what the sensor can physically see. The findings per model do not, because the model that was tested is two or three versions behind the one in the shop.

Which numbers can you trust?

Heart rate. The green light is doing something it is genuinely good at. Across the reviews, wearable heart rate sits at a mean bias of about 3% ^[1]. The Apple Watch meta-analysis puts the mean bias at 0.12 beats per minute, and no subgroup exceeded the 10% error threshold used to call a measurement valid ^[2].

One caveat that matters in practice. That same analysis reports limits of agreement from about minus 11 to plus 11 beats. The average is excellent and a single reading can still be well out, which is the argument for reading your resting heart rate as a weekly line rather than as this morning's number.

An irregular rhythm. This is the one that is worth the price of the device on its own. Pooled across the reviews, wearables detect arrhythmia with a sensitivity of 100% and a specificity of 95% ^[1]. A watch that tells you your rhythm looks wrong is a watch worth listening to.

Steps. Mean absolute percentage errors run from about minus 9 to plus 12, and most devices undercount ^[1], which is the right direction for a number people use as a target. The Apple Watch came in at about two steps per minute below the reference ^[2].

Blood oxygen. A mean absolute difference of up to 2.0% ^[1]. Enough to notice a pattern, not enough to act on a single reading, and not a medical device.

Which numbers should you ignore?

Calories burned. Errors from minus 21% to plus 15% across the reviews ^[1]. For the Apple Watch specifically, every subgroup of the energy expenditure analysis exceeded the 10% validity threshold, while none of the heart rate subgroups did ^[2].

Exercise intensity. The worst thing on the screen: a mean absolute error between 29% and 80%, depending on how hard the activity was ^[1]. Your watch does not know how hard you worked. It knows how much your wrist moved.

VO2max. Overestimated by about 15% on resting tests and 10% on exercise tests ^[1]. Systematically flattering, and always in the same direction, which at least leaves the trend over months worth something even though the absolute figure is not.

How accurate is sleep tracking?

Seventeen minutes is the right order of magnitude for a number you read as a monthly average and the wrong one for judging a single night. The umbrella review reads it the same way: a tendency to overestimate total sleep time, with errors typically above 10% ^[1].

Stages are harder, and one review put three named devices side by side.

Heart rate variability sits between the two groups. Pooled against ECG, portable devices show a small absolute difference, an effect size of 0.23, with high heterogeneity ^[5]. What moved the error was not which device you used. It was which HRV metric was reported and what position the person was in.

Which device is best per measurement?

This is the part every other list gets to first and this one has earned. Where a study put devices side by side, it says which won and on what. Where none did, it says so instead of guessing.

Rhythm: any device that offers it, and prefer one that will show you a trace. The 100% sensitivity figure is pooled across devices rather than won by one, so there is no evidence-based winner here. There is a reason to prefer a device that records something a doctor can look at, because a notification with no trace behind it starts a conversation nobody can finish.

Heart rate and steps: the Apple Watch, on a technicality that matters. Not because it beat the others, because almost nobody has run that comparison. Because it is the one consumer device with a meta-analysis of its own, 56 studies and 270 effect sizes, and it passed on both ^[2]. Being the most-checked device is a different claim from being the best one, and it is the only claim the literature supports.

Total sleep time: WHOOP. Out by 1.4 minutes against a sleep laboratory in the head-to-head, which is the best single result on this page ^[4].

Sleep stages: the Fitbit, and only if you accept how rough all of them are. Best REM agreement of the three and the highest sensitivity to deep and REM sleep ^[4]. The authors' own conclusion is that all three need to improve at staging, and that is the honest headline over this recommendation.

HRV: buy on the protocol, not the brand. The review found no difference between devices and a large difference between metrics and positions ^[5]. So the thing worth paying for is whatever you will actually use lying still, at the same time, every morning, reporting the same metric.

VO2max, calories and intensity: buy nothing. These are the three numbers with the worst evidence and they are on every box. A real VO2max test at a sports laboratory costs less than a watch and gives you a figure that is not inflated by a tenth.

And for everything else, buy on what no study can measure. Battery life, whether it is comfortable to sleep in, whether you can export your own data, and whether the numbers stop when you stop paying. Those decide whether you are still wearing it in six months, and a tracker in a drawer has an accuracy of zero whatever it scored against polysomnography.

Is it safe to wear one all day?

A separate question, so a separate search: (wearable* OR smartwatch* OR "fitness tracker*" OR wristband* OR "activity tracker*") AND ("contact dermatitis" OR "contact allergy" OR "skin sensitization" OR nickel OR acrylate* OR "adverse skin" OR "skin reaction*"), which returns 328 records. None of the accuracy conclusions above rest on it and it rests on none of them.

The answer is not the one most people expect. The documented harm is dermatological, and it comes from the strap and the glue rather than from anything the device emits.

The case literature names devices. Nickel in an Apple Watch, in a paper whose title points out that this is what European nickel-release rules exist to prevent ^[8]. Acrylates in the glue of a smartwatch screen protector, twice ^[10]. And one case where it did not stop at a rash: a nickel reaction to a smartwatch that resolved into contact leukoderma, permanent loss of pigment in the shape of the device ^[9].

Who checked the materials is worth a sentence of its own. The most careful safety assessment of these chemicals we found derives sensitisation thresholds for isobornyl acrylate, acryloylmorpholine and two others using in vitro methods ^[11]. It is good work. Its corresponding address is a manufacturer's biocompatibility team and one author is a paid consultant to that manufacturer, which is stated in the paper and belongs here too. Better than nobody checking, and it is still the industry marking its own homework.

Radio waves. Measured rather than argued about. Infrared thermography and field simulation of a wearable wireless device put the maximum average absorption at the skin at 50 to 54 mW/kg, with a temperature change of less than one degree ^[12]. For scale, exposure limits for a limb are set at 4 W/kg, so that is on the order of a hundredth of the limit. One study of one antenna rather than of the watch you own, and it is the measurement that exists.

The green light, which is what people actually ask about. We searched for studies of chronic exposure to a photoplethysmography sensor: LED light on skin and safety, photoplethysmography and adverse effects, wearable light exposure and skin damage.

What can be said is what surrounds the gap. Visible light at these wavelengths is delivered to skin deliberately and at far higher doses in photobiomodulation, which is a treatment rather than a hazard. That is an argument by analogy, not a result, and this site does not dress one up as the other. If you want the absence of that exposure, a device that reads movement only, or a chest strap you put on for a workout and take off, is the version of a tracker that has no light in it at all.

Does skin tone affect accuracy?

Photoplethysmography shines green light into skin and reads what comes back. Melanin absorbs green light. So the question is obvious, and it has been asked.

Inconclusive is the honest word and it is not the same as reassuring. If a device's core sensor may work less well on you, that is worth knowing before you spend three hundred euros on the numbers it produces, and it is on no box.

What we earn from this page

Nothing, and that is why it can be shaped like this.

We sell no trackers, take no commission on any of them, and there are no affiliate links here. Every list of the best fitness trackers you have read was paid per click on the buy button, which is why they all name a winner and none of them opens with the fact that nine in ten devices have never been validated.

We do intend to build our own, and it will be free. That is worth declaring now rather than later, because it is the obvious future reason to distrust this page. So here is the commitment attached to it. This page stays, in this shape, and our own device gets measured against the same reference standards and reported here on the same terms. If it is worse at something than a Fitbit, this page will say so.

A comparison you can only fail is not worth publishing, and one we quietly retire the day we have something to sell is worse than never having written it.

What we do not know

The device you can buy has almost certainly not been tested. Everything above is inference from measurements on older models with the same class of sensor. That inference is reasonable and it is still inference.

Almost nothing has been validated in the people most likely to be given one. The literature runs on healthy adults in laboratories, not on older people, or people with an arrhythmia, or the tremor and gait that change what an accelerometer sees.

The skin literature is case reports and one review that calls itself incomplete. How often a tracker causes a reaction is genuinely unknown, and the reviewers say so.

And nobody has looked at chronic light exposure at all. That is the largest gap on this page, it is the question people ask most, and the honest answer to it today is that there is no answer.