Can a smartwatch predict a migraine? What the data actually shows

Short answer

Not yet, and not well. The best published models can flag some migraine days above chance, but they miss most attacks — and the strongest single predictor in the best study was not a wearable signal at all. It was yesterday's headache. Meanwhile, the thing that already predicts attacks reasonably well is not on your wrist. It is you.

The most rigorous test so far followed 146 people with episodic migraine across roughly 21,550 days of diaries plus wearable data. Its best model reached an AUC of 0.84 for predicting the next calendar day — a number that sounds strong until you read the rest of the line: sensitivity 0.39, specificity 0.95. That means it caught fewer than 4 in 10 attacks. Over a three-day window it did better at catching them (sensitivity 0.58) and worse at avoiding false alarms (specificity 0.83). The authors' own conclusion was that low sensitivity limits clinical utility. (link.springer.com)

And if you have been told that you cannot possibly know an attack is coming — you were not imagining it, and it is not your fault that no device confirms it. Knowing an attack is on its way is a documented, measured phenomenon, and humans currently outperform the hardware at it. In the PRODROME screening dataset — 920 people, 4,802 logged early-symptom events — 81.5% were followed by a headache within 1–6 hours and 95.1% within 24 hours. (pmc.ncbi.nlm.nih.gov) An older electronic diary study found people correctly predicted the headache from 72% of entries with early symptoms. (neurology.org) Your sense that it is coming is a better instrument than any watch that has been tested.

If you want to look at what sits behind your own headache pattern rather than waiting for a device to tell you: take the 2-minute check-in — it maps the buildup in your nervous system behind the attacks, which is a different and more useful question than "will it be tomorrow."

What would a smartwatch actually be measuring?

Not the migraine. Migraine is a brain event; a watch sits on an artery in your wrist. Everything a consumer wearable contributes is indirect, and it comes from a short list:

• Heart rate and heart rate variability, as a proxy for autonomic balance

• Skin temperature, which shifts with circadian phase and hormonal cycle

• Electrodermal activity — skin conductance, a sympathetic arousal signal, present on some research devices and few consumer ones

• Sleep duration, timing and continuity, estimated from movement and heart rate

• Activity and movement

The theory behind using them is sound. The premonitory phase involves hypothalamic and brainstem circuits that also govern autonomic tone, sleep and appetite, which is why yawning, thirst and fatigue appear hours before pain. (nature.com) If those circuits shift before an attack, autonomic measurements might shift with them.

The gap between "might shift" and "tells you tomorrow" is where every one of these studies lives.

How good are the published models, really?

Two studies carry most of the weight, and they disagree about almost everything except the conclusion.

The large one. 146 people with episodic migraine, three months of daily headache diaries plus wearable sessions, around 21,550 days for next-day prediction. Standard machine learning barely beat a coin toss — a decision tree managed AUC 0.59, a foundation model 0.55. A time-series model built for irregular data reached AUC 0.84 for the next calendar day, with sensitivity 0.39 and specificity 0.95; over three days, AUC 0.76, sensitivity 0.58, specificity 0.83. The most important features were headache intensity, headache duration, and heart rate scores — in that order. (link.springer.com)

Sit with that feature ranking for a second. The two strongest predictors of tomorrow's migraine were properties of today's migraine. The wearable signal came third.

The small one. Ten people, wearing research-grade sensors, focused on nights before an attack. Electrodermal activity, skin temperature and accelerometer data separated pre-migraine nights best; a gradient-boosting model reached accuracy 0.806, precision 0.638, recall 0.595. The authors noted there was no control group without migraine, that ten people cannot support a generalisable model, and that the model "may not be sufficiently sensitive for clinical application." (pmc.ncbi.nlm.nih.gov)

Notice also what neither study used: a consumer smartwatch. Electrodermal activity is not on most wrists. The best-performing signals are the ones you do not own.

Why does AUC 0.84 not mean it works?

Because AUC answers a question you do not have.

AUC asks: if I hand the model one migraine day and one normal day, how often does it rank them correctly? That is a ranking score. It is not a promise about any single morning.

The numbers that describe your actual morning are these:

• Sensitivity — of the attacks that happen, what fraction does the alarm catch? At 0.39, it misses six out of ten.

• Specificity — of the normal days, what fraction does it correctly leave alone? At 0.95, it wrongly alarms on one normal day in twenty.

Now put your own numbers through it. Say you get four migraine days a month. In a 30-day month:

• The alarm catches about 1.6 of your four attacks. Two and a half arrive unannounced.

• It fires on about 1.3 of your 26 normal days for nothing.

So roughly three alarms a month, of which about half are real, while most of your actual attacks still show up without warning. That is not a device you can make decisions with — and making decisions is the entire point, because migraine treatment works better taken early. An alarm you cannot trust either way gives you nothing to act on.

The three-day version trades one problem for the other: sensitivity rises to 0.58, but specificity falls to 0.83, which on 26 normal days means roughly 4.4 false alarms a month. Catching more attacks costs you a standing state of alert.

Is a person better than a watch at this?

On the current evidence, yes — by a wide margin.

The PRODROME screening period logged 4,802 events in which someone said "a headache is coming." 95.1% were followed by a headache within 24 hours, 90.5% within 6 hours, 81.5% within 1–6 hours. Per person, a mean of 84.4% of their predictions landed within 1–6 hours, and 76.9% of participants were right at least three times in four. (pmc.ncbi.nlm.nih.gov)

Set that against sensitivity 0.39 and the comparison is not close.

Two honest caveats. First, those participants were recruited for being good at it — the study screened for people who said they could reliably identify attacks with early symptoms. This is not the accuracy of a randomly chosen person. Second, the two things are not measuring the same task: humans are reporting what they feel a few hours out, while the models are predicting a calendar day in advance with no symptom input. The models were, in a sense, given a harder job with worse information.

But that is exactly the useful conclusion. The models perform poorly partly because they were not given the premonitory symptoms — which the study authors list among their limitations. (link.springer.com) The signal that works is the one inside your experience, and so far nobody has built the system that combines it with the sensors properly.

What should you do with a wearable if you get migraines?

Stop asking it to forecast, and start using it as a record.

Track your own early symptoms, deliberately. The evidence says this is the highest-yield thing available. Rate four things twice a day — energy, concentration, neck, and light tolerance, 1 to 5 each — then mark attacks. Across the largest datasets, the most common pre-headache symptoms are light sensitivity (57.2%), fatigue (50.1%), neck pain (41.9%), sound sensitivity (33.9%) and dizziness (27.8%). (pmc.ncbi.nlm.nih.gov) Yours will be a subset of that list. Finding which subset is your personal forecasting model, and it beats the published ones.

Use the wearable for the things it genuinely measures. Sleep timing and consistency, alcohol effects on your night, activity load, resting heart rate as a rough load gauge. Sleep disruption and alcohol are among the most frequently named triggers — sleep disturbances at 70.1%, alcohol at 59.0% in a cohort of 632. (journals.sagepub.com) Unlike a migraine forecast, those are things a watch measures directly and reasonably well. What a watch still cannot do is tell you whether what you noticed was a trigger or the attack already starting.

Use it to reconstruct, not to predict. After an attack, look back at the two nights before. You are not looking for an alarm; you are looking for whether your own pattern shows anything repeatable. Over a few months, that is a real question with a real answer.

Be sceptical of any product that claims otherwise. If a device or app claims to predict your migraines, the questions to ask are: what was the sensitivity, in how many people, over how many days, and was the model validated on people it had not seen? Accuracy alone tells you nothing — a model that predicts "no migraine" every single day is over 85% accurate for most people, and useless.

What does the evidence not show?

That wearables cannot ever do this. The field is young, the datasets are small, and the best study is limited by a population recruited from a biofeedback trial — a group more engaged than average. (link.springer.com) Better labels and premonitory symptom input could plausibly change the picture.

That any particular signal is the one. Electrodermal activity and skin temperature looked best in a ten-person study with no control group. (pmc.ncbi.nlm.nih.gov) That is a hypothesis, not a finding.

That the premonitory phase itself is cleanly characterised. A systematic review found pooled prevalence of at least one early symptom ranging from 29% in population-based studies to 66% in clinic-based ones, with very high heterogeneity, and concluded the evidence was insufficient to characterise the phase reliably. (link.springer.com) The foundation everyone is building prediction on is less firm than it looks.

When should you see a doctor?

See a clinician if headaches are becoming more frequent, if you use acute painkillers on 10 or more days a month, if attacks are disabling, if the pattern has changed, or if you have never had a formal diagnosis. No amount of tracking substitutes for a diagnosis, and without one, prevention and early acute treatment are not really on the table.

Seek urgent care for a headache that peaks within a minute, a headache after head injury, headache with fever and a stiff neck, a first severe headache after age 50, or any headache with weakness, confusion, seizure, vision loss, or changed consciousness.

Before that appointment, it is worth knowing what your own pattern looks like: take the 2-minute check-in — it shows the buildup in your nervous system behind the headache pattern, which is the question a forecast can't answer anyway.

How to bring this up with your doctor — and what to ask for

Bring your early-symptom log, not your wearable export. Two weeks of twice-daily ratings plus attack dates is clinically readable. A year of heart rate graphs is not, and will mostly be ignored.

Say the sentence that opens the right door. "I can usually tell several hours before the pain starts. Is there anything I should be doing in that window?" That asks about early acute treatment, which is where the evidence actually is.

Ask these specifically. Should I take my acute medication when early symptoms start rather than waiting for pain? How many days a month can I use it before risking medication-overuse headache? Given my frequency, am I a candidate for preventive treatment? Is any of my tracking data useful to you, or should I stop bringing it?

Bring three numbers. Headache days per month, acute medication days per month, days you could not work or function.

If you are dismissed. "Can you note in my record that I asked about treating the premonitory phase?" A question on the record gets revisited more often than one that isn't. And ask about a referral to a headache specialist.

How Welltory helps

Welltory does not predict migraines. We are not able to, we said so in the data section with the specific reason, and we would rather publish that than a claim that makes a better headline.

What the app is genuinely good for, if you live with migraine, is the part that is hard to do by hand: a continuous record of sleep, resting heart rate, heart rate variability and daily load, measured the same way every day, against your baseline rather than a population average. That is the raw material for looking backwards at the days around an attack — and looking backwards is a question wearable data can actually answer.

And one finding from our own cohort is worth taking personally. Brain fog is reported by 49% of the people in our data who report migraine, against 25% of everyone else, and that difference held up under our verification check when most of what we measured did not. If you have been treating your foggy days as a discipline problem, roughly half the people in your situation report the same thing.

How we made it

Everything quantitative comes from published work by other groups: the time-series forecasting study (Faisal et al. 2026), the pre-migraine night sensor study (Kapustynska et al. 2024), the PRODROME screening period, the electronic diary prediction study (Giffin et al. 2003), and the REFORM trigger cohort. The counter-evidence is included deliberately — the meta-analysis questioning premonitory prevalence estimates, and each prediction study's own statement of its limitations.

We ran our own feasibility check before writing and are reporting its result rather than hiding it: across all 135 columns of Welltory's daily panel there is no migraine event field, so the label a forecasting model requires does not exist in our data and no model was built. The cohort comparison shown uses 4,145 users with wearable-quality data, 396 of whom report migraine in an onboarding survey, with every metric tested against a verification gate that requires a difference to hold within strata of how many other conditions a person reports. Brain fog passed; resting heart rate failed and is reported as failing.

Discounts for blog readers: up to 36% off

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

Explore your data