Can a Wearable Predict a Migraine Attack? What a New Wrist-Sensor Study Found

Why Predicting an Attack Matters

Ask people with migraine what makes the condition so hard to live with and the pain is only part of the answer. The other part is not knowing when the next attack will land. You can't plan a meeting, a flight, or a family dinner around something that shows up without notice.

Timing also affects treatment. Acute medications tend to work best when taken early. Once an attack is fully underway, the trigeminovascular pain pathways can become sensitized, and that sensitization makes standard acute treatment less effective. Pain relief is more often incomplete, the headache is more likely to come back, and repeat doses become more common, which feeds the risk of medication overuse headache.

Many migraine attacks are preceded by a prodrome (also called the premonitory phase), a stretch of hours before the headache when the brain is already changing course. Yawning, neck stiffness, fatigue, food cravings, and trouble concentrating are common signs. The problem is that these symptoms are vague, they overlap with ordinary daily life, and most people do not recognize them reliably.

A new study in Cephalalgia, the journal of the International Headache Society, tested a different approach. Instead of asking people to notice their own warning signs, researchers in Lithuania and Denmark had participants wear a wrist sensor around the clock and used machine learning to look for changes in the body's autonomic signals in the hours before an attack.

How the Study Worked

The team recruited adults with episodic migraine from the headache clinic at Vilnius University Hospital Santaros Klinikos. Participants had 4 to 14 migraine days per month and were not using any migraine preventive. They also could not be taking medications known to affect the autonomic nervous system, including blood pressure medications, antidepressants, benzodiazepines, and hormonal contraceptives.

Each person wore an Empatica EmbracePlus device on the wrist of their non-dominant hand 24 hours a day, removing it only to charge, until they had recorded at least three migraine attacks. They logged attacks in a paper diary or the Migraine Buddy app and pressed a marker on the device when an attack began. A headache specialist reviewed every entry.

The researchers then labeled the sensor data. Time windows in the 2 to 12 hours before a confirmed attack were marked as "pre-attack," and matched windows from the same time of day on attack-free days were marked as "control." Matching by time of day mattered, because heart rate and skin temperature swing naturally across the day and night.

The key design choice was to build a separate model for each person. For every participant, the computer tested six machine learning methods and several prediction windows, picked the best combination using about three quarters of that person's days, and then checked how well it performed on the remaining quarter it had never seen. The team also built one shared model trained on everyone's data to see whether a single algorithm could work across people.

What the Sensor Picked Up

A usable personal model could be built for 25 of the 27 participants (93%). For the other two, the model either missed every pre-attack window or performed too poorly to count. Across the 25 people with working models, the results were:

Performance varied a lot from person to person. For about two thirds of participants the model discriminated well, and for a smaller group it did not.

How Well the Personal Models Performed

Model performance for each of the 27 participants, grouped using standard cutoffs for prediction models.

No single algorithm worked best for everyone. A simple method called Gaussian naive Bayes was chosen most often (44% of working models), followed by logistic regression (20%) and random forest (16%). Even within one person, swapping the algorithm changed results substantially. Shorter warning windows were linked to slightly higher accuracy, which makes sense: predicting what happens in 2 hours is easier than predicting what happens in 12.

Personalized vs One Model for All

The shared model, trained on everyone else and tested on a person it had never seen, did much worse. It beat chance for only 9 of the 27 participants (33%). For the other two thirds, its predictions were close to random.

Personal Models vs a Shared Model

This matches an earlier Finnish study that used nighttime wearable data and found the same split, with personal models performing well and a general model performing poorly. The physical warning signs of an attack seem to look different from one person to the next. One person's prodrome may show up mostly as a shift in skin conductance, another's as a change in heart rate or temperature. For any future device, that likely means a learning period where it gets to know your patterns before its alerts mean much.

The Prodrome Most People Miss

One of the most useful findings for patients had nothing to do with the sensor. When participants were asked in an open-ended way whether they noticed any warning symptoms before their migraine attacks, only 5 of 27 (18.5%) said yes. When they were handed a structured list of 95 possible prodromal symptoms, 26 of 27 (96.3%) recognized at least one.

The unprompted figure lines up with population studies, where roughly 29% of people with migraine report premonitory symptoms. The gap suggests that most people do have a prodrome but don't connect those symptoms to the attack that follows. That is the main reason researchers are interested in objective signals: a sensor doesn't need you to notice that you yawned more than usual this afternoon.

What Early Warning Could Change

An early warning is only useful if there is something effective to do with it. Here the timing is good. The PRODROME trial, published in The Lancet in 2023, tested ubrogepant (Ubrelvy) taken during the prodrome, before any headache began. People who took ubrogepant (Ubrelvy) at the first warning signs were less likely to go on to develop a moderate or severe headache than when they took placebo. That trial only enrolled people who could reliably recognize their own prodrome and required careful training, which is exactly the group this new study suggests is small.

A reliable sensor could, in theory, extend prodromal treatment to people who never notice their warning signs. It could also help people plan: rest, eat, hydrate, cancel a late night, or keep acute medication close by.

The authors point out that any real system will face this trade-off. Catching more attacks means more false alarms, and fewer false alarms means missing some attacks. Someone with 12 migraine days a month and a history of rebound headache may want a strict alert. Someone with 4 migraine days a month may prefer a sensitive one. A future device would likely need settings tailored to attack frequency, medication use, and personal preference.

Where This Stands Today

The authors describe their results as internally validated proof-of-concept evidence, and several features of the study support that cautious framing.

The next step is prospective testing: larger groups, more diverse patients, real-time alerts, and a check on whether those alerts lead to better outcomes, such as fewer severe attacks or fewer missed workdays.

Key Takeaways

For now, the most practical lesson is the one about awareness. Most people in this study did have warning signs, and a simple checklist revealed them. Tracking your own prodrome and talking with your clinician about early or prodromal treatment options is something you can start this week, while the wearable research catches up.