For decades, migraine has largely been diagnosed the old-fashioned way: doctors ask patients what their headaches feel like.
There is no blood test, scan or other established biomarker that can definitively identify the disorder. Instead, clinicians rely on a familiar constellation of symptoms — intense headaches, often accompanied by nausea and sensitivity to light or sound.
But new research suggests that migraine may leave a much broader biological footprint than those symptoms reveal.
Researchers in Norway have used machine learning to identify people with migraine from information that did not include their headache symptoms — and then used artificial intelligence to tease out distinct groups of migraine patients.
The findings, published in the journal Neurology, raise the possibility that migraine is not a single disorder at all, but “a spectrum disorder that can be identified and stratified using broad clinical and biological data—not solely through headache-specific symptoms,” researchers wrote.
“There are many inconsistencies with the diagnosis of migraines. It appears that AI could help us make the correct diagnosis,” said Anker Stubberud, a physician and headache researcher at the Norwegian University of Science and Technology (NTNU).
That could eventually matter for treatment. If different forms of migraine have different biological signatures, doctors may one day be able to predict which patients will respond to which therapies “so that people can receive the best possible treatment,” said Stubberud, rather than relying largely on trial and error.
The researchers analysed data from the Trøndelag Health Study, a large Norwegian population study whose clinical information was collected in the 1990s and 2000s. Their diagnostic analysis included 43,197 people, including nearly 9,000 classified as having migraine.
The researchers fed the AI model dozens of pieces of information about people's health and lives — including demographic characteristics, mental health, cardiovascular and musculoskeletal conditions, sleep, exercise, medication use and other factors — alongside genetic information.
Crucially, the model was not given the defining characteristics of the headache itself.
Even so, the best-performing model achieved an area under the curve of 0.80 in the held-out test set, a measure indicating reasonably strong discrimination between people with migraine and headache-free controls. Adding genetic information produced only a marginal improvement over clinical information alone.
“The fact that AI could identify migraine so accurately without knowing anything about the headache itself suggests that the condition leaves traces that extend far beyond the attacks,” said Stubberud.
That result is important for what it says about migraine biology.
Age was the most important predictor, followed by factors including neck pain, menstruation and nausea, according to the researchers' analysis of which features were driving the model's predictions.
Four types of migraines
The study then took a second step: rather than asking the machine to distinguish migraine from non-migraine headache, researchers asked it to look for naturally occurring groups within the data.
Among 12,185 people with sufficiently complete headache information, the algorithm identified one cluster of 1,425 people in which 94% met the researchers' criteria for migraine. A much larger cluster contained people whose headaches were more often classified as non-migraine.
The migraine-like group could itself be divided into four subgroups.
One consisted exclusively of men. Another was characterised by prominent neck pain. A third had more musculoskeletal pain alongside anxiety and depression. The fourth looked more like what doctors might recognise as “classic” migraine, with people in this group experiencing migraine aura – a set of temporary neurological symptoms which can occur before or during an attack, such as flashing lights, zigzag patterns, blind spots or shimmering areas, but they can also involve tingling or numbness, and difficulty speaking.
Those groups also showed differences in their genetic signals. The researchers found that machine-learning-based genetic risk scores distinguished the groups better than conventional polygenic risk scores.
“This strengthens the hypothesis that migraine is not a single disease, but rather a diverse group of different biological conditions,” said Stubberud.
The implication is potentially bigger than a better diagnostic test.
The findings could help explain a longstanding frustration for patients and doctors: why a treatment that works remarkably well for one person with migraine can do little for another.