AI identifies migraine as spectrum disorder with distinct biological subtypes

For as long as doctors have treated migraines, they have done so by listening. A patient describes the pain, the nausea, the way light feels like a weapon, and a clinician nods and writes down a diagnosis. There is no blood test to confirm it, no scan that lights up with certainty. The disorder has resisted the kind of biological fingerprinting that modern medicine has learned to read in so many other conditions. But researchers in Norway have now used artificial intelligence and genetic data from more than 43,000 people to suggest that migraine may be far more than what patients report feeling—that it is, in fact, a spectrum of distinct biological conditions masquerading as a single disease.

The work began with a simple question: could a machine identify migraine patients without ever being told what their headaches felt like? Researchers fed an AI model dozens of pieces of information about people's lives and bodies—age, mental health status, cardiovascular and musculoskeletal conditions, sleep patterns, exercise habits, medication use, and genetic markers—but deliberately withheld any description of the headaches themselves. The model, trained on data from the Trøndelag Health Study, a large Norwegian population study spanning the 1990s and 2000s, succeeded with striking accuracy. Among nearly 9,000 people classified as having migraine and more than 34,000 without, the AI distinguished between the two groups at a level that suggested something real and measurable was there, even when the defining symptom was hidden. "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 Anker Stubberud, a physician and headache researcher at the Norwegian University of Science and Technology.

What the model found most predictive was not dramatic: age mattered most, followed by neck pain, menstruation, and nausea. Yet the combination of these signals, woven together with genetic information, painted a picture of migraine as something systemic, something that touches the body in ways that go unnoticed until a headache arrives to announce it. Adding genetic data to the clinical information improved the model's performance only marginally, suggesting that the biological signature of migraine is written as much in how people live and what their bodies do as in their DNA.

The researchers then took a second step. Rather than asking the machine to sort people into migraine or non-migraine, they asked it to find natural groupings within the data itself. Among 12,185 people with complete headache information, the algorithm identified four distinct clusters of migraine patients. One consisted entirely of men. Another was marked by prominent neck pain. A third showed elevated musculoskeletal pain alongside anxiety and depression. The fourth resembled what neurologists have long recognized as classic migraine—patients who experienced aura, the temporary neurological symptoms that can precede or accompany an attack: flashing lights, zigzag patterns, blind spots, tingling, numbness, or difficulty speaking. These four groups also carried different genetic risk profiles, suggesting that what doctors have treated as a single condition may actually be four separate biological phenomena.

The implications reach far beyond diagnosis. For decades, patients and clinicians have encountered a maddening reality: a medication that eliminates migraines in one person does almost nothing for another. The current approach is trial and error, a process that can stretch across months or years while a patient cycles through treatments. If migraine truly comprises distinct biological subtypes, then the path forward becomes clearer. Doctors might one day match a patient's genetic and clinical profile to their migraine subtype, then select a therapy known to work for that particular form of the disorder. "So that people can receive the best possible treatment," Stubberud said, rather than gambling on which drug might help. The research, published in the journal Neurology, does not yet offer that level of precision. But it suggests the possibility is no longer theoretical. The question now is whether the medical system can move fast enough to make use of what the machines have begun to see.