Migraine Is Not One Disease, but Many: New Insight from a Machine Learning Study

Diagnostic modeling using non-headache features, along with phenotypic modeling, supports the idea of migraine as a spectrum disorder.

That migraine is not a single condition – what drives the disease at the cellular and molecular level in one person may differ from what drives it in another – is a key insight to have emerged from decades of research in the field. But that heterogeneity may not be reflected in current diagnostic criteria. A new machine learning study now provides fresh evidence that migraine may be better viewed as a spectrum disorder, rather than as a single entity. In the process, it confirms that the research community’s interest in using artificial intelligence (AI) to advance understanding of migraine continues to pay dividends.

Researchers in Norway applied machine learning (ML) to data from the Trøndelag Health Study (known as HUNT), a large population-based cohort study from Trøndelag County in Norway. They found that ML diagnostic models trained on general non-headache clinical data could accurately classify those who had migraine from those who did not. Meanwhile, ML phenotypic models identified four relatively distinct subclusters of migraine patients, providing new data to support the idea that not all migraine patients are the same.

The study “suggests that the migraine diagnosed by [International Classification of Headache Disorders (ICHD)] criteria may represent the phenotypic core of a broader biological continuum and that a meaningful percentage of patients we see in practice, particularly those with prominent musculoskeletal pain and psychiatric comorbidities, may represent a distinct subtype,” wrote Marina Romozzi, Università Cattolica del Sacro Cuore, Rome, Italy, in an accompanying editorial. Romozzi was not involved with the new study.

“I think this is important and clinically relevant work,” wrote Masahito Katsuki, Dublin City University, Ireland, in an email to Migraine Science Collaborative (MSC).

Katsuki, who is a headache specialist as well as a researcher using AI approaches to better understand headache but was not part of the current work, highlighted the study’s use of non-headache data in the diagnostic modeling part of the research. That approach, he said, rather than replacing diagnosis, could serve as a screening tool that might help identify individuals who merit more detailed headache assessment even though headache may not be the main thing that brought them to the doctor’s office.

The study and editorial appeared in the July 14, 2026, issue of the journal Neurology.

Diagnostic modeling
In an interview with MSC, study author Erling Tronvik, Norwegian University of Science and Technology, Trondheim, Norway, said that two main questions motivated the new research.

“The first was whether there was something more than the headache attack itself that could be used to predict migraine. The second question was whether the machine learning method on its own could be used to group migraine or headache characteristics,” he said.

In a previous study from the same group, including corresponding author Anker Stubberud, the investigators used HUNT data to show that ML analysis of genome-wide genotype data outperformed polygenic risk scoring (PRS) in distinguishing those with migraine from headache-free controls; PRS sums all the risk alleles a person carries but does not capture interactions between genes, known as non-additive genetic effects. In another recent HUNT study, the team found that ML trained on a combination of non-headache clinical data and genetic data better predicted new-onset migraine compared to using either type of data alone.

Building upon that past work, the researchers embarked on their new ML investigation, using data from the second and third waves of the HUNT study. More than 43,000 people were included in the diagnostic ML models (almost 9,000 with migraine according to a modified version of the first and second ICHD, and roughly 34,000 headache-free controls).

The researchers trained and tested these diagnostic models using more than 5 dozen non-headache features in 20 categories, including sleep, exercise, musculoskeletal factors, and cardiovascular factors, to name just a few.

“The deliberate exclusion of headache characteristics from the predictor set is conceptually and methodologically important,” wrote Romozzi in the accompanying editorial. “Most ML studies in migraine have relied on headache-related features as predictors, an approach that is inherently circular. As these features define the diagnosis itself, models trained on them do not uncover underlying biology but rather reproduce existing diagnostic criteria.”

Using only the non-headache clinical features, the diagnostic model achieved an area under the curve (AUC) of 0.79. AUC is a metric that is used to assess how well an ML model performs, and in the current instance, specifically its ability to distinguish migraine from headache-free controls.

“An area under the curve of 0.5 means that you are predicting by chance – it’s like throwing a two-sided dice – and an area under the curve of 1.0 is a perfect separation between the groups that you’re looking at,” Tronvik said. So, “in machine learning terms, 0.79 is quite a good result.”

Using genotype data along with the non-headache clinical data only slightly improved the result (an AUC of 0.80). “This doesn’t necessarily mean that genetics is unimportant in causing migraine, just that, in the prediction model, it doesn’t add very much in addition to the clinical data,” according to Tronvik. To explain this outcome, the authors hypothesized in their paper that the clinical data may already capture much of the genetic risk.

Finally, the researchers identified the non-headache features that best predicted migraine. Age, sex-related features, neck pain, nausea, sleepiness, low systolic or high diastolic blood pressure, and good self-reported health and mental health emerged as the most predictive features.

Features for predicting migraine. The features are ordered from the most important (top) to least important (bottom). Age is the most important predictor, followed by neck pain, menstruation, and nausea. For each feature, the impact of that feature toward having or not having migraine is plotted for each participant as a single dot. Dots placed to the right of the vertical axis predict having migraine, whereas dots placed to the left of the vertical axis predict not having migraine. Redder color of the dots indicates a higher value (i.e., high age), while bluer color of the dots indicates a lower value (i.e., low age). As an example, high age (bright red dots to the left of the vertical axis) seems to be predictive of not having migraine, while the presence of neck pain, menstruation, or nausea (red dots to the right of the vertical axis) predicts having migraine. CONOR = Cohort of Norway; dBP = blood pressure; HADS = Hospital Anxiety and Depression Scale; HDL = high-density lipoprotein; sBP = systolic blood pressure. Credit: Image and caption from Danelakis et al. Neurology. 2026 Jul 14;107(1):e218076.

ML phenotypic modeling
The investigators next sought to determine if ML trained on headache data, as well as on the most predictive features from the diagnostic models, could enable identification of particular subgroups, or clusters. More than 12,000 people were included in this phenotyping analysis (approximately 4,400 with migraine and 7,700 with non-migraine headache).

Here, the ML identified two clusters. The first cluster included 1,425 people, 94% of whom had migraine. The second cluster comprised 10,760 people, 71% of whom had non-migraine headache.

The researchers performed additional analysis on the cluster of 1,425 individuals to see if subclusters could be identified. This approach was not successful using headache characteristics.

However, when the team used the 10 most important non-headache features from the diagnostic modeling, four distinct subclusters emerged. One of the subclusters included only men, while the other three subclusters included only women.

For the subclusters with women, one was characterized by neck pain, and the absence of general gastrointestinal symptoms, which the researchers suggest implicates a cervicogenic component for the women in this particular subcluster. The second subcluster of women featured general musculoskeletal pain, lower self-reported health, and higher anxiety and depression, which could reflect central sensitization and comorbid psychiatric disorders, according to the researchers. Finally, the third subcluster of women was characterized by lower anxiety and depression, and pulsating pain – this subcluster, the investigators wrote, most closely resembles “classic” migraine according to ICHD criteria, with little neuropsychiatric comorbidity.

Finally, when looking at the two main clusters – the 1,425 individuals who primarily had migraine, and the 10,760 people who mostly had non-migraine headache – the investigators saw no differences between groups according to their PRS.

But there were differences according to ML-based genetic risk scores, suggesting that the non-additive genetic effects that ML-based analysis is able to discern were playing an important role. The four subclusters also differed in their ML-based genetic risk scores.

In terms of study limitations, which the authors acknowledge, the clinical data used in the study is of an older vintage, having been collected from 1995 to 2008. In addition, the genotype platform the group used is older and not as comprehensive compared to newer platforms. And, the data from HUNT are health survey data, which rely on self report.

Katsuki told MSC that one open question is whether the findings would generalize to other cohorts, while another “is whether the identified subgroups are clinically meaningful. In particular, future studies should determine whether these groups differ in prognosis, chronification, disability, or treatment response.”

Similarly, “[t]he key question is no longer whether ML can identify migraine, but whether the subtypes it reveals correspond to true biological entities that can inform prognosis and guide treatment decisions,” Romozzi wrote in the accompanying editorial.

More – and better-linked – data
Tronvik told MSC that he and his colleagues continue to use ML to better understand migraine, in particular by linking other data sets to the HUNT data.

“We have data basically from the whole Norwegian population now, [including] medication data from the prescription registry, labor data, socioeconomic data, and diagnostic registers, and all of these [data sources] have also been linked to the HUNT data. So now we can actually start looking at maybe a 10- to 20- to 30-year perspective and look at dynamic changes over time in individual patients,” he said.

Having this volume of data bodes well for ML studies, “because machine learning has its main advantage when we look at large numbers,” according to Tronvik.

Speaking of the data, Tronvik added, “I think machine learning will definitely have a big place in the future, but what we lack at the moment is sufficient data, [and] international data, because machine learning is never better than the data that is put into it. So my hope for the future is that we are able, in the headache field, to collaborate better, both on harmonizing data and also on linking data so that we can use machine learning in a much more fruitful way,”

As far as the new study goes, Tronvik emphasized that it offers “support [for] moving away from a one-size-fits-all diagnosis towards data-driven phenotyping. And in the future, that could make us better at tailoring both treatments and diagnosis for individual patients.”

Along those lines, Tronvik said that he and his colleagues are now using ML in a clinical trial of 200 migraine patients to see whether the method can enable prediction of what medication will be best for an individual patient.

Neil Andrews is a science journalist and executive editor of the Migraine Science Collaborative.

Image credit: 123RF Stock Photo.

References
Machine diagnostics and machine phenotyping of migraine: A HUNT study.
Danelakis et al.
Neurology. 2026 Jul 14;107(1):e218076.

Is migraine a single entity? Machine learning reveals underlying heterogeneity.
Romozzi M.
Neurology. 2026 Jul 14;107(1):e218177.