New study identifies four unique migraine phenotypes using artificial intelligence

A recent study found that artificial intelligence algorithms can accurately identify people with migraines without relying on any information about their head pain. Published in the journal Neurology, the research suggests that migraines exist on a spectrum of distinct biological and clinical subtypes rather than as a single condition.

Migraine is a common neurological condition characterized by intense, pulsating head pain often accompanied by nausea and sensitivity to light and sound. The condition frequently causes severe disruptions to daily life, limiting a person’s ability to work or engage in regular activities. The medical field currently diagnoses the disorder based entirely on a patient’s self-reported symptoms during clinical visits.

There are no established biological markers, such as a blood test or a specialized brain scan, to definitively confirm a migraine diagnosis. Because symptoms and treatment responses vary widely among patients, medical professionals debate whether migraine is a single uniform disease. Some researchers suspect it is a spectrum of related syndromes.

These varying syndromes likely result from combinations of genetic vulnerabilities and environmental factors. Previous research into the genetics of migraine has identified many associated DNA variants. However, adding up these individual genetic risks only explains a small portion of how the condition is inherited.

This gap in understanding suggests that genetic factors might interact with each other and with environmental triggers in ways that simple statistical models cannot easily capture. Antonios Danelakis, a researcher at the Norwegian University of Science and Technology, led a team to explore these hidden patterns in patient data. They hypothesized that advanced artificial intelligence could capture the full biological footprint of the condition.

Free daily newsletter

To investigate this, the researchers conducted a large study using data from the Trøndelag Health Study in Norway. This massive public health initiative gathered extensive information through physical examinations, blood samples, and self-administered questionnaires. This initial analysis included 43,197 individuals who had provided genetic samples and completed the required health surveys. The participants were classified as either having migraines or being headache-free based on traditional medical criteria.

The team built predictive machine learning models to differentiate the migraine sufferers from the healthy controls. They trained these algorithms using genetic data and general clinical information. This clinical data included details about sleep patterns, cardiovascular health, mental health, socioeconomic status, and regular medication use.

The researchers intentionally excluded any data about the participants’ actual headaches from the computer models. They wanted to test whether the algorithms could detect a migraine diagnosis entirely through other biological and lifestyle markers. The algorithms were trained on one portion of the participants and then evaluated on a separate, unseen group to verify their accuracy.

A specific type of machine learning algorithm performed the best on the test group. The algorithm successfully identified individuals with migraines with a high degree of accuracy. This success indicated that the condition is associated with a wide constellation of clinical and biological markers outside of the head.

The researchers analyzed the algorithm to see which features were most predictive of a migraine diagnosis. The top factors included age, biological sex, neck pain, and nausea. Other highly predictive factors were sleepiness, blood pressure variations, and self-reported mental health status.

Adding genetic data to the clinical data provided only a marginal improvement in the algorithm’s accuracy. The researchers noted that rich clinical data likely already captures the physical manifestations of a person’s genetic risk. Because the clinical data reflects the downstream effects of DNA, the genetic information itself became slightly redundant for diagnostic purposes.

In the second phase of the research, the team conducted another large study focusing on 12,185 individuals who reported experiencing headaches. This time, they used unsupervised machine learning models. Instead of giving the algorithm traditional medical definitions to look for, they allowed the computer to find natural groupings in the data.

The researchers fed the algorithm the participants’ specific headache characteristics, along with the most predictive general health features identified in the first analysis. They used clustering algorithms to see how the patients grouped together mathematically based on shared data points. To better understand these mathematical groupings, the team used an advanced dimensional reduction technique that translated the highly complex data into two-dimensional visual maps.

The unsupervised models divided the participants into two main clusters. The first cluster contained 1,425 individuals, and the algorithm grouped them based on shared traits. When the researchers checked this group against traditional medical guidelines, they found that nearly all of these individuals met the official criteria for a migraine diagnosis.

The second cluster contained 10,760 individuals. Most of the people in this larger group (71%) had non-migraine headaches. This division demonstrated that the artificial intelligence could naturally separate migraine sufferers from other headache patients without being explicitly taught the medical rules.

The researchers then instructed the algorithm to break the migraine cluster down into smaller subgroups. They based this subclustering on the general health features like neck pain and mental health. This process revealed four distinct subtypes of migraine sufferers.

The first subtype consisted entirely of men. The second subtype was characterized by prominent neck pain and an absence of general stomach issues, possibly representing patients whose migraines have a strong connection to neck tension.

The third subtype featured widespread muscle and bone pain, combined with lower self-reported health and higher levels of anxiety and depression. The researchers suggested this group might represent a phenotype where the nervous system is highly reactive to pain. The fourth subtype fit the classic picture of a pulsating migraine but lacked the psychological and widespread physical pain symptoms seen in the third group.

To analyze genetic profiles, the team conducted genome-wide association studies comparing the data-driven migraine group with those labeled as having migraine by traditional criteria. This process scanned genetic variants across tens of thousands of participants to find DNA patterns linked to each group, and found a high degree of overlap in the identified risk loci. When the team compared genetic risk scores across the four subgroups, the machine learning-based genetic risk scores were distinctly different across the subclusters, supporting the idea that these clinical subgroups reflect genuine biological differences.

The study relied on self-reported questionnaires to determine the participants’ initial medical status. Self-reporting can lead to misclassification if participants misremember their symptoms or misunderstand the questions. For instance, people with infrequent migraines might have incorrectly answered negatively to questions about having headaches within the past year, falsely placing them in the healthy control group.

The clinical data was collected between 1995 and 2008. Lifestyle behaviors, mental health trends, and typical sleep patterns have likely shifted since that time, meaning the data might not perfectly reflect a modern population. The genotyping technology used during those years was also less comprehensive than modern laboratory platforms.

The identified relationships between clinical features and migraine subtypes are correlations. The findings do not mean that these general health features directly cause migraines, only that they tend to occur together in specific patterns. The researchers noted that these models need to be validated in entirely different populations before being used in standard medical practice.

Future research should investigate whether these newly identified migraine subtypes respond differently to specific medications. A deeper understanding of these categories could help doctors prescribe targeted treatments rather than relying on a trial-and-error approach. Scientists could also use wearable devices and smartphone applications to track patient symptoms in real time, providing even more detailed data for future algorithms.

The study, “Machine Diagnostics and Machine Phenotyping of Migraine: A HUNT Study,” was authored by Antonios Danelakis, Håkon Kvisle Abildsnes, Fahim Faisal, Marte-Helene Bjørk, Dominic Giles, Knut Hagen, Tjaṧa Kumelj, Manjit Matharu, Parashkev Nachev, Erling Tronvik, Bendik S. Winsvold, and Anker Stubberud.