Clinical Measurements Plus Machine Learning Place Migraine on a Spectrum
By Kayt Sukel | August 12, 2026 | Posted in News
A new machine learning model trained on routine clinical data situates multiple diseases – including migraine – along a continuum. Model outputs were associated with a host of relevant health metrics.
Traditional medical conceptions of disease take a binary perspective: People either have a disease or they don’t. Yet anyone who lives with a chronic condition, whether it be migraine or diabetes, arthritis or depression, implicitly understands that diseases in the real world don’t operate via a simple on/off switch.
In fact, these conditions can develop gradually. People with the same diagnosis may have a very different experience of the disease. Symptoms fluctuate over time. Further, the dynamic nature of most chronic diseases can make it difficult for patients to find relief, especially before they receive an official diagnosis. It also poses challenges for physicians seeking to offer their patients the most effective treatments.
But now, a new study uses artificial intelligence (AI) to better capture the reality of how diseases manifest in the real world. Researchers including corresponding author Ron Do, Icahn School of Medicine at Mount Sinai, New York City, US, trained machine learning models on routine clinical data – the vital signs and laboratory tests typically captured during standard care. Then they used these models to develop a disease-spectrum score called SHIMMER (spectral health index from machine measurements of electronic records) for seven different major diseases, including migraine.
The approach worked: For all of the conditions considered in the study, SHIMMER was associated with many health metrics, including risk factors, biomarkers, disease course, disease severity and sequela, and treatment, successfully placing the conditions along a continuous disease spectrum.
This type of approach could help doctors detect disease earlier on and improve risk stratification, resulting in better patient outcomes.
Chia-Chun Chiang, a neurologist and headache expert at Mayo Clinic, Rochester, Minnesota, US, who is using AI to study migraine and vascular neurology but was not involved with the new research, called the study “innovative” and “[a better reflection of] how clinicians think about disease.”
“Rather than simply asking whether someone has a disease or not, SHIMMER estimates where a person falls along a spectrum using routine laboratory tests, vital signs, and basic demographic information that are already collected during everyday medical care,” she wrote in an email to Migraine Science Collaborative. “For migraine, this is particularly exciting because we do not have an objective biomarker. The finding that a score based on routine clinical data was associated with migraine severity, medication use, and age of onset suggests that these common measurements may contain useful information about underlying disease burden.”
The study appeared in the June 12, 2026, issue of Med, a Cell Press journal.
Using routine data and machine learning for a more sophisticated view of disease
The current study builds on the researchers’ previous work, which also used machine learning models trained on data from electronic health records (EHRs), to characterize the spectral nature of coronary artery disease (CAD) and autoimmune diseases.
For the new study, which would focus on a broader range of diseases, “we trained our models solely using a minimal set of routine clinical measurements that are widely available. This approach allows for ease of implementation and portability in health systems, addressing two shortcomings in the field that have hindered real-world usage of clinical machine learning,” the authors wrote in their paper.
Yohannes Woldeamanuel, a physician and researcher at Mayo Clinic, Phoenix, US, who is involved with AI development at his institution to help develop models that can diagnose migraine and classify different headache disorders but was not involved with the current study, said it is a compelling approach.
“We are sitting on a goldmine of data in the EHR that nobody bothers to look at … and AI could potentially help us look at these different data points, whether they are metabolic, inflammatory, or carcinogenic, and help us to understand the spectrum or trajectory of these different diseases,” he said in an interview with Migraine Science Collaborative.
The investigators used EHR data from almost 400,000 participants from two different datasets, including the BioMe Biobank and UK Biobank. The data included lab test results, vital signs, and anthropometric data. The group used diagnosis codes and other information to define disease cases and controls.
The group then trained machine learning models on the EHR data for seven diseases of interest, including atrial fibrillation, CAD, type 2 diabetes, rheumatoid arthritis, schizophrenia, breast cancer – and migraine. Subsequently, a SHIMMER score was generated; scores could range from 0, indicating minimal disease, to 1, which indicated maximal disease. Then, the question was whether the SHIMMER scores were associated with pertinent health metrics for each of the seven conditions.
That indeed turned out to be the case. For instance, for atrial fibrillation, increases in SHIMMER scores were associated with increases in BMI and age, two risk factors for the condition, in both the BioMe Biobank and UK Biobank cohorts. As another example, in the case of biomarkers for rheumatoid arthritis, the SHIMMER score tracked with changes in autoantibodies, inflammatory markers, and cell counts.
SHIMMER also tracked with disease course – for example, for all the diseases, there was a progressively earlier age of onset with increasing SHIMMER. SHIMMER also reflected disease severity and sequela, as well as the prevalence of treatment and medications. For the latter, in the case of CAD, higher SHIMMER scores were associated with a higher likelihood of using statins, aspirin, and nitrates, and undergoing surgical procedures like angioplasty or stenting.

Machine learning on a small set of routine clinical measurements that are ubiquitous in health systems generates SHIMMER, a digital marker that can profile an individual’s disease state for a variety of conditions. The bottom inset depicts a schematic illustration of how SHIMMER relates to disease-relevant biomarkers, complications, and survival along a continuous spectrum. The x-axis represents increasing SHIMMER values, while the y-axes represent biomarker magnitude, complication probability, and survival probability, respectively. Curve colors correspond to SHIMMER strata (blue, lowest strata; yellow, middle strata; red, highest strata) and match the color gradient shown above. This inset is intended as a conceptual visualization rather than an empirical data plot. CBC, complete blood count; CMP, comprehensive metabolic panel. Created with BioRender. Image and caption credit: Forrest et al. Med. 2026 Jun 12;7(6):101150.
What about migraine?
The study included many people with migraine, including more than 2,200 in the BioMe Biobank (8% of that cohort), and more than 18,500 in the UK Biobank (3.7% of that group) – and the general picture held: SHIMMER scores were associated with many relevant health metrics. As with the other conditions, SHIMMER was higher for diagnosed cases of migraine than for controls, and was also associated with migraine onset, severity, and treatment.
The performance of SHIMMER was not as good for migraine compared to some of the other diseases. But the authors wrote in their paper that this was expected, considering that migraine is a condition that has a “reliance on clinical history and episodic symptoms rather than robust routine laboratory biomarkers.” They added that this relatively weaker performance “[does] not detract from SHIMMER’s ability to capture meaningful disease gradients. SHIMMER produces continuous values between 0 and 1, and is intended as a disease-spectrum index rather than a calibrated estimate of absolute risk.”
A proof of concept, not a diagnosis tool
Chiang said one of the strengths of the SHIMMER model is its simplicity.
“Instead of relying on specialized imaging, genetic testing, or disease biomarkers, it uses routine laboratory tests, vital signs, and demographic information that are already available in most healthcare systems,” she said. “That makes the approach much easier to implement in real-world practice.”
Yet, both she and Woldeamanuel said that developing separate models for each disease is not ideal, considering the multimorbidity seen in the real world.
“I’ve never seen a patient with migraine who doesn’t have some other issue and, unfortunately, it’s rarely discussed outside of the headache clinic,” said Woldeamanuel. “They go to a different doctor or clinic for their joint pain or gastrointestinal issues. And we know there are extremely similar underlying mechanisms for many of these conditions.
An AI model, he said, could perhaps detect patterns in those similar underlying mechanisms, providing more insight into how such mechanisms influence a wide range of different chronic disease states.
In their paper, the authors point to the issue of multimorbidity, writing that future work “could include joint modeling of multiple SHIMMER scores, longitudinal trajectory analysis, and approaches to partition shared versus disease-specific components of risk, further enhancing interpretability in multimorbid populations.”
Importantly, the authors said that their work is a proof-of-concept study, and one that is not intended to supplant diagnosis.
“In future clinical settings,” they wrote, “we envision that SHIMMER may be integrated as a background decision-support layer within EHR systems, leveraging routine laboratory tests and vital signs already collected in standard care. Rather than serving as a diagnostic replacement, SHIMMER could enhance early detection and risk stratification by flagging individuals with elevated disease-specific SHIMMER for targeted evaluation.”
Chiang said she hopes future prospective studies will be able to demonstrate that incorporating SHIMMER into clinical practice will improve patient outcomes. In the meantime, she is encouraged that the study supports what she already sees in the clinic every day.
“This study supports the idea that disease exists on a continuum rather than as a simple yes-or-no diagnosis. Individuals without a formal diagnosis but with intermediate or high SHIMMER scores often showed characteristics consistent with early or under-recognized disease,” she said. “From a headache perspective, it is encouraging that routine clinical data alone carried meaningful information about migraine despite the absence of an established laboratory biomarker. That raises the possibility that commonly collected healthcare data may provide additional insights into migraine biology and disease burden.”
Kayt Sukel is a freelance writer based in Kansas.
Image credit: 123RF Stock Photo.
Reference
Capturing multi-disease states on a spectrum with machine learning and routine clinical data.
Forrest et al.
Med. 2026 Jun 12;7(6):101150.

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Sign Up For An MSC Newsletter
Share this article: