AI and a blood test may predict multiple sclerosis course better than symptom labels

The questions arrive quickly after a diagnosis of multiple sclerosis: How fast will this progress? Will I lose my mobility? Should I start the strongest drugs now, or wait?

For decades, neurologists have had few concrete answers. They could describe a patient’s disease as “relapsing-remitting” or “progressive,” but those labels—based largely on symptoms—offered only rough clues about the future.

Now a team led by University College London says artificial intelligence and a simple blood test can do better.

A new way to sort MS

In a study published Dec. 2 in the journal Brain, researchers used machine learning to sift through MRI scans and blood samples from 634 people with MS. The algorithm identified two distinct biological types of the disease that cut across the textbook categories and behaved very differently over time.

“MS is not one disease and current subtypes fail to describe the underlying tissue changes, which we need to know to treat it,” said Dr. Arman Eshaghi, a neurologist and imaging scientist at UCL and co-founder of the spinout company Queen Square Analytics, in a statement about the work.

By combining brain imaging with a blood marker of nerve damage, he said, “we have been able to show two clear biological patterns of MS for the first time. This will help clinicians understand where a person sits on the disease pathway and who may need closer monitoring or earlier, targeted treatment.”

The findings arrive as neurologists seek more precise ways to manage a condition that affects an estimated 2.8 million people worldwide, many of them young adults. Multiple sclerosis, an autoimmune disease in which the immune system attacks the brain and spinal cord, can follow wildly different courses from one person to the next. Some remain relatively stable for years; others experience rapid loss of function despite treatment.

Historically, doctors have classified MS into clinically isolated syndrome, relapsing-remitting MS, secondary progressive MS and primary progressive MS, based on when symptoms appear and how they evolve. Those definitions were formalized in expert consensus guidelines in the 1990s and updated in 2014.

But the new research—and an accompanying commentary in Brain—argue that these categories do not capture the underlying biology.

Traditional labels are “a clinical scaffold built on overt symptomatology — inherently retrospective and poorly equipped to capture the latent biological diversity among patients,” wrote Drs. Julia Brummer and Verena Fleischer in the commentary, calling the new work “among the first to integrate fluid biomarkers with imaging data in a sufficiently generalizable and clinically interpretable manner.”

How the model works

At the heart of the study is an AI model known as SuStaIn, short for Subtype and Stage Inference. Developed originally to study neurodegenerative diseases such as Alzheimer’s, SuStaIn is an unsupervised algorithm: it is not told in advance which patients have which type of MS. Instead, it looks for patterns in how different biological features become abnormal as disease advances.

For this analysis, the team trained the model on detailed MRI measures and blood samples from 189 people with relapsing-remitting or secondary progressive MS enrolled in clinical trials sponsored by German drugmaker Merck KGaA. The scans measured, among other things, the volume of white matter lesions, the integrity of the corpus callosum—a major fiber bundle connecting the brain’s hemispheres—and the size of deep gray matter structures and regions of the cortex involved in memory and emotion.

The blood samples were tested for serum neurofilament light chain (sNfL), a protein released into cerebrospinal fluid and blood when nerve fibers are damaged. Over the past decade, sNfL has emerged as a general marker of neuro-axonal injury. In MS, higher levels have been linked to new inflammatory lesions on MRI, clinical relapses, faster brain atrophy and worse disability. Levels typically fall when disease-modifying drugs are working.

Using these combined data, the AI model learned two recurring trajectories of disease.

Two biological patterns

The “early-sNfL” subtype

In what the researchers called the early-sNfL subtype, patients showed a sharp rise in sNfL levels early in the course of their disease. At the same time, their MRIs revealed rapid accumulation of lesions and early damage to the corpus callosum.

These patients had more active inflammation and, on average, a more aggressive course, with a 2.44-fold higher risk of developing new contrast-enhancing lesions compared with the other group.

The “late-sNfL” subtype

In the late-sNfL subtype, the earliest changes were structural shrinkage in the limbic cortex and deep gray matter—regions associated with emotion and cognition—while sNfL remained relatively low. The biomarker rose only later, suggesting a slower, more insidious pattern of neurodegeneration with less overt inflammatory activity.

When the team applied the trained model to an independent cohort of 445 people with newly diagnosed MS or a first demyelinating episode, it again separated patients into these two patterns, suggesting the subtypes are reproducible and identifiable early.

What it could mean for prognosis and treatment

Adding the blood marker improved the model’s ability to track clinical disability. In the training set, the correlation between the AI-derived disease stage and scores on the Expanded Disability Status Scale (EDSS), a standard MS disability measure, nearly doubled when sNfL was included alongside MRI, rising from 0.231 to 0.420. In the validation cohort, the correlation was weaker but still improved when sNfL was used.

Importantly for treatment decisions, the two biological types responded differently to therapy. People with the early-sNfL pattern saw larger drops in active lesion counts and sNfL levels over time when treated, even as their brains shrank faster overall. Those in the late-sNfL group had lower inflammatory activity to begin with and more gradual changes.

The work adds to a growing body of research suggesting MS should be stratified by underlying mechanisms rather than only by symptoms and relapse patterns.

“This research adds to growing evidence supporting a move away from the existing descriptors of MS (like ‘relapsing’ and ‘progressive’) and towards terms that reflect the underlying biology of the condition,” said Caitlin Astbury, senior research communications manager at the MS Society in the United Kingdom. “This could help identify people at an increased risk of progression — and allow people to be offered more personalised treatment.”

Caveats and conflicts to navigate

Yet the study also highlights challenges as AI and biomarkers move closer to the clinic.

All of the data came from phase II and III drug trials, which typically exclude people with significant other illnesses and those with very advanced disability. Primary progressive MS—a form characterized by steady worsening from onset—was underrepresented. Experts say the subtypes will need to be tested in broader, real-world populations, including more diverse ethnic groups, before they can be widely applied.

The underlying imaging and clinical data are also proprietary to Merck, limiting independent replication. The study was funded by the company, and several co-authors are Merck employees. Others, including Eshaghi and co-authors Prof. Frederik Barkhof and Prof. Daniel C. Alexander, hold equity in Queen Square Analytics, which develops AI tools for analyzing brain scans and stands to gain commercially if such models become standard in trials or clinical practice.

The authors disclosed these relationships in the journal and said the funder had input on study design and data interpretation. There is no evidence the funding altered the results, but specialists say the commercial incentives underscore the need for transparent validation and regulatory oversight as decision-support tools emerge.

Even if the science holds up, many hospitals do not yet have the infrastructure to run complex AI models or access to standardized sNfL testing. While automated immunoassays for sNfL have recently entered the clinical market, there is not yet a universally accepted cut-off to define “high” levels, and results can be influenced by age and other conditions.

Advocates caution that advanced imaging and biomarker pipelines could deepen disparities if only well-resourced centers can use them to triage patients to high-cost therapies.

What comes next

For patients, the promise is earlier, more tailored information. A person newly diagnosed might one day be told not just that they have relapsing-remitting MS, but that their disease follows an early-sNfL pattern that typically demands swift, aggressive treatment—or a late-sNfL course that progresses more quietly and may call for different strategies.

For now, neurologists say, AI-derived subtypes are best viewed as an additional layer of evidence rather than a replacement for clinical judgment.

The new work suggests there is more than one way MS unfolds in the brain—and that those invisible pathways can be mapped. Whether those maps will change everyday care will depend on how quickly the models can be validated outside of drug trials, integrated into health systems and made available beyond the few centers at the forefront of precision neurology.