AI Is Getting More Out of MS Data, but Clinical Use Lags Behind

Artificial intelligence (AI) has moved quickly into multiple sclerosis (MS) research, with new tools to segment lesions, flag paramagnetic rim lesions (PRLs), and extract prognostic signals from scans that clinicians already order. At MSToronto 2026, the 10th Joint ACTRIMS-ECTRIMS Meeting, a scientific session on AI applications from diagnosis to prognosis, along with poster presentations and interviews at the meeting, showed how far the tools have come and how much of the evidence needed to use them in practice is still missing.

Michael G. Dwyer, PhD, associate professor of neurology and biomedical informatics at the University at Buffalo, State University of New York, walked through the barriers that separate an accurate model from one clinicians can use. Arman Eshaghi, MD, PhD, a National Institute for Health and Care Research (NIHR) Advanced Fellow at the UCL Queen Square Institute of Neurology, described the shift from carefully curated research data to the scans produced in routine care.^1,2

"We have seen an explosion of tools recently that support both the improvement of the quality of the images [and] the analysis themselves," Alessandro Cagol, MD, of Translational Imaging in Neurology (ThINk) at the University of Basel in Switzerland, said in an interview at the meeting. “This is something that is really changing our practice.”

One of the Most and Least Tractable Diseases for AI

Dwyer described MS as one of the most tractable diseases for AI and, at the same time, one of the least. On the tractable side, MS generates repeated, quantitative measurements across MRI, optical coherence tomography (OCT), serum biomarkers, cognition, and digital outcomes, he noted. Diagnosis and monitoring rely on objective data. Trials, registries, and electronic health record (EHR) archives span decades, and treatment decisions are tied to measured disease activity.¹

On the other side, MS varies widely in its biology and clinical course, and its pathology is spread across the brain, spinal cord, optic pathway, and networks. The key outcomes, such as progression, cognition, and quality of life, are the least directly observed, he noted. Even the word "lesion," according to his presentation, covers multiple biological entities.¹

The labels that models learn from are another weak point. Dwyer cited data showing that 40% of new lesions were marked by 2 or fewer of 4 experts and that inter-rater agreement for PRL phenotyping ranged from an intraclass correlation coefficient of 0.71 to 0.85 in a study he cited.¹ In a registry analysis, only 17.3% of disability progression events were identified by all 48 definitions of confirmed progression that the investigators compared.³ Identifying secondary progressive MS carries about 2.9 years of uncertainty, according to Dwyer's presentation, and a separate registry study tested 576 candidate definitions of its onset.^1,4

1,600 Authorized Devices, Few Externally Validated MS Models

Those weaknesses have not kept AI tools from reaching regulators. More than 1,600 AI-enabled devices have FDA authorization, most of them in radiology, and at least 340 high-risk devices carry CE approval in Europe, according to Dwyer's presentation. Yet on a validation ladder that runs from cross-validation through held-out testing, external institutions, different protocols and populations, prospective deployment, and demonstrated clinical benefit, "most published MS models [are] at 1 or 2," he noted.¹

Dwyer also pointed to the small size of public MS imaging data sets, which have included 19 subjects in one benchmark challenge, 53 in another, and 100 scan pairs in a third. Most MS models are trained on single-center data, full release of code and data is rare, and large clinical registries are rarely linked to imaging, he noted. Population, scanner, and site imbalances can lead to biased models, and in MS, site and scanner are potential proxies that a model can learn instead of disease biology.¹

Adoption is another barrier. MS lesion-tracking tools found more disease activity, with new lesions in 30% vs 15% of patients and treatment changed in 11%, according to Dwyer's presentation, but AI training is uncommon.¹ In one survey he cited, 82.1% of 201 physicians in Saudi Arabia trusted AI-based recommendations in clinical decision-making, yet 25.9% had used AI in the previous year.⁵ Both under-use and over-reliance are problems, he noted. Dwyer's recommendations included requiring TRIPOD+AI or CLAIM reporting and external validation for MS AI publications, building shared multi-rater reference data sets for the 2024 McDonald criteria biomarkers, and building AI literacy into MS fellowships and continuing education.¹

What the Tools Can Find, and How Far They Have Been Tested

Even with those gaps, studies at the meeting showed what AI can draw from imaging that clinicians already collect. Eshaghi noted that routine care produces far more scans than research studies, citing a model trained on 5.24 million MRI and CT volumes from 20 years of routine care in a single US health system. He also pointed to a gap in what is automated. Spinal cord measures explain more of physical disability than brain measures, according to an unpublished meta-analysis he presented, yet automated tools have reached regulatory approval only for brain lesions and brain volume.²

Cagol and colleagues found that a deep learning model reading routine FLAIR scans of the thalamus added predictive value for disability worsening beyond thalamic volume alone. The finding held in an external cohort of 504 patients followed for a median of 7.6 years, the kind of outside test Dwyer said most MS models lack.⁶

PRLs, a marker of chronic active inflammation added to the 2024 McDonald criteria, are a newer target. In a study presented by Àlex Rovira, MD, of Vall d'Hebron University Hospital in Barcelona, Spain, a model detected them most reliably when given 3 MRI sequences rather than a single sequence.⁷ Federico Spagnolo, also of ThINk at the University of Basel, and colleagues found that starting from a model pretrained on MRI from more than 11,000 people improved classification when annotated examples were limited.⁸ Both were evaluated on held-out test sets, and neither reported testing at an outside center.

Beyond Imaging: Guidelines and Drug Discovery

Foundation models can help when annotated MS imaging is scarce. Clinicians interviewed at the meeting also saw uses for AI beyond imaging, particularly where evidence is thin. In rarer neuroimmunologic diseases, the shortage starts with the treatment evidence itself. Amanda Piquet, MD, director of the Autoimmune Neurology Program at the University of Colorado Anschutz, said the field lacks the treatment structure that MS has.

"Unlike MS, where we have 20-some FDA-approved therapies and people kind of have algorithms, I feel like autoimmune neurology is a little more of the wild, wild West," Piquet said in an interview at the meeting. "If AI can be used somehow to integrate and organize data to help us make guidelines, that's where I think it can be helpful."

Drug discovery is another possible use. Nick Cunniffe, PhD, a consultant neurologist at the Cambridge MS Clinic in the United Kingdom who leads clinical remyelination trials, said AI could screen large compound libraries for existing drugs that act on a target pathway, the approach behind repurposed-drug trials in MS. "It basically allows us to take shortcuts," Cunniffe said, though he added that "we don't really know where AI is going to be in the next five, 10 years."

Clinicians Remain in the Loop

Whatever form the tools take, Dwyer's closing recommendations placed responsibility on researchers, MS centers, and societies alike, from sharing data under consent to involving people with MS in defining AI goals, disclosures, and governance.¹ Gianmarco Bellucci, MD, a postdoctoral scholar at the University of California, San Francisco (UCSF), described AI as a tool that will become routine for clinicians without replacing them.

"I see it [as a] fundamental tool that any clinician will have to deal with in the future that will help us focus more on patients and less on administrative procedures," Bellucci said in an interview with NeurologyLive. "From a clinical standpoint, I still think that it's a long way for clinicians to be substituted by AI tools."

Cagol described a field in rapid motion. "It's changing really week after week," he said.

Click here for more coverage of MSToronto 2026.

REFERENCES
1. Dwyer MG. Overcoming barriers for AI and big data. Presented at: MSToronto 2026, the 10th Joint ACTRIMS-ECTRIMS Meeting; October 21-23, 2026; Toronto, Canada. Session S13.2.
2. Eshaghi A. Using MRI from clinical practice to inform prognosis. Presented at: MSToronto 2026, the 10th Joint ACTRIMS-ECTRIMS Meeting; October 21-23, 2026; Toronto, Canada. Session S13.1.
3. Kalincik T, Cutter G, Spelman T, et al. Defining reliable disability outcomes in multiple sclerosis. Brain. 2015;138(pt 11). doi:10.1093/brain/awv258
4. Lorscheider J, Buzzard K, Jokubaitis V, et al. Defining secondary progressive multiple sclerosis. Brain. 2016;139(pt 9):2395-2405.
5. Alhazmi F. Understanding physician attitudes toward AI in clinical decision-making: cross-sectional study. JMIR Form Res. 2025;9:e79730. doi:10.2196/79730
6. Cagol A, Ocampo-Pineda M, Montobbio N, et al. Deep learning-derived conventional MRI signatures of thalamic pathology add predictive value for disability progression in MS. Presented at: MSToronto 2026, the 10th Joint ACTRIMS-ECTRIMS Meeting; October 21-23, 2026; Toronto, Canada. Abstract P0173.
7. Rovira À, Hwang J, Won J, et al. Optimizing MRI input modalities for automated detection of paramagnetic rim lesions in MS. Presented at: MSToronto 2026, the 10th Joint ACTRIMS-ECTRIMS Meeting; October 21-23, 2026; Toronto, Canada. Abstract O048.
8. Spagnolo F, Gordaliza PM, Lu PJ, et al. Foundation model-enhanced classification of paramagnetic rim lesions in multiple sclerosis. Presented at: MSToronto 2026, the 10th Joint ACTRIMS-ECTRIMS Meeting; October 21-23, 2026; Toronto, Canada. Abstract P0331.