News|Articles|October 28, 2026

Machine Learning Spots Early Signs of MS and NMOSD Years in Advance

At MSToronto 2026, investigators applied machine learning to health records, retinal scans, and blood samples to flag multiple sclerosis and neuromyelitis optica spectrum disorder years before diagnosis or disability worsening, and identified the symptoms and conditions that tend to come first.

Machine learning models applied to electronic health records (EHRs), retinal imaging, and blood samples picked up signs of multiple sclerosis (MS) and neuromyelitis optica spectrum disorder (NMOSD) years before diagnosis, according to studies presented at MSToronto 2026, the 10th Joint ACTRIMS-ECTRIMS Meeting. Much of the work focused on the prodrome, the period of nonspecific symptoms and rising health care use that can precede a diagnosis by years, and on whether algorithms can find the people in that window who warrant a closer look.

"MS is often preceded by a measurable healthcare footprint, but diagnosis still begins after a clinical threshold is crossed," Ali Manouchehrinia, PhD, of Karolinska Institutet in Stockholm, Sweden, noted in his presentation.¹

The studies also point to what tends to come first. Before MS, the models gave the most weight to frequent health care visits and to clusters of sensory, bladder, and psychiatric symptoms, particularly in younger women.^1,2,3 Before NMOSD, autoimmune conditions such as lupus and Sjögren syndrome and a rise in infections stood out.⁴ In people already diagnosed with MS, changes in the retina preceded disability worsening by about 9 years.⁵

AI Helped Define an NMOSD Prodrome

Only 2 earlier studies had examined the period before an NMOSD diagnosis, according to the presentation by Gianmarco Bellucci, MD, a postdoctoral scholar at the University of California, San Francisco (UCSF).⁴ Bellucci and Sergio E. Baranzini, PhD, used AI at 2 stages to look for that window in the UCSF EHR, which covers 7 million patients. First, an agentic AI analysis followed by clinical adjudication identified people who later developed NMOSD.⁴ The resulting cohort of 317 people, including 102 with at least 5 years of records before diagnosis, was matched 1:100 to the general population and to immunologically healthy controls, with comparison groups of people with prediagnostic MS (n = 3048), neuro-lupus (n = 685), neuro-Sjögren (n = 369), and neurosarcoidosis (n = 76).⁶

Health care encounters rose across the 5 years before diagnosis, with an adjusted rate ratio of 2.1 (95% CI, 1.9-2.3) vs the general population and 5.0 (95% CI, 4.5-5.5) vs healthy controls. In the final year, the ratios reached 3.0 (95% CI, 2.4-3.8) and 7.3 (95% CI, 5.9-9.1), respectively.⁶

"A measurable prodrome precedes the first NMOSD attack: healthcare utilization rises about 6 years before onset; qualitatively, the phenome shift becomes significant 3-4 years before onset," Bellucci noted in his presentation.⁴

Five years before diagnosis, 153 conditions, 123 drugs, and 49 abnormal laboratory results were enriched compared with healthy controls, including autoimmune diseases, vitamin D deficiency, C-reactive protein, and immune-cell abnormalities. Compared with people who went on to develop MS, those who developed NMOSD had more autoimmune comorbidity, including thyroiditis, lupus, and Sjögren syndrome, along with area postrema and visual features.⁶ Bellucci's presentation identified lupus and Sjögren syndrome as core comorbidities preceding NMOSD and pointed to a peak in infectious-disease encounters before diagnosis, with suggestive signals for cytomegalovirus.⁴

The second use of AI came in prediction. A random-forest model using EHR data alone reached an area under the receiver operating characteristic curve (AUROC) of 0.67 one year before diagnosis and 0.69 five years before. Embedding the EHR data in SPOKE, a biomedical knowledge graph, raised the AUROC by about 12 points on average and distinguished NMOSD from prodromal MS with an AUROC of 0.87, compared with 0.74 for EHR data alone.⁶

"[AI tools will] help us grasp more rapidly a bigger amount of information that will eventually help us in optimizing the treatment or the diagnostic workups for our patients," Bellucci said in an interview with NeurologyLive.

A Deep Learning Model Turns Primary Care Records Into MS Risk Scores

Manouchehrinia and colleagues set out to turn the MS prodrome into individual risk estimates. They trained a deep learning survival model, a gated recurrent unit network, on quarterly Stockholm general practitioner records from 2010 to 2025, using a 30-quarter lookback window to predict incident MS within 1 to 5 years.² They then distilled it into a simpler model with what the presentation called clinician-friendly decision logic.¹

In a development cohort of 122,451 individuals with 2401 incident MS cases, C-statistics ranged from 0.80 at 1 year to 0.73 at 5 years in internal testing, and from 0.75 to 0.73 in an independent cohort from the Swedish National Patient Register.² In held-out test data with 472 future MS cases and 531,438 controls, referring the top-ranked 500 people per year for evaluation would have identified 20 future cases (4.2%), and referring 5000 per year would have identified 71 (15.0%), according to Manouchehrinia's presentation.¹

What Tends to Come First

The models also showed which signals carried the most weight. Important predictors in Manouchehrinia's model included health care use, investigation history, sensory symptoms, dizziness, adjustment and stress disorders, and prescriptions for analgesics and psychoactive medications. Predicted 5-year risk was higher in women and in people with a first- or second-degree relative with MS.²

"Model priority rises most clearly when prodromal symptom groups are present, especially in younger women with multiple active symptom groups," Manouchehrinia noted in his presentation.¹ Those groups were sensory, visual, and balance disturbances; pain, headache, and musculoskeletal symptoms; psychiatric symptoms, fatigue, and sleep disturbances; urinary and bladder dysfunction; gastrointestinal and bowel symptoms; and injuries, falls, and instability.¹

In a second presentation, Manouchehrinia described an interpretable machine learning approach that mapped how diagnoses and treatments connect over time in Swedish registry data from 56,548 individuals. Patterns enriched before an MS diagnosis included neurologic abnormalities and movement disorders, along with prescriptions for baclofen, tolterodine, oxybutynin, and other anticholinergic drugs, reflecting symptom management and bladder dysfunction.³

The early picture may also differ by race. In a US case-control study drawn from the TriNetX US Collaborative Network, Christopher Orlando, of the Keck School of Medicine of the University of Southern California, and colleagues matched 30,400 people with MS 1:1 with controls. Musculoskeletal conditions were more common 2 to 5 years (odds ratio [OR], 2.5) and 5 to 10 years (OR, 1.9) before diagnosis. In the 2 to 5 years before diagnosis, they were recorded in 53% of Black individuals and 52% of Asian individuals compared with 45% of White individuals (P <.001), and genitourinary conditions in 43% of Black individuals compared with 33% of White individuals (P <.001). The investigators noted that these patterns "may help inform future early detection strategies."⁷

The Retina Changes About 9 Years Before Disability Worsens

Early signals are not limited to the years before diagnosis. Giacomo Boffa, MD, of the University of Genoa in Italy, and colleagues applied a Gaussian process progression model, a data-driven machine learning method that reconstructs the order of disease events, to longitudinal optical coherence tomography (OCT) and Expanded Disability Status Scale (EDSS) data from 584 patients with relapsing-remitting MS, and validated it in an independent cohort of 210 patients.⁵

The peripapillary retinal nerve fiber layer (pRNFL) reached its peak rate of change earliest, at a mean disease duration of 4.5 years, followed by the ganglion cell-inner plexiform layer at 11.1 years and the EDSS at 13.3 years. Maximal pRNFL change preceded EDSS worsening by about 9 years, and the external cohort reproduced the same order.⁵ "The peak rate of pRNFL change occurred substantially earlier along the reconstructed disease trajectory than the peak worsening of EDSS," Boffa noted in his presentation.

Blood Shows Biological Aging Before Symptoms

Blood-based measures also picked up change ahead of clinical signs. In 2 proteomic data sets that included 117 samples collected before symptom onset (median lead time, 4.0 years), Pavan Bhargava, MBBS, of Johns Hopkins University in Baltimore, Maryland, and colleagues found systemic (β = 1.0; 95% CI, 0.4-1.7; P = .002) and brain (β = 2.4; 95% CI, 1.2-3.7; P <.001) proteomic age acceleration before onset, with no acceleration in immune age. Larger systemic and muscle age gaps were associated with greater disease severity.⁸

In a separate study, Jakob Siljebo, of Uppsala University in Sweden, and colleagues applied a convolutional neural network to serum metabolomics from 938 people with newly diagnosed MS and 954 matched controls. The model distinguished MS from controls with an AUC of 0.99, but metabolomic age acceleration was not detectable at diagnosis (P = .515), which the investigators noted suggests that aging-related metabolic changes may emerge later in the disease course.⁹

From Signal to Screening

None of these tools is ready for clinical use. "The next step is not yet clinical deployment, but prospective utility testing," Manouchehrinia noted in his presentation.¹ Bellucci listed validation in external cohorts with large-scale EHR data as the next step for the NMOSD work.⁴

Bellucci said the gains will depend on clinicians who understand the tools. "There will need to be a fruitful interaction between experts, clinicians who are aware of what they can do with AI tools, what they cannot do, what they cannot ask, and this type of tool, in order to improve the patient quality of life in general," he said.

Click here for more coverage of MSToronto 2026.