Medical Records and Retinal Images Revealed MS and NMOSD Years Before Diagnosis
Studies presented at MSToronto 2026 show that models trained on medical records, retinal images and blood tests flag multiple sclerosis and NMOSD long before clinical diagnosis — but the findings are retrospective, and no effect on outcomes has been demonstrated.
Machine learning models applied to electronic health records, retinal images and blood tests picked up signs of multiple sclerosis (MS) and neuromyelitis optica spectrum disorder (NMOSD) several years before diagnosis, according to studies presented at MSToronto 2026, the tenth joint ACTRIMS–ECTRIMS meeting. The coverage from NeurologyLive is dated October 28, 2026, and all figures in this article come from that secondary conference report — the conference abstracts and the studies themselves were not available for verification.
The Karolinska model: primary care data as a risk score
A team at Karolinska Institutet in Stockholm, led by Ali Manouchehrinia, trained a deep learning survival model — a so-called gated recurrent unit network — on quarterly primary care data from Stockholm covering the period 2010 to 2025. The model used a 30-quarter lookback window to predict a new MS diagnosis within 1 to 5 years. The model was then distilled into a simpler variant with what the presentation called "clinician-friendly decision logic."
The strongest predictors were healthcare utilization, examination history, sensory symptoms, dizziness, adjustment and stress disorders, and prescriptions for analgesics and psychoactive medications. Predicted five-year risk was higher among women and among people with a close relative with MS — a first- or second-degree relative.
"MS is often preceded by a measurable healthcare footprint, but diagnosis still comes only after a clinical threshold is crossed," Manouchehrinia said in the presentation, according to NeurologyLive.
Performance metrics for the Karolinska model are not given in the source material. That means it is currently impossible to assess how accurate the risk scores actually were.
The NMOSD prodrome: measurable six years before onset
A presentation by Bellucci and colleagues mapped the prodromal phase before NMOSD. "A measurable prodrome occurs before the first NMOSD attack: healthcare utilization increases roughly six years before onset; qualitatively, the phenotypic shift becomes significant 3–4 years before onset," Bellucci said in the presentation. In an interview with NeurologyLive, he said that AI tools will help optimize treatment and diagnostic workup.
The numbers behind it: healthcare contacts rose throughout the five years before diagnosis, with an adjusted rate ratio of 2.1 (95% CI 1.9–2.3) versus the general population and 5.0 (95% CI 4.5–5.5) versus healthy controls. In the final year, the ratios rose to 3.0 (95% CI 2.4–3.8) and 7.3 (95% CI 5.9–9.1).
A random forest model based on records data alone achieved an AUROC of 0.67 one year before diagnosis and 0.69 five years before — that is, discrimination far from perfect, but above pure chance. By distributing the records data into SPOKE, a biomedical knowledge graph, the AUROC rose by an average of around 12 percentage points. For distinguishing NMOSD from prodromal MS, the model achieved an AUROC of 0.87, versus 0.74 for records data alone. This is this publication's analysis: the SPOKE lift is substantial, but the AUROC levels for the model without the knowledge graph are modest, and the source material says nothing about prospective validation or planned clinical use.
What the models weighted most heavily
Before MS, the models weighted frequent healthcare visits and clusters of sensory, bladder and psychiatric symptoms most heavily, particularly among younger women. Before NMOSD, autoimmune conditions such as lupus and Sjögren's syndrome stood out, along with an increase in infections.
A third finding concerns already-diagnosed MS patients: changes in the retina preceded disability worsening by around nine years. This is the longest time horizon in the story, but it applies to disease course after diagnosis, not to the timing of diagnosis itself.
The caveats
All the findings are retrospective conference presentations, known through NeurologyLive's report rather than through primary sources. Retrospective prodrome findings show that the signals exist — not that earlier intervention changes outcomes. No prospective validation or clinical use of the models has been documented, and the source material provides no basis for assuming that earlier detection improves prognosis. The AUROC values for the NMOSD model without the knowledge graph (0.67–0.69) are modest, and the performance figures for the Karolinska model are entirely absent from the available source material.