AI in MS, NMOSD, autoimmune encephalitis and MOGAD: promise still outruns proof

Artificial intelligence is being evaluated across nearly every corner of central nervous system autoimmune medicine, from automated MRI lesion segmentation in multiple sclerosis to relapse-risk prediction in neuromyelitis optica spectrum disorder and outcome modelling in autoimmune encephalitis. But a comprehensive review published in EMBO Molecular Medicine concludes that, despite a rapidly growing literature, the field has yet to deliver a single prospectively validated model for individual treatment selection in any of the four major CNS autoimmune diseases it examines. The review, led by Zijian Cao and Fu-Dong Shi with colleagues, maps where computational methods have genuinely advanced patient care and where enthusiasm has outpaced evidence.

The four conditions at the centre of the analysis—multiple sclerosis, neuromyelitis optica spectrum disorder, autoimmune encephalitis, and myelin oligodendrocyte glycoprotein antibody-associated disease, or MOGAD—share a stubborn clinical problem. Their symptoms overlap, their biomarkers are imperfect, and their courses unfold over years, demanding longitudinal observation rather than single snapshots. Misdiagnosis, delayed recognition, uncertain prognosis and difficulty choosing or adjusting treatment remain common. Because relevant signals are scattered across MRI, optical coherence tomography, electrophysiology, serum and cerebrospinal fluid biomarkers, treatment histories and long-term outcomes, the authors argue that managing these diseases has become a multimodal, longitudinal inference problem rather than a single-biomarker classification task—precisely the kind of problem machine learning is built for.

The evidence, however, is strikingly uneven. Multiple sclerosis dominates the published record, particularly in neuroimaging, where early applications focused on automated lesion segmentation and MRI-based monitoring. Some segmentation and disease-monitoring tools have undergone external, multicentre or prospective workflow evaluation, yet the review stresses that evidence their use actually improves treatment decisions or patient outcomes remains limited. By contrast, studies in NMOSD, autoimmune encephalitis and MOGAD are smaller, more heterogeneous, and frequently lack external validation. A meta-analysis of MRI-based studies in NMOSD reported pooled accuracy of roughly 82 percent, with sensitivity of about 83 percent and specificity of about 80 percent, for distinguishing NMOSD from multiple sclerosis—but many cohorts emphasise established aquaporin-4 antibody-positive disease and typical MS comparators, leaving seronegative and diagnostically uncertain patients poorly represented.

The methodological critique is blunt. Systematic appraisals cited in the review show that high reported performance is often accompanied by design features that inflate the risk of bias. Spectrum bias arises when models are trained on well-characterised cases and healthy controls rather than patients presenting with real-world diagnostic uncertainty. Data leakage occurs when information from the same patient, examination, site or preprocessing step influences both model development and evaluation—a particular hazard when multiple lesions, scans or visits from one patient are treated as independent observations. Small rare-disease cohorts, high-dimensional radiomic features and absent independent validation amplify overfitting, while annotation variability means that even lesion boundaries and diagnostic labels depend on reader judgment. Random internal splitting, the authors note, does not establish transportability across scanners, centres, demographic groups or healthcare systems.

To discipline interpretation, the review proposes a six-level molecular evidence ladder that separates what computational studies actually demonstrate from what they merely suggest. Level one is statistical association: transcriptomic classifiers and metabolomic profiles that identify molecular features linked to disease stage or severity, such as stage-specific blood-transcriptome signatures and reduced hippurate in multiple sclerosis. Level two is causal prioritisation, where Mendelian randomisation and Bayesian colocalisation strengthen target nomination—EVI5, OGA, TNFRSF14, AGER, CD58 and TNF in MS reach this level—without establishing biological mechanism. Level three is mechanistic hypothesis generation from single-cell, spatial and ligand-receptor analyses. Levels four through six demand experimental perturbation, druggability assessment and, ultimately, demonstrated therapeutic efficacy. The verdict is unambiguous: no candidate discussed in the review has demonstrated therapeutic efficacy in CNS autoimmune disease on the basis of computational evidence alone.

The same caution applies to clinical prediction. Most available models estimate prognosis under observed care rather than the comparative effects of alternative treatments, and the review insists that high predicted risk does not establish that any particular therapy will improve an individual patient’s outcome. Prognostic models, it argues, should not be interpreted as validated tools for treatment selection. Treatment-effect estimation instead requires explicit causal methods—target-trial emulation, causal forests, careful handling of confounding by indication—and prospective interventional evaluation before model outputs can guide individualised therapy. Even in NMOSD, where a multicentre deep-survival model improved relapse-risk prediction over conventional Cox modelling in antibody-positive disease, the authors caution that treatment-related predictors within such models should not be read as evidence of treatment-effect heterogeneity.

Disease-specific gaps sharpen the picture. In MOGAD, the MOG-AR relapse-risk score was developed and internally validated in a registry-based multicentre Chinese cohort, but external validation in a 284-patient UK cohort showed only modest discrimination for three-year relapse risk, with an area under the receiver operating characteristic curve of 0.67 for the continuous score and 0.58 for the corresponding risk grade, alongside overestimation of relapse probabilities. In autoimmune encephalitis, evidence is geographically concentrated and heterogeneous in antibody subtype and outcome definition, EEG-based models remain underdeveloped, and no validated model estimates the comparative benefit of second-line immunotherapy after inadequate first-line response. Across all four diseases, diagnostic and prognostic applications outnumber comparative treatment-effect studies, and recovery-oriented modelling in MOGAD lags behind relapse prediction despite a clinically important subset of patients developing incomplete or severe residual disability after attacks.

Some of the most forward-looking work concerns infrastructure rather than algorithms. Because these diseases are rare, multicentre data pooling is essential, yet privacy regulation—including China’s Personal Information Protection Law and the European General Data Protection Regulation—constrains data sharing. Federated learning, which trains models across institutions without centralising raw patient data, has direct research evidence in multiple sclerosis, including federated lesion segmentation on multicentre clinical datasets and personalised federated learning for two-year disability-progression prediction. These studies demonstrate research feasibility rather than routine clinical implementation, and the review identified no disease-specific federated-learning study in NMOSD, MOGAD or autoimmune encephalitis, making application to those rarer conditions forward-looking. The authors also warn against importing evidence from adjacent fields: the BraTS initiative proves federated feasibility in brain-tumour segmentation, but should not be treated as direct neuroimmunology evidence.

Foundation models and large language models receive a similarly measured treatment. A single-centre retrospective study evaluated an LLM for antibody-positive autoimmune encephalitis but lacked external or prospective validation, and the review found no prospective clinical validation of foundation models or medical LLMs anywhere in CNS autoimmune disease. Domain shift, rare-disease under-representation, hallucination, poor calibration and computational burden may all limit reliability. Agentic and retrieval-augmented systems, the authors suggest, should in the near term be confined to source-grounded evidence retrieval and hypothesis prioritisation rather than autonomous diagnosis or treatment selection, with outputs traceable to supporting sources, version-controlled and reviewable by clinicians.

The review’s overall conclusion is that current evidence supports assistive research and selected workflow-evaluated applications, mainly in MS imaging, but not autonomous diagnosis or treatment selection. Near-term value, it argues, is more plausibly found in multimodal evidence integration, standardised image and biomarker assessment, risk stratification, molecular hypothesis prioritisation and the evaluation of candidate trial-enrichment strategies and outcome measures—such as quantitative MRI composites that correlate with cognitive and physical disability, though correlation alone does not establish surrogacy. Translation into routine care will require harmonised multicentre evidence, equitable evaluation across ancestry, sex, age and geography, and lifecycle governance including post-deployment surveillance for performance drift and subgroup safety. Until those conditions are met, the authors write, claims of direct clinical benefit should remain proportionate to the level of validation achieved—a sober benchmark for a field whose published performance metrics so often outrun its proven impact on patients.

Subject of Research: Applications and validation status of artificial intelligence in multiple sclerosis, NMOSD, autoimmune encephalitis and MOGAD

Article Title: Artificial intelligence in central nervous system autoimmune diseases: from clinical decision support to mechanistic discovery and translation

Article References: Cao, Z., Zheng, P., Qi, Z., Yao, Y., Xie, D., He, Y., Zong, S., Zhang, Z., & Shi, F.-D. (2026). Artificial intelligence in central nervous system autoimmune diseases: from clinical decision support to mechanistic discovery and translation. EMBO Molecular Medicine. https://doi.org/10.1038/s44321-026-00527-0

Image Credits: AI Generated

DOI: 10.1038/s44321-026-00527-0

Keywords: artificial intelligence, multiple sclerosis, NMOSD, autoimmune encephalitis, MOGAD, machine learning, MRI, federated learning, biomarkers, clinical decision support, neuroimmunology, drug discovery