AI Shows Promise for Multiple Sclerosis Diagnosis, Review Finds Validation Gaps

Artificial intelligence models can diagnose multiple sclerosis with striking accuracy when fed medical images and genetic data, but most of the studies behind those numbers are too small and too weakly validated to reach the clinic, according to a systematic review published in Frontiers in Medicine. The review, conducted under PRISMA guidelines, sifted 2,462 records from PubMed, Scopus, Web of Science and IEEE Xplore down to 26 peer-reviewed studies and assessed how machine learning and deep learning systems use MRI scans, retinal OCT imaging, blood transcriptomics, SNP genotyping and multimodal clinical datasets to detect and predict the disease.

The headline numbers are eye-catching. Among the included studies, some AI models exceeded 95 percent diagnostic accuracy using imaging and genomic data. An OCT study using swept-source imaging reported 97.2 percent accuracy with an area under the curve of 0.995, while a machine learning model combining transcriptomic and phenotypic features reached 97 percent classification accuracy. A deep learning model analyzing retinal vessel texture achieved 98 percent accuracy on its dataset. Yet the review’s authors repeatedly caution that these figures come largely from small, single-center cohorts validated only internally, meaning the reported performance reflects research-level optimization rather than clinically deployable systems.

The modality breakdown reveals where the field is heading. Optical coherence tomography dominated the included literature with 13 studies, reflecting the appeal of retinal imaging as a non-invasive, inexpensive window onto neuroaxonal loss. OCT delivers highly consistent, low-dimensional data, and models built on features such as retinal nerve fiber layer thinning and ganglion cell and inner plexiform layer thickness regularly surpassed 85 percent accuracy. Genomic studies accounted for six papers, applying random forests, gradient boosting and ensemble methods to high-dimensional transcriptomic and SNP data. Only three MRI-based studies made the cut, a scarcity the authors attribute to data complexity, acquisition costs and interoperability problems across centers. Four further studies used multimodal or clinical data.

Each modality carries distinct strengths and failure modes. Deep learning architectures such as ResNet-50 and U-Net excel at extracting spatial patterns from MRI, detecting gray matter atrophy and lesions that handcrafted radiomic features can miss, but their performance is hampered by small datasets, variability in acquisition protocols and reliance on internal cross-validation. In one MRI study, a random forest model achieved an AUC of 0.90 with n = 116 participants, while a ResNet-50 analyzing 3D T1-weighted MRI in 181 patients produced AUCs ranging from 0.69 to 0.81 depending on the imaging plane. Genomic models, by contrast, wrestle with polygenicity and the mismatch between feature dimensionality and sample size, which inflates overfitting risk and population bias. One ensemble transcriptomic study reported 94.3 percent sensitivity, but without external validation.

Interpretability emerged as a persistent weak point. Only a handful of the reviewed studies deployed explainable AI techniques, such as LIME applied to SNP-based classification or feature-importance rankings in random forests, and none assessed whether those explanations were stable or clinically meaningful. The review notes a fundamental trade-off: the deep learning models that achieve the best predictive performance tend to be the least transparent, while simpler, more interpretable models may fail to capture the disease’s heterogeneity. No study evaluated how algorithmic explanations actually affect physician confidence or decision-making, a gap the authors identify as central to clinical adoption.

The review also stresses that cross-study comparisons should be read descriptively rather than as head-to-head evidence of superiority. The 26 studies differ enormously in cohort size and composition, disease phenotype, control-group definition, preprocessing, validation strategy and endpoint, spanning disease classification, lesion reconstruction, disability prediction and treatment-response estimation. Differences in accuracy or AUC may therefore reflect methodological and endpoint variability as much as intrinsic model quality. Several studies reported more modest results, including a genomic artificial neural network with 64.7 percent accuracy and a clinical statistical model with a concordance index at or below 0.65, underscoring that high performance is not universal.

On the question of what comes next, the authors outline emerging techniques intended to close the validation and privacy gaps. Federated learning would allow hospitals to train shared models without exchanging sensitive patient data, an approach particularly relevant given that regulations such as GDPR and HIPAA restrict cross-border sharing of genetic information. Foundation models trained on large unlabeled imaging and omics datasets through self-supervised learning, multimodal representation learning that fuses MRI, OCT, genomic and clinical inputs into shared latent representations, and causal inference methods that separate correlation from mechanism all feature in the review’s forward-looking section. Information-theoretic feature selection is highlighted as a mathematically grounded way to tame ultrahigh-dimensional genomic data.

The authors also propose a unified framework for MS diagnosis that layers interpretability, external validation and bias control onto a standard pipeline of multimodal data acquisition, feature extraction, integration and classification. They argue that evaluation criteria must extend beyond accuracy to include fairness, robustness and utility, noting that models trained on homogeneous, predominantly European datasets can produce biased results when applied to more diverse populations. Lifestyle, environmental and socioeconomic factors that influence disease development, such as smoking, vitamin D exposure and healthcare access, are largely absent from current models, a limitation the review says may lead to imprecise predictions and underdiagnosis in underrepresented groups.

The review acknowledges its own constraints. Restricting inclusion to peer-reviewed, English-language articles may have omitted relevant preprints and non-English research, and the heterogeneity of assessment measures made quantitative meta-analysis infeasible. Most included studies prioritized point-in-time diagnosis over longitudinal monitoring, so the review draws no conclusions about AI’s ability to predict disease trajectory or treatment response over time. Still, the synthesis offers a clear map of a fast-moving field: AI can extract meaningful biomarkers from retinal scans, brain images and gene expression profiles, and multimodal integration appears to outperform single-modality approaches, but translating those research results into trustworthy clinical tools will require large, diverse, externally validated cohorts and explainable models that clinicians can actually interrogate.

Subject of Research: A systematic review of artificial intelligence models for multiple sclerosis diagnosis using medical imaging, genomic and multimodal data.

Article Title: Artificial intelligence-driven multiple sclerosis diagnosis using genomic and medical imaging data: a review

Article References: Wahab Sait, A. R., & Alkhurayyif, Y. (2026). Artificial intelligence-driven multiple sclerosis diagnosis using genomic and medical imaging data: a review. Frontiers in Medicine, 13, Article 1812134. https://doi.org/10.3389/fmed.2026.1812134

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1812134

Keywords: multiple sclerosis, artificial intelligence, machine learning, medical imaging, MRI, optical coherence tomography, genomics, deep learning, diagnosis, systematic review, interpretability, federated learning