Can Machine Learning Detect and Predict Outcomes in Multiple Sclerosis? A Review of Diagnostic and Prognostic Approaches

Multiple sclerosis (MS) is a chronic inflammatory and neurodegenerative disease of the central nervous system, and achieving an accurate diagnosis remains a persistent clinical challenge. The 2017 McDonald criteria, which integrate clinical findings with magnetic resonance imaging (MRI) evidence of lesion dissemination in space and time, have improved diagnostic sensitivity and facilitated earlier diagnosis; however, their accurate application remains highly dependent on clinician expertise, and misdiagnosis continues to occur, particularly when MS must be differentiated from other white matter or inflammatory conditions (1).

Over the past decade, researchers have developed machine learning (ML) and deep learning (DL) methods to assist in this diagnostic process. Early applications focused primarily on automating the labor-intensive task of MS lesion segmentation on MRI, which has traditionally been performed manually and is subject to substantial inter-rater variability (2, 3). As computational methods have advanced, their applications have expanded to full diagnostic classification, including differentiation of MS from healthy controls and, more recently, from clinically similar disorders such as neuromyelitis optica spectrum disorder (NMOSD) (1, 4). In parallel, researchers have increasingly explored non-MRI modalities, including optical coherence tomography (OCT) and blood and cerebrospinal fluid (CSF) biomarkers, as additional and less invasive substrates for ML-based detection (5, 6).

Beyond diagnosis, ML has also been applied to patients with established MS to forecast disability accumulation and symptom burden, i.e., tasks that are prognostic rather than diagnostic. Nevertheless, these applications share the same underlying methodology and are frequently discussed alongside diagnostic applications in the literature (4, 7). Disability, most commonly quantified using the Expanded Disability Status Scale (EDSS), is challenging to predict because it reflects the heterogeneous and non-linear accumulation of impairment across multiple neurological domains. Recent cohort studies have applied ensemble ML methods to predict future EDSS scores, disease-specific symptom severity, and changes in disability during clinically important periods, such as the postpartum period, often achieving higher classification accuracy than continuous regression approaches for disability scores.

Despite the growing body of research across both domains, the translation of ML tools into routine clinical practice remains limited. Reported performance varies considerably among studies, and this heterogeneity is frequently attributed to differences in dataset size, imaging or clinical feature protocols, and the selection of comparison or outcome groups (1, 7). The present review synthesizes current evidence on ML-based detection and classification of MS together with ML-based prediction of disability progression and symptom severity to clarify (i) which data modalities and model families have demonstrated the strongest performance in each domain, (ii) how performance differs across tasks of increasing clinical difficulty, and (iii) which methodological barriers must be addressed before these tools can be considered for routine clinical deployment.

This is a narrative, rather than a systematic, review. Sources were identified through purposive searches of PubMed and Google Scholar (last search: July 2026) using combinations of the terms “multiple sclerosis,” “machine learning”, “deep learning”, “diagnosis”, “prognosis”, and “disability progression”, supplemented by manual screening of the reference lists of recently identified reviews. We retained 25 sources published between 2012 and 2026 based on their direct relevance to ML-based MS diagnosis or disability/symptom prognosis, methodological clarity, and, for primary studies, availability of quantitative performance metrics. This purposive approach does not guarantee exhaustive coverage and reflects the authors’ judgment regarding representative and influential work rather than a systematic sample.

Chronological Evolution of the Field

Table 1 summarizes 25 key sources in chronological order, illustrating how the field has evolved from early, narrowly focused lesion-segmentation tools to broader diagnostic frameworks and, more recently, to prognostic and disability-prediction applications.

As listed in Table 1, the earliest work (2012) focused almost exclusively on automated lesion segmentation using MRI (3, 8, 9), reflecting the practical need to reduce the burden associated with manual lesion annotation. Between 2018 and 2021, the field shifted toward DL architectures capable of full diagnostic classification rather than segmentation alone, alongside early consensus efforts to standardize non-MRI modalities, such as OCT (2, 10, 12). From 2022 onward, the literature has increasingly emphasized translational challenges, including generalizability, differential diagnosis, and multimodal data integration, rather than purely technical improvements in model performance (1, 13, 14, 17). The most recent sources (2024-2026) reflect a further expansion in scope along two directions: toward fluid biomarkers and retinal imaging on the diagnostic side (6, 19, 22) and toward ensemble ML models for predicting disability progression, postpartum disability change, and symptom-specific severity on the prognostic side (23-25). This latter shift suggests that, once a diagnosis is established, the same model families originally developed for classification are increasingly being repurposed to address a clinically distinct question: how will this patient’s disease evolve? It should be noted that 22 of the 25 sources summarized in Table 1 are literature reviews or methodological surveys rather than primary studies. In particular, the performance ranges reported in Table 2 are derived from two review articles (1, 7) rather than recalculated from the primary studies they cite and should therefore be interpreted as review-level summaries rather than independently verified findings from the primary studies.

ML Approaches for MS Diagnosis

MRI-based Classification: Conventional ML versus DL

MRI remains the principal data source for ML-based MS detection (1, 4, 10). Two broad methodological families predominate in this literature. DL approaches [most commonly convolutional neural networks (CNNs) and U-Net variants] operate directly on raw T2-weighted, fluid-attenuated inversion recovery, or susceptibility-weighted images and have achieved accuracies ranging from 70.2% to 98.8% in distinguishing MS from healthy controls (1). In contrast, conventional ML approaches [including support vector machines, random forests (RFs), and logistic regression] typically rely on engineered features derived from quantitative, diffusion-weighted, or resting-state functional MRI and have reported somewhat narrower accuracy ranges of 83.7-90.0% (1). Notably, the higher peak performance observed with DL models does not necessarily indicate consistent superiority, as performance in both methodological families depends heavily on dataset composition, imaging modality, and the definition of the diagnostic task (13, 21).

The diagnostic reference standard itself is also evolving. The 2017 McDonald criteria, against which nearly all models discussed in this review were trained and validated, are themselves undergoing revision: proposed updates incorporate the central vein sign and paramagnetic rim lesions as imaging markers of MS-specific pathology, kappa free light chains as a fluid biomarker, and the optic nerve as an additional topography for demonstrating dissemination in space. Several of these candidate markers are precisely the targets of the lesion-subtype segmentation tools described in Section 3.2. Models trained and benchmarked under the current criteria may therefore require retraining or revalidation as the reference standard they were developed to approximate continues to evolve.

Automated Lesion Segmentation

Lesion segmentation is the most methodologically mature ML application in MS, with a research history extending back to at least the early 2010s (3, 8). Manual lesion assessment is time-consuming and subject to inter-rater variability, which prompted the early adoption of automated approaches (2). U-Net-type CNN architectures now predominate in this subfield, and current tools target not only conventional T2 lesions but also gadolinium-enhancing lesions, central vein sign-positive lesions, paramagnetic rim lesions, and cortical and spinal cord lesions (19). Despite this progress, no segmentation method has achieved universal acceptance in clinical practice, and expert expectations regarding reliability continue to exceed the performance delivered by most current algorithms (2, 15). It is important to distinguish regulatory clearance from demonstrated clinical utility: several MS lesion quantification tools have obtained Conformite Europeenne marking or U.S. Food and Drug Administration (FDA) clearance, but such clearance primarily certifies technical performance and safety rather than demonstrating benefits to patient outcomes or constituting formal endorsement by clinical practice guidelines. Cleared tools also remain sensitive to scanner hardware, acquisition protocols, and field strength and typically require post-deployment performance monitoring when used outside the conditions under which they were validated.

Differential Diagnosis: MS versus Clinical Mimics

A clinically important extension of ML-based classification is the differentiation of MS from disorders with overlapping clinical presentations, most notably NMOSD (9, 16). RF models using gray-matter measures alone have achieved approximately 74% accuracy in distinguishing MS from NMOSD. In contrast, multiparametric models incorporating MRI with clinical and neuropsychological data have achieved approximately 88% accuracy (1). This gap illustrates a recurring theme: models trained only to distinguish MS from healthy controls may perform considerably worse when applied to realistic differential-diagnosis comparisons, such as vascular white-matter lesions or other inflammatory disorders (1).

Beyond MRI: OCT, Retinal Imaging, and Fluid Biomarkers

The field is expanding beyond structural MRI. OCT, which quantifies retinal nerve fiber and ganglion cell layer thickness, has shown promising preliminary results for distinguishing MS from healthy controls. However, sample sizes remain small, and standardization of quality control is still evolving (5, 12). Consensus efforts, such as the artificial intelligence (AI)-extended OSCAR-IB criteria (OCT Standardization of Image Acquisition and Analysis in MS) criteria, aim to address this standardization gap (12). DL applied to retinal imaging is increasingly being positioned as a non-invasive, low-cost complement to neurological assessment (22). Separately, ML applied to blood and CSF biomarkers remains comparatively underexplored relative to imaging-based approaches, despite its potential to provide an additional minimally invasive diagnostic substrate (6).

Clinical Integration and Multimodal Approaches

Several reviews argue that multimodal integration, i.e., combining MRI with clinical, demographic, OCT, or fluid biomarker data, is the most credible path toward improved diagnostic utility, as no single marker captures the full heterogeneity of MS pathology (6, 17, 18). At the same time, reviews assessing real-world clinical translation caution that most current AI tools remain assistive rather than autonomous, requiring human oversight and further validation before routine deployment (14, 19). This assistive positioning is consistent with the regulatory status of most tools in this space: regulatory clearance indicates that a tool performs a specified technical function safely and reproducibly, rather than demonstrating that it improves diagnostic accuracy or patient outcomes in clinical practice. Guideline endorsement of autonomous decision-making remains absent across the tools discussed in this review.

ML for Disability Progression and Symptom Prediction

Once a diagnosis of MS is established, a distinct but methodologically related question arises: how will the patient’s disability and symptom burden evolve? Three recent cohort studies illustrate this prognostic application of ML, summarized chronologically in Table 2.

EDSS and Disability Progression Prediction

Ozdogar et al. (24) developed an XGBoost-based model (CorBoost) to predict 2-year EDSS scores in 231 people with MS, using 126 baseline demographic, clinical, physical, and cognitive features that were reduced to 42 through correlation-based feature selection. The model achieved a coefficient of determination (R²) of 0.97 and a mean absolute error (MAE) of 0.24 EDSS points, outperforming the RF and linear regression comparators. Baseline EDSS, disease duration, and age were the most influential predictors. Notably, the model achieved this performance using only routinely collected clinical data, without MRI or fluid biomarker input. This very high R² should nonetheless be interpreted with caution because EDSS is strongly autocorrelated over a 2-year interval and baseline EDSS was the leading predictor; thus, much of the explained variance may reflect a patient’s baseline status rather than genuine predictive insight into disease evolution. The reported MAE of 0.24 EDSS points is notably below the scale’s own inter-rater measurement error, further suggesting that the model’s practical incremental value over baseline disability alone may be more modest than the headline R² implies.

Symptom-specific Severity Prediction

Kaya et al. (23) applied DecisionTree, RF, and XGBoost models to predict restless legs syndrome (RLS) severity in a small cohort of 23 people with MS, using sleep, mobility, and quality-of-life questionnaire data. XGBoost performed best (MAE: 1.94; accuracy: 92.65%), with the Pittsburgh Sleep Quality Index emerging as the dominant predictor across all three algorithms, consistent with conventional correlation analysis. Reporting both a continuous error metric (MAE) and a categorical accuracy measure indicates that the authors evaluated continuous as well as categorical formulations of RLS severity; readers are referred to the primary source for the exact sex composition of the 23-person cohort. This study illustrates that ML can extend beyond global disability metrics to model specific, clinically burdensome MS symptoms, although the very small sample size limits generalizability.

Context-specific Disability Prediction: The Postpartum Period

Emec et al. (25) modeled postpartum disability change (ΔEDSS) in a large real-world cohort of 662 women with MS contributing 909 pregnancies, comparing regression models (RF, XGBoost, Elastic Net, Support Vector Regression) with classification models predicting improved, stable, or worsened disability categories. Classification models substantially outperformed regression models (test accuracy, 85-88%; F1: 0.84-0.87; versus test R²=0.31-0.39, for regression), a pattern the authors attributed to the discretized, threshold-like distribution of clinically meaningful EDSS change. Postpartum relapse was the strongest predictor of disability worsening, followed by disease duration and age at pregnancy, with obstetric variables such as delivery type and breastfeeding contributing secondary predictive value.

It should be noted that references 23, 24, and 25 share substantial overlap in authorship and appear to draw on a largely common patient population and research group, so they are better interpreted as a coherent program of work from one center rather than as three independent replications. Independent multicenter prognostic studies are still needed before the pattern described below can be generalized with confidence. Across all three studies, tree-based ensemble methods (RF, XGBoost) consistently matched or outperformed simpler linear models, and categorical outcome framing tended to generalize better than continuous regression, mirroring a pattern also noted in the diagnostic literature, where task definition strongly influences reported performance (13, 21).

Diagnostic and Prognostic Performance Summary

Table 2 summarizes the three prognostic and disability-prediction studies described above, whereas Table 3 consolidates diagnostic performance measures by task and data modality.

Table 4 summarizes qualitative risk-of-bias judgments for the prognostic studies; similarly, the diagnostic performance measures in Table 3 vary in the completeness of their reporting. Table 5 makes this limitation explicit by summarizing, for each row of Table 3, the validation strategy, aggregate sample size, and class balance reported in the underlying source. When a source did not report one of these elements, we state this explicitly rather than leaving the field blank, because the absence of this information is itself relevant to interpreting the reported accuracy measures.

Challenges and Limitations

Several methodological limitations recur across both domains. Most studies rely on small, single-center datasets, thereby limiting generalizability across scanners, acquisition protocols, and patient populations; this limitation is particularly apparent in symptom-prediction studies involving cohorts as small as 23 patients (1, 10, 23). Diagnostic models are frequently validated against healthy controls rather than clinically representative comparator groups, whereas prognostic regression models often exhibit substantial overfitting between training and test datasets (1, 16, 25). DL as well as ensemble tree-based models remain relatively opaque, which can hinder clinician confidence, interpretability, and regulatory evaluation (10, 19). Finally, prospective multicenter validation, which is a prerequisite for broader clinical deployment, remains uncommon, with most available evidence derived from retrospective, single-center study designs (1, 17, 24, 25).

To move this observation from assertion toward demonstration, Table 4 provides a brief, Prediction Model Risk Of Bias Assessment Tool (PROBAST)-informed qualitative appraisal of the three primary prognostic studies discussed in Section 4. This appraisal is based solely on information reported in the published articles or summarized above and should not be considered a substitute for a formal PROBAST or TRIPOD+AI audit conducted with access to the full study protocols; rather, it is intended to illustrate, without exhaustively documenting, the methodological domains most likely to influence generalizability.

Future Directions

Priorities for future work include: (i) developing multimodal models that integrate MRI with OCT, fluid biomarkers, and clinical history to improve upon single-modality systems (6, 18); (ii) establishing open, standardized, multicenter datasets and implementing federated learning to address generalizability limitations across diagnostic as well as prognostic applications (19); (iii) placing greater emphasis on model interpretability and prospective validation in clinically realistic differential-diagnosis and longitudinal-prognosis settings (1, 19); and (iv) externally validating prognostic models across healthcare systems while incorporating imaging or fluid biomarkers alongside the clinical variables evaluated to date (24, 25). Given that DL as well as ensemble tree-based models remain largely opaque (Section 6), explainable artificial intelligence frameworks, such as SHAP and LIME, are increasingly being proposed as potential approaches to bridging the gap between model outputs and clinical trust. By attributing a model’s output to specific input features on a per-patient basis, these methods could enable clinicians to scrutinize individual predictions against established MS pathophysiology rather than treating model outputs as opaque scores, potentially supporting regulatory evaluation and guideline acceptance of these tools.

ML can detect MS primarily by identifying lesions and disease-associated patterns on MRI, with an expanding evidence base for applications involving OCT and fluid biomarkers. Once a diagnosis is established, the same broad model families (particularly tree-based ensembles such as RF and XGBoost) can generate clinically meaningful predictions of disability progression, context-specific disability change, and symptom-specific severity, often achieving higher reported performance when outcomes are framed categorically rather than continuously. Across both diagnostic and prognostic applications, the strongest and most consistent evidence supports computer-assisted analysis rather than autonomous clinical decision-making: performance remains sensitive to dataset size and quality, single-center study designs, and the clinical realism of validation settings. Headline accuracy measures should also be interpreted cautiously when model performance may be driven substantially by strongly autocorrelated baseline variables (Section 4.1) rather than by a novel predictive signal. Continued progress will depend less on isolated algorithmic improvements and more on multimodal data integration, standardized open datasets, and rigorous prospective, multicenter validation across the diagnostic as well as prognostic dimensions of MS care.