Large-scale online assessment uncovers a distinct Multiple Sclerosis subtype with selective cognitive impairment

Abstract

Cognitive impairments in Multiple Sclerosis (MS) are prevalent and disabling yet often unaddressed. Here, we optimised automated online assessment technology for people with MS and used it to characterise their cognitive deficits in greater detail and at a larger population scale than previously possible. The study involved 4526 UK MS Register members over three stages. Stage 1 evaluated 22 online cognitive tasks and established their feasibility. Based on MS discriminability a 12-task battery was selected. Stage 2 validated the resulting battery at scale, while Stage 3 compared it to a standard neuropsychological assessment. Clustering analysis identified a prevalent MS subtype exhibiting significant cognitive deficits with minimal motor impairment. Disability in this group is currently unrecognised and untreated. These findings underscore the importance of cognitive assessment in MS, the feasibility of integrating online tools into patient registries, and the potential of such large-scale data to derive insights into symptom heterogeneity.

Fig. 1. Cognitron task design and participant flow.

A Overview of the 24 tasks from the Cognitron platform employed in the study. Twenty-two tasks were evaluated in Stage 1 using a random task sampling approach. Two additional tasks were introduced in Stage 2. Tasks circled in red were selected in Stage 1 to form the C-MS battery used in Stages 2 and 3. B Recruitment steps from the wider UK Multiple Sclerosis (MS) Register cohort and analysis steps. The total number of people with MS (pwMS) in each group is reported. The number of pwMS available for each task varies due to random task sampling and/or assessment completeness and is detailed in Tables S1 and S2. The groups used for each analysis are indicated by boxes with different colours.

Fig. 2. Multiple Sclerosis (MS) discriminability across task performance metrics and task selection for the C-MS battery.

A MS discriminability in terms of accuracy, cost in accuracy (top), response time, and cost in response time (bottom) across tasks for Stage 1 participants (N = 3048; task-specific Ns in Table S1). MS discriminability is defined as the median deviation from expected (DfE) score across participants with 95% confidence intervals for each performance metric. Tasks selected for inclusion in the C-MS battery are reported in green. The primary performance metric for each task is indicated by a blue square. The orange line represents the control median. B Heatmap of factor loadings across primary performance metrics. Tasks selected for each factor are indicated by green squares. C Heatmap of correlations between task performance metrics (top: accuracy and cost in accuracy; bottom: response time and cost in response time) and patient-reported outcomes (PROs) including MSIS-29 motor, MSWS-12, HADS-A, HADS-D, FSS, EQ5D and webEDSS. Only statistically significant (p ≤ 0.05) correlations are shown. D Grid indicating with a green check mark whether a task was selected for the C-MS battery. Source data are provided as a Source Data file.

Fig. 3. Replication of Multiple Sclerosis (MS) discriminability across stages and comparative validity with standard neuropsychological assessment.

A Scatter plots showing MS discriminability in terms of accuracy (left) and response time (right) across tasks for Stage 1 (x-axis, N = 3048) and the independent sample from Stage 2 (y-axis, N = 1425). MS discriminability is defined as the median deviation from expected (DfE) score across participants for each performance metric and is evaluated separately for each stage. Two-tailed Pearson correlation coefficients and corresponding p-values between values at Stage 1 and Stage 2 for accuracy and response time are reported on the plots. B Scatter plots showing global composites of cognitive performance (left) and hand motor function (right) across Stage 3 participants (N = 31), as measured by the standard neuropsychological assessment (x-axis) and the C-MS battery (y-axis). Global cognitive composites were derived by averaging standardised performance scores across tasks within each battery. Nine-Hole Peg Test scores correspond to the less impaired hand. Two-tailed Pearson correlation coefficients and corresponding p-values between global cognitive composites and hand motor function scores from the standard and online assessments are reported on the plots. Source data are provided as a Source Data file.

Fig. 4. Associations of cognitive performance with disease subtype and duration, and task discriminability of early-stage people with Multiple Sclerosis (pwMS) and controls.

A Mean deviation from expected (DfE) scores with 95% confidence intervals for accuracy (top) and response time (bottom) across all tasks, stratified by disease subtype (shades of blue) and disease durations (shades of red), for participants who reported both subtype and onset year in the questionnaire (N = 2689; task-specific Ns in Table S2). Two-way permutational ANOVA was performed separately for each performance metric to test for main and interaction effects of subtype and disease duration. B Median DfE scores with 95% confidence intervals for accuracy (left) and response time (right) across tasks for participants with early-stage MS (between 0 and 5 years post disease onset, N = 298; task-specific Ns in Table S10). Two-tailed Wilcoxon signed-rank tests against zero were used to assess statistical significance. Resulting p-values are reported in the figure. P-values < 0.0001 are reported as <0.0001 due to precision limits based on 10,000 permutations. Source data are provided as a Source Data file.

Fig. 5. Unsupervised clustering of people with Multiple Sclerosis (pwMS) based on digital markers of motor and cognitive impairment and characterisation of resulting clusters.

A Heatmap (top) showing standardised motor and cognitive scores for all participants (N = 1180), sorted by cluster membership. Clusters were derived using K-Means clustering applied to 13 cognitive features (one per task) and 2 motor features (one per patient-reported outcome—PRO). A 4-cluster solution was selected as the most valid and stable after comparing different clustering methods and cluster numbers. Bottom heatmaps display the distribution of participant characteristics—sex, age, education, MS subtype and disease duration—across clusters. B Median scores with 95% confidence intervals for each cluster (Minimal Motor + Moderate Cognitive: N = 307; Severe Motor + Mild Cognitive: N = 403; Minimal Motor + No Cognitive: N = 301; Severe Motor + Severe Cognitive: N = 169) across clustering variables (motor and cognitive features) and additional clinical variables (HADS-D, HADS-A, FSS, EQ5D and webEDSS). Source data are provided as a Source Data file.

Fig. 6. Summary of statistical analyses.

Overview of the 11 main analysis steps performed in the study. The data used for each step are indicated by colour coding.