# Machine learning integration of MRI and gait reveals mobility phenotypes in multiple sclerosis

## TL;DR

The study uses machine learning to identify four mobility phenotypes in multiple sclerosis patients by combining MRI and gait data, revealing hidden patterns of gait impairment.

## Contribution

A novel intermediate mobility phenotype was identified, characterized by increased cadence and shorter strides, linked to MRI signs of structural impairment.

## Key findings

- Four distinct gait clusters were identified using machine learning, reflecting varying degrees of mobility impairment and MRI features.
- Cluster 3 showed increased cadence and shorter strides, with higher lesion burden and lower brain parenchymal fraction, suggesting compensatory gait.
- Unstable clusters correlated with progressive MS, higher disability scores, and longer disease duration, aligning with MRI and clinical data.

## Abstract

Mobility impairment is a hallmark of disease worsening in multiple sclerosis (MS), yet its phenotypic diversity and pathophysiology mechanisms are not completely understood. Conventional gait assessments often rely on subjective clinical measures, which may not fully capture the complexity of gait abnormalities. The integration of advanced quantitative gait analysis, quantitative from MRI, and machine learning (ML) may reveal unique mobility phenotypes, potentially reflecting underlying disease mechanisms and heterogeneity. In this study, we aimed to identify and characterize mobility phenotypes among people with MS (pwMS) using a mixed approach with spatiotemporal gait parameters and MRI-derived features, supported by unsupervised ML clustering. 1026 pwMS underwent comprehensive gait assessments and quantitative MRI between 2018 and 2023. Principal component analysis was applied for dimensionality reduction and k-means clustering to identify distinct phenotypes. Clusters were compared using demographic, clinical, and MRI features, with statistical comparisons performed using Kruskal–Wallis and Chi-square tests. Four gait clusters were identified. Cluster 1 (faster stable, 47.8%), demonstrated the most efficient gait features and highest grey matter fractions. Cluster 4 (slow severely unstable, 7.4%) showed profound disability, shortest strides, lowest velocity, and greatest variability. Intermediate clusters 2 (slower stable, 32.3%) and 3 (moderately unstable, 12.6%) had similar velocity but differed in cadence and stride length. Cluster 3, marked by shorter steps and increased cadence, showed higher lesion burden and lower brain parenchymal fraction, suggesting emerging structural impairment and possible compensatory gait. Clinical measures aligned with these findings: unstable Clusters 3 and 4 had the highest proportion of progressive MS, worst disability scores, longest disease duration, and greatest self-reported gait impairment. Integrating quantitative MRI metrics with spatiotemporal gait analysis has the potential to phenotype clinical impairments in pwMS. ML-driven analysis identified a novel intermediate cluster, distinguished by a gait with increased cadence and shorter strides, alongside distinct MRI abnormalities. This pattern may reflect a potential adaptation within the mobility spectrum, not yet conclusively discernible by human raters but detectable through ML.

Inojosa et al. report that integrating quantitative MRI metrics with spatiotemporal gait analysis in people with multiple sclerosis identified four distinct mobility phenotypes using machine learning. This data-driven approach identifies subtle gait impairments and compensatory patterns, offering new insight into disease-related motor dysfunction and potential targets for early intervention.

Graphical Abstract