BACKGROUND:

Cognitive impairment is a hallmark of multiple sclerosis (MS) but is usually an under-recorded symptom of disease progression. Identifying the predictive signatures of cognitive decline in people with MS (pwMS) over time is important to ensure effective preventative treatment strategies. Structural and functional brain characteristics as measured by various magnetic resonance (MR) methods have been correlated with variation in cognitive function in MS, but typically these studies are limited to a single MR modality and/or are cross-sectional designs. Here we assess the predictive value of multiple different MR modalities in relation to cognitive decline in pwMS over 5 years.

METHODS:

A cohort of 43 pwMS was assessed at baseline and 5 years follow-up. Baseline (input) data consisted of 70 multi-modal MRI measures for different brain regions including magnetic resonance spectroscopy (MRS), diffusion tensor imaging (DTI) and standard volumetrics. Age, sex, disease duration and treatment were included as clinical inputs. Cognitive function was assessed using the Audio Recorded Cognitive Screen (ARCS) and the Symbol Digit Modalities Test (SDMT). Prediction modelling was performed using the machine learning package - GLMnet, where a penalised regression was applied to identify multi-modal signatures with the most predictive value (and the least error) for each outcome.

RESULTS:

The multi-modal approach to neuroimaging was able to accurately predict cognitive decline in pwMS. The best performing model for change in total ARCS (tARCS) included 16 features from across the various MR modalities and explained 54 % of the variation in change over time (R2=0.54, 95 % CI=0.48-0.51). The features included nine MRS, four volumetric and two DTI parameters. The model also selected disease duration, but not treatment, as a predictive feature. By comparison, the best model for SDMT included several of the same above features and explained 39 % of the change over time (R2=0.39, 95 % CI=0.48-0.51). Conventional volumetric measures were about half as good at predicting change in tARCS score compared to the best multi-modal model (R2=0.26 95 % CI0.22-0.29). The clinical interpretation of the best predictive model for change in tARCS showed that cognitive decline could be predicted with >90 % accuracy in this cohort (AUC=0.92, SE=0.86 - 0.94).

CONCLUSION:

Multi-modal MRI signatures can predict cognitive decline in a cohort of pwMS over 5 years with high accuracy. Future studies will benefit from the inclusion of even more MR modalities e.g., functional MRI, quantitative susceptibility mapping, magnetisation transfer imaging, as well as other potential predictors e.g., genetic and environmental factors. With further validation, this signature could be used in future trials with high-risk patients to personalise the management of cognitive decline in pwMS, even in the absence of relapses.