To predict whether people with multiple sclerosis (PwMS) exhibit longitudinal cognitive decline from baseline clinical MRI data using artificial intelligence and state-of-the-art automated processing methods for measuring atrophy and structural network disruption.

A retrospective analysis of 147 PwMS was conducted to predict cognitive decline over several years (mean=3.9±2.2) from baseline clinical T2-FLAIR MRI. The previously validated NeuroSTREAM and Network Modification tools were applied on baseline T2-FLAIR MRI to automatically measure lateral ventricular volume and patterns of structural network white matter tract disruption, respectively. This baseline data was then applied to predict cognitive decline, defined as a drop in 4-points or greater on the Symbol Digit Modalities Test (SDMT). The data was divided randomly into two discrete subsets (n=117; n=30). The first was used to train a modular deep neural network to predict cognitive decline and the other was used to test/validate the predictive accuracy of the model.

44.8% of study participants (n=50 training; n=16 testing) exhibited cognitive decline over the period of observation. After training the modular deep neural network, we achieved moderate predictive accuracy in the separate test/validation set (accuracy=63%; sensitivity=0.65; specificity=0.86; false positive=2; false negative=9; F1=0.62). These results compare favorably to previously reported R² values of 0.20-0.35 on research-quality full connectomic investigations.