Classifying multiple sclerosis patients on the basis of SDMT performance using machine learning.

Objective: To build a model to predict cognitive status reflecting structural, functional, and white matter integrity changes in early multiple sclerosis (MS).

Methods: Based on Symbol Digit Modalities Test (SDMT) performance, 183 early MS patients were assigned "lower" or "higher" performance groups. Three-dimensional (3D)-T2, T1, diffusion weighted, and resting-state magnetic resonance imaging (MRI) data were acquired in 3T. Using Random Forest, five models were trained to classify patients into two groups based on 1-demographic/clinical, 2-lesion volume/location, 3-local/global tissue volume, 4-local/global diffusion tensor imaging, and 5-whole-brain resting-state-functional-connectivity measures. In a final model, all important features from previous models were concatenated. Area under the receiver operating characteristic curve (AUC) values were calculated to evaluate classifier performance.

Results: The highest AUC value (0.90) was achieved by concatenating all important features from neuroimaging models. The top 10 contributing variables included volumes of bilateral nucleus accumbens and right thalamus, mean diffusivity of left cingulum-angular bundle, and functional connectivity among hubs of seven large-scale networks.

Conclusion: These results provide an indication of a non-random brain pattern mostly compromising areas involved in attentional processes specific to patients who perform worse in SDMT. High accuracy of the final model supports this pattern as a potential neuroimaging biomarker of subtle cognitive changes in early MS.