Predictive models of multiple sclerosis-related cognitive performance using routine clinical practice predictors

Abstract

Background: The application of machine learning (ML) to predict cognitive evolution is exceptionally scarce. Computer-based self-administered cognitive tests provide the opportunity to set up large longitudinal datasets to aid in developing ML prediction models of risk for Multiple Sclerosis-related cognitive decline.

Objective: to analyze to what extent clinically feasible models can be built with standard clinical practice features and subsequently used for reliable prediction of cognitive evolution.

Methods: This prospective longitudinal study includes 1184 people with MS who received a Processing Speed (PS) evaluation at 12 months of follow-up measured by the iPad®-based Processing Speed Test (PST). Six of the most potent classification models built with routine clinical practice features were trained and tested to predict the 12-month patient class label (PST worsening (PSTw) versus PST stable). A rigorous scheme of all the preprocessing steps run to obtain reliable generalization performance is detailed.

Results: Based on a 12-month reduction of 10% of the PST raw score, 187/1184 (15.8%) people with MS were classified as PSTw. The trees-based models (random forest and the eXtreme Gradient Boosting) achieved the best performance, with an area under the receiver operating characteristic curve (AUC) of 0.90 and 0.89, respectively. The timing of high-efficacy disease-modifying therapies (heDMTs) was identified as one of the top importance predictors in all the models evaluated.

Conclusion: Using trees-based machine learning models to predict individual future information processing speed deterioration in MS could become a reality in clinical practice.

Keywords: Cognitive dysfunction; High efficacy disease-modifying therapies; Information processing speed; Machine learning; Multiple sclerosis; Processing speed test.