Predicting Risk of Multiple Sclerosis Worsening

Abstract

This paper describes our participation in the first two tasks of the iDPP@CLEF 2023 challenge focused on providing clinicians with AI-based methods for better prediction of Multiple Sclerosis progression. We evaluate several standard and transformer-based methods, e.g., Random Forest, Gradient Boosting, and SurfTRACE transformer, to address the risk and cumulative probability of Multiple Sclerosis worsening. The considerable performance increase was achieved by (i) hyper-parameter fine-tuning, (ii) validation procedure, and (iii) data pre-processing. The best method based on the Random Forest algorithm scored first place in Task 1 and 2 (sub-task A) with a C-Index of 0.834, and a mean AUROC score of 0.881, respectively, while reducing the runner-up’s error by 16.2% and 2.3%, respectively. Our methods purely designed and optimized for sub-task A and submitted into sub-task B showed considerable robustness towards overfitting on a specific dataset as achieved third and second place and achieving 0.601 C-Index and second in Task 2, sub-task B of 0.607 mean AUROC score.