Predicting Cognitive Impairment Level after a Serious Game-based Therapy in Chronic Stroke Survivors

Abstract

Accurate prediction of post-treatment outcomes of a rehabilitation program can enable personalized treatments for stroke survivors. Inspired by the clinical effectiveness of our recently developed game-based cognitive rehabilitation tool - Neuro-World- in improving chronic stroke survivors' cognitive level, in this work, we introduce a machine learning approach to predict the post-treatment cognitive level based on information available at baseline. A total of 14 chronic stroke survivors received our game-based therapy for 12 weeks. Subjects were assessed for their cognitive impairment level using the Mini-Mental Status Examination (MMSE) and game performance at baseline and follow-up. Our analytic results show that we can estimate subjects' post-treatment MMSE scores based on their baseline MMSE scores and game performance with a normalized root mean square error (NRMSE) of 9.8% (equivalently, an RMSE of 0.58), which is far more accurate than the population-based prediction method (i.e., averaging). This study opens up new research and clinical opportunities as it allows to stratify our game-based rehabilitation tool to those who are expected to benefit from the treatment, thus enabling optimal, individually-tailored therapeutic programs.