Machine Learning in Neuro-Rehabilitation Video Game

In this study, we investigated the potential use of Machine Learning algorithms (ML) to predict the outcome of home-based neuro-rehabilitation video game intervention and its advantage in supporting clinical decision-making. We adopted Support Vector Machines (SVM) and K-Nearest Neighbors (KNN) to develop multidimensional functions (multi-variable Kernel functions) since both algorithms were considered significant and active analysis agents for prediction and classification. Supervised SVM and KNN algorithms were trained using the upper extremity (arm, forearm, and hand) joints’ kinematic data and hand gestures of participants while interacting with the developed video games. Data collected from healthy and Multiple sclerosis (MS) participants were compared and used to develop the predictive algorithm. Pre- and post-rehabilitation data of MS subjects were investigated and used to assess the subject’s functional improvements following the program. Bayesian optimization, Sigmoid, polynomial, and Gaussian Radial Basis functions were utilized for training and predicting outcomes. The results showed that the first two kernel regressions had the best performance regarding predictability and cross-validation loss. KNN’s prediction accuracy was exceeded by 91.7% versus SVM, which was 88.0%. The effectiveness of the rehabilitation program was assessed through Spatiotemporal control and motor assessment scale presenting 40% improvement. Our findings suggest that ML has a great potential to be used for decision-making in neuro-rehabilitation programs.