Linear and Non-Linear Predictive Models in Predicting Motor Assessment Scale of Stroke Patients Using Non-Motorized Rehabilitation Device

Abstract

Various predictive models, both linear and non-linear, such as Multiple Linear Regression (MLR), Partial Least Squares (PLS), and Artificial Neural Network (ANN), were frequently employed for predicting the clinical scores of stroke patients. Nonetheless, the effectiveness of these predictive models is somewhat impacted by how features are selected from the data to serve as inputs for the model. Hence, it's crucial to explore an ideal feature selection method to attain the most accurate prediction performance. This study primarily aims to evaluate the performance of two non-motorized three-degree-of-freedom devices, namely iRest and ReHAD using MLR, PLS and ANN predictive models and to examine the usefulness of including a hand grip function with the assessment device. The results reveal that ReHAD coupled with non-linear model (i.e. ANN) has a better prediction performance compared to iRest and at once proving that by including the hand grip function into the assessment device may increase the prediction accuracy in predicting Motor Assessment Scale (MAS) score of stroke subjects. Furthermore, these findings imply that there is a substantial association between kinematic variables and MAS scores, and as such the ANN model with a feature selection of twelve kinematic variables can predict stroke patients' MAS scores.