ESTIMATION OF BALANCE STATUS IN PATIENTS WITH HEMIPARESIS: AN ARTIFICIAL NEURAL NETWORK IMPLEMENTATION

Abstract

Objective:Although Balance Evaluation Systems Test (BESTest) is an important balance assessment tool to differentiate balance deficits, it is time consuming and tiring for hemiparetic patients. Using artificial neural networks (ANNs) to estimate balance status can be a practical and useful tool for clinicians. The aim of this study was to compare manual BESTest results and ANNs predictive results and to determine the highest contributions of BESTest sections by using ANNs predictive results of BESTest sections. Methods:66 hemiparetic individuals were included in the study. Balance status was evaluated using the BESTest. 70% (n = 46), of the dataset was used for learning, 15% (n = 10) for evaluation, and 15%(n = 10) for testing purposes in order to model ANNs. Multiple linear regression models (MLRs) were used to compare with ANNs. Results:The results of the study showed that ANNs(root mean square error-RMSE:4.993) were better than MLR (RMSE:7.031) model to estimate balance status of patients with hemiparesis. The BESTest sections making lowest and highest contribution to BESTest total score was found to be Stability Limits/Verticality and Stability in Gait sections, respectively. As the highest and the lowest contribution of sections items were investigated it was found that error(RMSE) values were small indicating the success of ANN modeling. Discussion:The results obtained from this study showed that RMSE values of ANNs were better than the ones found in literature. It is believed that this study can lead to constitute a shorter, more sensitive and more practical mini subset of BESTest for physiotherapists to differentiate balance problems while carrying the whole philosophy of the full BESTest.