Muscle activity prediction using wavelet neural network

Abstract

The purpose of this study was to develop a multi dimensional wavelet neural network (WNN) approach in order to predict human lower extremity muscle activities based on ground reaction forces (GRF) and joint angles. For this purpose, four healthy subjects were taken from a previous study. The proposed approach consisted of two main parts: 1) input variable selection (IVS) and 2) network training. First, mutual information (MI) method was used to determine nine inputs including three dimensional GRFs and six joint angles as WNN inputs to predict seven number of outputs. The network was trained based on batch descent gradient algorithm using inter subject data space which provided by leave-one-out (LOO) technique. The WNN predictions for the left-out subject were compared with inverse dynamics calculations based on root mean square error (RMSE) and its percentage as well as Pearson correlation analysis (p). Results showed that multi dimensional WNN was capable to model the highly nonlinear relationship between GRF and joint angles as inputs and muscle activities as outputs.