Residual analysis of ground reaction forces simulation during gait using neural networks with different configurations.

Abstract

The aim of the study was to analyze and compare the residuals obtained from ground reaction force (GRF) models developed using two different neural network configurations (one network with three outputs; and three networks with one output each), based on accelerometer data. Seventeen healthy subjects walked along a walkway, with a force plate embedded, with a three dimensional accelerometer attached to the shank. Multilayer perceptron networks (MLP) models were developed with the 3D accelerometer data as inputs to predict the GRF. The residuals of these models were evaluated graphically and numerically to verify the fitting. A visual analysis of the simulated signals suggests the model was able to adequately predict the GRF. The errors and correlations found in the MLP models for the 3D GRF is at least similar to other studies, although some of them showed higher errors. There was not difference between the two MLP configurations. However, despite the high correlation coefficient and closeness to a normal probability distribution, the residual analysis still presented a higher kurtosis and skewness, suggesting that the inclusion of other variables and the increase of the validation sample size could increase the fitting of the simulation.