THE VALIDATION OF AN ARTIFICIAL NEURAL NETWORK TO PREDICT POWER OUTPUT FROM ROWING KINEMATICS

Abstract

The purpose of this research was to develop an individualised artificial neural network (ANN) to predict the rowing performance of a rower using the joint angles produced during the movement. Five novice rowers each completed a 2000m row on a Rowperfect ergometer, during this the kinematics were captured using a 200Hz motion analysis system. The power output of each stroke was obtained from the Rowperfect software. Each ANN was developed to be a fully connected feed-forward back-propagation network trained using the Levenberg-Marquardt method. The input parameters of the network were five joint angles produced during each stroke and the output parameter was the power output produced by the stroke. The results showed an average Pearson correlation co-efficient of 0.83±0.09 (P<0.01) when comparing the actual power output and the ANN predicted power. These significant correlations reveal that the ANN is accurate in predicting power output from joint kinematic data. A Bland & Altman analysis of the data reveals that the power output of the rower can be predicted to an average of 21±6W within a 95% confidence interval. To further develop the research an increased number of rowers will be analysed to develop a more powerful statistical analysis ensuring that generalised movement pattern outputs can be predicted.