Predicting the initiation of minimum-jerk submovements in three-dimensional target-oriented human arm trajectories

Target-oriented human arm trajectories can be represented as a series of summed minimum-jerk submovements. Under this framework, corrections for errors in reaching trajectories could be implemented by adding another submovement to the ongoing trajectory. It has been proposed that a feedback-feedforward error-detection process continuously evaluates trajectory error, but this process initiates corrections at discrete points in time. The present study demonstrates the ability of a feed-forward Artificial Neural Network (ANN) to learn the function of this error-detection process. Experimentally recorded human target-oriented arm trajectories were decomposed into submovements. It was assumed that the parameters of each submovement are known at their onset. Trained on these parameters, for each of three participants, an ANN can predict presence of corrections with sensitivity and specificity > 80%, and can predict their timing with R(2) > 40%.