Application of a Neural ODE to Classify Motion Control Strategy using EEG

Speed-accuracy trade offs exist in a variety of functional tasks, which may require differences in control strategies in future neuroprosthetic devices. It is the goal of this work to evaluate the predictability of different motor control strategies during wrist rotation tasks. Participants were asked to perform a series of discrete wrist rotations. This motion data was clustered into segments of either speed or range of motion oriented control strategy, controlling for age cohort and motion type. Competing neural ordinary differential equation (NODE) and random forest (RF) models were evaluated to explore the feasibility of classifying control strategy using cortical data alone. In comparison to traditional ML techniques, such as RF models, the NODE model provided achieved comparable classification accuracy at a fraction of the time. Furthermore, the use of a single motor cluster or two frontal clusters provided similar accuracy to the full data from 4 clusters, which may due to increased information from these cortical areas. This study provided a promising initial demonstration of the benefits of NODE models for future brain-computer-interface applications that require near real-time classification.