Modeling Neuromotor Adaptation to Pulsed Torque Assistance During Walking

Multiple mechanisms of motor learning contribute to the response of individuals to robot-aided gait training, including error-based learning and use-dependent learning. Previous models described either of these mechanisms, but not both, and their relevance to gait training is unknown. In this paper, we establish the validity of existing models to describe the response of healthy individuals to robot-aided training of propulsion via a robotic exoskeleton, and propose a new model that accounts for both use-dependent and error-based learning. We formulated five state-space models to describe the stride-by-stride evolution of metrics of propulsion mechanics during and after robot-assisted training, applied by a hip/knee robotic exoskeleton for 200 consecutive strides. The five models included a single-state, a two-state, a two-state fast and slow, a use-dependent learning (UDL), and a newly-developed modified UDL model, requiring 4, 9, 5, 3, and 4 parameters, respectively. The coefficient of determination (R2) and Akaike information criterion (AIC) values were calculated to quantify the goodness of fit of each model. Model fit was conducted both at the group and at the individual participant level. At the group level, the modified UDL model shows the best goodness-of-fit compared to other models in AIC values in 15/16 conditions. At the participant level, both the modified UDL model and the two-state model have significantly better goodness-of-fit compared to the other models. In summary, the modified UDL model is a simple 4-parameter model that achieves similar goodness-of-fit compared to a two-state model requiring 9 parameters. As such, the modified UDL model is a promising model to describe the effects of robot-aided gait training on propulsion mechanics.