Prediction of EMG Activation Profiles from Gait Kinematics and Kinetics during Multiple Terrains

Continuous myoelectric prediction of intended limb dynamics has the ability to provide transparent control of a prosthesis by the user. However, the impact on these models of adding a human user into the control loop is less clear. Here, the ability of a User Response Model (URM) to continuously predict EMG activity from gait kinematics and kinetics collected during three mobility tasks (level-ground walking, stair ascent, and stair descent) was examined. Multiple-input, multiple-output NARX-based URMs were developed with two outputs (ankle plantarflexor and dorsiflexor) and variable inputs (ankle kinetics, and shank and/or ankle kinematics). Accuracy in predicting the tibialis anterior and medial gastrocnemius EMG was comparable across URMs regardless of the number of inputs. Stair descent had the lowest accuracy among the mobility tasks. No significant differences in normalized root-mean-square error and cross-correlation were found between URMs with five and nine inputs. A URM that continuously predicts EMG activity from gait kinetics and kinematics could be used to simulate human-in-the-loop myoelectric control of a transtibial prosthesis and examine the stability of the system to changes in the environment or due to control errors.