A Hybrid Kinematic and Machine Learning Approach to Future Joint Angle Estimation at the Ankle

Deploying a reduced sensor count imposes limitations on the accuracy of machine learning-based joint angle estimation models when informing lower-limb assistive devices. Thus, developing additional, biomechanically meaningful input parameters from these sensors and subsequently informing joint angle estimation models with these features may further reduce model error for limited sensor applications. As such, this study explored the effects of including simple, kinematically extrapolated joint angle estimations as Random Forest model input features when using only historical sagittal ankle angles to predict future ankle angles. Results indicated that including N ≥ 1 KE estimations significantly reduced the joint angle estimation error of the Random Forest models across a variety of estimation horizons without meaningfully increasing the model runtime for exoskeleton applications (N = 0: trun = 1.89 ms; N= 25: trun = 2.91 ms). Near future horizons (thzn = 50 -100 ms) only saw increased benefit from a small number of KEs, while larger estimation horizons (thzn = 200 -250 ms) saw benefit from the inclusion of higher counts of KEs. Such results indicate that this simple methodology may be implemented into some single sensor ankle exoskeleton applications to reduce model error without meaningfully increasing computational demand.