Continuous Prediction of Joint Angular Positions and Moments: A Potential Control Strategy for Active Knee-Ankle Prostheses

Abstract

Transfemoral amputation substantially impairs locomotion. To restore the lost locomotive capability, amputees rely on knee-ankle prostheses. Theoretically, active knee-ankle prostheses may better support natural gait than their passive counterparts by replacing the missing muscle function. The control algorithms of such active devices need to comprehend the user's locomotive intention and convert them into control commands for actuating the prosthesis. For an active knee-ankle prosthesis, the gait variables to be controlled to allow the desired locomotion could be the knee angle, knee moment, ankle angle, ankle moment. In this paper, a random forest regression model is employed for the continuous prediction of these gait variables for level ground walking at self-selected normal speed. Experimentally obtained thigh kinematics were the input to the random forest model. The proposed method could predict the angles and moments of the knee and ankle with high accuracy (mean $R^{2}$ value of 0.97 for ankle angle, 0.98 for ankle moment, 0.99 for knee angle, 0.95 for knee moment across four able-bodied subjects). The proposed strategy shows potential for continuously controlling an active knee-ankle prosthesis for transfemoral amputees, whose thigh angular motion can be used to infer the required prosthetic moments or angles.