Data-driven Gait-predictive Model for Anticipatory Prosthesis Control.

Area of Science:

• Biomechanics
• Robotics
• Machine Learning

Background:

• Physiological movement relies on pre-planning using sensory information and environmental cues for efficient musculoskeletal tissue utilization.
• Prosthetic devices require advanced control systems to compensate for inherent system delays.
• Anticipatory control is crucial for seamless integration of prosthetic limbs with user intent.

Purpose of the Study:

• To develop a gait predictive model for anticipating lower-limb joint movements.
• To predict ankle angles and moments 50 milliseconds in advance using past trajectory data.
• To enable anticipatory control for lower-limb prostheses.

Main Methods:

• A gated recurrent neural network (GRN) based model was employed.
• The model utilized past trajectory data of input signals for prediction.
• Shank angle was used as a single input signal for prediction.

Main Results:

• High accuracy in predicting ankle angles ($R^{2} > 0.91$) was achieved.
• Accurate prediction of ankle moments ($R^{2} > 0.91$) was also observed.
• Predictions were validated on walking trials at a self-selected comfortable speed.

Conclusions:

• The developed gait predictive model demonstrates high accuracy in forecasting lower-limb joint kinematics and kinetics.
• This predictive capability is valuable for enhancing anticipatory control in lower-limb prosthesis applications.
• Real-time prediction of user locomotive intent can optimize prosthetic joint actuation.