Learning Post-Stroke Gait Training Strategies by Modeling Patient-Therapist Interaction.

Area of Science:

• Robotics
• Rehabilitation Engineering
• Biomechanics

Background:

• Robot-aided gait training requires integrating physical therapist expertise for safety and efficacy.
• Current methods often lack the nuanced, adaptive strategies employed by human therapists during gait assistance.

Purpose of the Study:

• To develop a computational model that captures physical therapists' decision-making processes during manual gait assistance.
• To encode therapist strategies into a control framework for wearable robots used in stroke rehabilitation.

Main Methods:

• Collected kinematic and force data from physical therapists assisting patients with lower-limb movement.
• Utilized a wearable sensing system with a force sensing array to measure assistive forces.
• Developed a virtual impedance model incorporating key gait features (knee extension, weight-shifting) to predict therapist's assistive torque.

Main Results:

• Identified knee extension and weight-shifting as critical features influencing therapist assistance.
• The virtual impedance model accurately predicted high-level therapist behaviors (r2 = 0.92) and nuanced stride-level actions (r2 = 0.53).
• Demonstrated the model's ability to capture therapist strategies over full training sessions and individual strides.

Conclusions:

• Directly encoding physical therapist decision-making into robot control offers a novel approach for gait rehabilitation.
• This method enhances the safety and effectiveness of human-robot interaction in wearable robotics for stroke recovery.
• The developed model provides an intuitive framework for characterizing and estimating therapist assistance strategies.