Pathology-Informed Personalized Exoskeleton Assistance for Post-Stroke Gait Rehabilitation via Simulation-to-Real...

Area of Science:

• Robotics
• Biomedical Engineering
• Rehabilitation Science

Background:

• Post-stroke gait impairment presents significant heterogeneity, challenging standardized exoskeleton control.
• Deep reinforcement learning (DRL) shows promise for adaptive assistance but is limited by scarce pathological gait data and a lack of interpretable transfer frameworks.

Purpose of the Study:

• To develop a data-efficient, pathology-informed DRL framework for personalized exoskeleton assistance in stroke rehabilitation.
• To address the constraints of limited clinical gait data and enhance the interpretability of transfer learning in this context.

Main Methods:

• Combined neuromuscular-inspired parametric augmentation (NIPA) with parameter-efficient transfer learning.
• NIPA synthesized pathological gait trajectories by modeling weakness, stiffness, and abnormal synergies.
• A policy was pretrained in simulation and adapted to clinical data via transfer learning, freezing feature extractors and fine-tuning output heads.

Main Results:

• The proposed framework significantly outperformed zero assistance, rule-based control, and DRL from scratch.
• Reduced total Mean Squared Error (MSE) from 14.8681 to 11.9369 (p=5.96×10-8) and improved reward from -21.2264 to -18.4798 (p=3.76×10-4).
• Demonstrated significant reductions in hip and knee MSE, indicating improved gait tracking accuracy and smoothness in stroke survivors.

Conclusions:

• The framework effectively reduces the need for extensive pathological training datasets.
• It enhances offline trajectory-level personalization of exoskeleton assistance using limited clinical data.
• Provides an interpretable foundation for characterizing post-stroke gait heterogeneity, aiding individualized rehabilitation planning.