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

Background/Objectives: Post-stroke gait impairment is highly heterogeneous, which limits the effectiveness of standardized exoskeleton control strategies. Deep reinforcement learning offers a route to adaptive assistance, but its use in stroke rehabilitation is constrained by limited pathological gait data and the lack of interpretable transfer frameworks. We developed a data-efficient, pathology-informed reinforcement learning framework for personalized exoskeleton assistance under limited clinical gait data. Methods: The framework combines neuromuscular-inspired parametric augmentation (NIPA) with parameter-efficient transfer learning. NIPA synthesizes pathological gait trajectories by modeling weakness, stiffness or contracture, and abnormal synergies. A policy is first pretrained in simulation and then adapted to clinical gait data by freezing a shared feature extractor and fine-tuning the output heads. The framework was evaluated on a public clinical gait dataset of 50 stroke survivors using tracking error, reward, smoothness, generalization, and data efficiency as main outcomes. Results: The proposed method outperformed zero assistance, rule-based control, and reinforcement learning from scratch on the test set. Compared with scratch, it reduced total 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). Hip MSE decreased from 5.9544 to 4.0143 (p=7.51×10-8) and knee MSE decreased from 6.5507 to 5.4507 (p=1.51×10-5), with significant improvements in repeated experiments. Conclusions: The proposed framework reduces reliance on large pathological training datasets and improves offline trajectory-level personalization under limited clinical data. It also provides an interpretable basis for quantitative characterization of post-stroke gait heterogeneity and may support individualized rehabilitation assessment and assistance planning.