Personalized Gait Phase Estimation for Post-Stroke Individuals via Supervised Domain Adaptation with Limited Gait Data.

Accurate gait phase estimation is important for characterizing locomotor progression and enabling phase-dependent monitoring and intervention across gait analysis, rehabilitation, neuroprosthetic, and wearable assistive systems. However, this remains challenging in post-stroke individuals because pathological gait patterns vary substantially across subjects, while patient-specific gait data are often limited. To address this issue, this study proposes a personalized gait phase estimation framework based on supervised domain adaptation (SDA) with a multi-phase maximum mean discrepancy (MP-MMD) objective. This objective aligns source and target latent features separately during stance and swing using stride-specific toeoff labels. A shared feature extractor and domain-specific regressors are jointly trained using data from non-disabled individuals and limited labeled data from each post-stroke participant to construct a subject-specific model. Leave-One-Trial-Out evaluation of seven participants yielded a Root Mean Square Error (RMSE) of 4.37±1.26% and a Maximum Absolute Error (MaxAE) of 8.64±1.90%. After Holm correction, the proposed SDA method achieved significantly lower RMSE than Source Only and Fine-Tuning and lower MaxAE than all three baselines (all adjusted p = 0.047). Relative to Target Only and Fine-Tuning, RMSE showed numerical reductions of 13.32% and 57.51%, respectively, and MaxAE decreased by 26.50% and 72.17%. Compared with conventional global MMD, MP-MMD yielded numerical reductions of 2.87% in RMSE and 2.77% in MaxAE. These results constitute a proof-of-concept for data-efficient within-subject personalization of gait phase estimation.