Abstract
This study aims to analyze long-term gait recovery patterns in sub-acute post-stroke hemiplegic patients by applying end-to-end deep learning (DL)-based clustering to sagittal joint-angle trajectories throughout the gait cycle. To address the data scarcity issue in long-term follow-up patient gait trajectory datasets, two time-series data augmentation methods, TimeVAE and Diffusion-TS, were employed to generate high-fidelity synthetic joint-angle trajectories. The augmented dataset were subsequently analyzed using a Deep Temporal Clustering (DTC) model, which effectively captured individualized longitudinal recovery patterns by jointly learning temporal representations and cluster assignments. Based on the clustering evaluation criteria, the model identified six clusters as the optimal grouping. These clusters were statistically well represented by distinct kinematic characteristics. This study represents the first attempt to analyze longitudinal gait recovery patterns in post-stroke patients using a deep clustering model. While exploratory in nature, it provides a conceptual basis for future longitudinal research in stroke rehabilitation.