Deep Temporal Clustering for Long-Term Gait Recovery Patterns of Post-Stroke Patients using Joint Kinematic Data

This study aims to analyze long-term gait recovery patterns of post-stroke hemiplegic patients by applying a deep learning-based clustering algorithm using joint kinematic data. An end-to-end deep learning model was employed to autonomously identify long-term gait recovery patterns from time-series joint angle data, eliminating the need for manual feature extraction. Gait data were collected from 31 patients using a motion capture system over a six-month rehabilitation period. The Deep Temporal Clustering (DTC) algorithm effectively clustered temporal data, identifying five distinct gait recovery groups with an average silhouette score of $\mathbf{0. 4 2 5 6}$. This study emphasizes the capability of deep learning to cluster long-term gait recovery patterns, providing valuable information to develop personalized rehabilitation strategies.