Kim Gyeongmin, Kim Hyungtai, Kim Yun-Hee, Kim Seung-Jong, Choi Mun-Taek
Department of Intelligent Robotics, Sungkyunkwan University, Suwon 16419, Republic of Korea.
School of Mechanical Engineering, Sungkyunkwan University, Suwon 16419, Republic of Korea.
Bioengineering (Basel). 2025 Jan 11;12(1):55. doi: 10.3390/bioengineering12010055.
Rehabilitation of gait function in post-stroke hemiplegic patients is critical for improving mobility and quality of life, requiring a comprehensive understanding of individual gait patterns. Previous studies on gait analysis using unsupervised clustering often involve manual feature extraction, which introduces limitations such as low accuracy, low consistency, and potential bias due to human intervention. This cross-sectional study aimed to identify and cluster gait patterns using an end-to-end deep learning approach that autonomously extracts features from joint angle trajectories for a gait cycle, minimizing human intervention. A total of 74 sub-acute post-stroke hemiplegic patients with lower limb impairments were included in the analysis. The dataset comprised 219 sagittal plane joint angle and angular velocity trajectories from the hip, knee, and ankle joints during gait cycles. Deep temporal clustering was employed to cluster them in an end-to-end manner by simultaneously optimizing feature extraction and clustering, with hyperparameter tuning tailored for kinematic gait cycle data. Through this method, six optimal clusters were selected with a silhouette score of 0.2831, which is a relatively higher value compared to other clustering algorithms. To clarify the characteristics of the selected groups, in-depth statistics of spatiotemporal, kinematic, and clinical features are presented in the results. The results demonstrate the effectiveness of end-to-end deep learning-based clustering, yielding significant performance improvements without the need for manual feature extraction. While this study primarily utilizes sagittal plane data, future analysis incorporating coronal and transverse planes as well as muscle activity and gait symmetry could provide a more comprehensive understanding of gait patterns.
中风后偏瘫患者步态功能的康复对于提高活动能力和生活质量至关重要，这需要全面了解个体步态模式。以往关于使用无监督聚类进行步态分析的研究通常涉及手动特征提取，这会引入诸如准确性低、一致性差以及由于人为干预导致的潜在偏差等局限性。本横断面研究旨在使用一种端到端深度学习方法来识别和聚类步态模式，该方法能从步态周期的关节角度轨迹中自动提取特征，最大限度地减少人为干预。共有74名患有下肢损伤的亚急性中风后偏瘫患者纳入分析。数据集包含219条在步态周期中来自髋、膝和踝关节的矢状面关节角度和角速度轨迹。采用深度时间聚类通过同时优化特征提取和聚类以端到端的方式对其进行聚类，并针对运动步态周期数据进行超参数调整。通过这种方法，选择了六个最优聚类，轮廓系数为0.2831，与其他聚类算法相比，这是一个相对较高的值。为了阐明所选组的特征，结果中呈现了对时空、运动学和临床特征的深入统计。结果证明了基于端到端深度学习的聚类的有效性，在无需手动特征提取的情况下实现了显著的性能提升。虽然本研究主要利用矢状面数据，但未来纳入冠状面和横断面以及肌肉活动和步态对称性的分析可能会提供对步态模式更全面的理解。
