Kuch Andrian, Schweighofer Nicolas, Finley James M, McKenzie Alison, Wen Yuxin, Sánchez Natalia
Department of Physical Therapy, Chapman University, Irvine, CA.
Division of Biokinesiology and Physical Therapy, University of Southern California, Los Angeles, CA.
bioRxiv. 2024 Oct 30:2024.10.28.620665. doi: 10.1101/2024.10.28.620665.
Heterogeneous types of gait impairment are common post-stroke. Studies used supervised and unsupervised machine learning on discrete biomechanical features to summarize the gait cycle and identify common patterns of gait behaviors. However, discrete features cannot account for temporal variations that occur during gait. Here, we propose a novel machine-learning pipeline to identify subgroups of gait behaviors post-stroke using kinematic time series data.
We analyzed ankle and knee kinematic data during treadmill walking data in 39 individuals post-stroke and 28 neurotypical controls. The data were first input into a supervised dual-stage Convolutional Neural Network-Temporal Convolutional Network, trained to extract temporal and spatial gait features. Then, we used these features to find clusters of different gait behaviors using unsupervised time series k-means. We repeated the clustering process using 10,000 bootstrap training data samples and a Gaussian Mixture Model to identify stable clusters representative of our dataset. Finally, we assessed the kinematic differences between the identified clusters using 1D statistical parametric mapping ANOVA. We then compared gait spatiotemporal and clinical characteristics between clusters using one-way ANOVA.
We obtained five clusters: two clusters of neurotypical individuals (C1 and C2) and three clusters of individuals post-stroke (S1, S2, S3). C1 had kinematics that resembled the normative gait pattern. Individuals in C2 had a shorter stride time than C1. Individuals in S1 had mild impairment and walked with increased bilateral knee flexion during the loading response. Individuals in S2 had moderate impairment, were the slowest among the clusters, took shorter steps, had increased knee flexion during stance bilaterally and reduced paretic knee flexion during swing. Individuals in S3 had mild impairment, asymmetric swing time, had increased ankle abduction during the gait cycle and reduced dorsiflexion bilaterally during loading response and stance. Every individual was assigned to a cluster with a cluster membership likelihood above 93%.
Our results indicate that joint kinematics in individuals post-stroke are distinct from controls, even in those individuals with mild impairment. The three subgroups post-stroke showed distinct kinematic impairments during specific phases in the gait cycle, providing additional information to clinicians for gait retraining interventions.
步态障碍的异质性类型在中风后很常见。以往研究使用有监督和无监督机器学习方法处理离散生物力学特征，以总结步态周期并识别常见的步态行为模式。然而，离散特征无法解释步态过程中发生的时间变化。在此，我们提出一种新颖的机器学习流程，使用运动学时间序列数据来识别中风后的步态行为亚组。
我们分析了39名中风后个体和28名神经典型对照在跑步机行走数据期间的踝关节和膝关节运动学数据。数据首先输入到一个有监督的双阶段卷积神经网络 - 时间卷积网络中进行训练，以提取时间和空间步态特征。然后，我们使用这些特征通过无监督时间序列k均值法来寻找不同步态行为的聚类。我们使用10000个自助训练数据样本和高斯混合模型重复聚类过程，以识别代表我们数据集的稳定聚类。最后，我们使用一维统计参数映射方差分析评估已识别聚类之间的运动学差异。然后，我们使用单因素方差分析比较聚类之间的步态时空特征和临床特征。
我们获得了五个聚类：两个神经典型个体聚类（C1和C2）以及三个中风后个体聚类（S1、S2、S3）。C1的运动学类似于正常步态模式。C2中的个体步幅时间比C1短。S1中的个体有轻度损伤，在负重反应期间双侧膝关节屈曲增加。S2中的个体有中度损伤，是聚类中最慢的，步幅较短，双侧站立期间膝关节屈曲增加，摆动期间患侧膝关节屈曲减少。S3中的个体有轻度损伤，摆动时间不对称，在步态周期中踝关节外展增加，负重反应和站立期间双侧背屈减少。每个个体都被分配到一个聚类成员可能性高于93%的聚类中。
我们的结果表明，中风后个体的关节运动学与对照组不同，即使是那些轻度损伤的个体。中风后的三个亚组在步态周期的特定阶段表现出明显的运动学损伤，为临床医生进行步态再训练干预提供了额外信息。
