Multi-joint gait clustering for children and youth with diplegic cerebral palsy

Abstract

Background: Clinical management of children and youth with cerebral palsy (CP) is increasingly supported by computerized gait analysis. Methods have been developed to reduce the complexity of interpreting biomechanical data and quantify meaningful movement patterns. However, few methods are inclusive of multiple joints and planes of motion, and consider the entire duration of gait phases; potentially limiting insight into this heterogeneous pathology. The objective of this study was to assess the implementation of k-means clustering to determine clusters of participants with CP based on multi-joint gait kinematics. Methods: Barefoot walking kinematics were analyzed for a historical cohort (2007-2015) of 37 male and female children and youth with spastic diplegic CP [male n = 21; female n = 16; median age = 12 (range 5-25) years; Gross Motor Function Classification System Level I n = 17 and Level II n = 20]. Mean stance phase hip (sagittal, coronal, transverse), knee (sagittal), and ankle (sagittal) kinematics were time (101 data points), mean and range normalized. Normalized kinematics data vectors (505 data points) for all participants were then combined in a single data matrix M (37x505 data points). K-means clustering was conducted 10 times for all data in M (2-5 seeds, 50 repetitions). Cluster quality was assessed using the mean Silhouette value (s) and cluster repeatability. The mean kinematic patterns of each cluster were explored with respect to a dataset of normally developing (ND) children using Statistical Parametric Mapping (SPM, alpha 0.05). Differences in potentially confounding variables (age, height, weight, walking speed) between clusters (C) were assessed individually in SPSS (IBM, USA) using Kruskal-Wallis H tests (alpha 0.05). Results: Four clusters (n 1 = 5, n 2 = 12, n 3 = 12, n 4 = 8) provided the largest possible data separation based on high cluster repeatability (96.8% across 10 repetitions) and comparatively greater cluster quality [s (SD), 0.275 (0.152)]. Participant data with low cluster quality values displayed a tendency toward lower cluster allocation repeatability. Distinct kinematic differences between clusters and ND data were observable. Specifically, C1 displayed a unique continuous hip abduction and external rotation pattern. In contrast, participants in C2 moved from hip adduction (loading response) to abduction (mid to terminal stance) and featured a unique ankle plantarflexor pattern during pre-swing. C3 was characterized by gait deviations in the sagittal plane of the hip, knee and ankle only. C4 displayed evidence for the most substantial hip and knee extension, and ankle plantarflexion deficit from midstance to preswing. Discussion: K-means clustering enabled the determination of up to four kinematic clusters of individuals with spastic diplegic CP using multi-joint angles without a priori data reduction. A cluster boundary effect was demonstrated by the Silhouette value, where data with values approaching zero were more likely to change cluster allocation. Exploratory analyses using SPM revealed significant differences across joints and between clusters indicating the formation of clinically meaningful clusters. Further work is needed to determine the effects of including further topographical classifications of CP, additional biomechanical data, and the sensitivity to clinical interventions to assess the potential for informing clinical decision-making.