Cluster analysis-based classification of healthy female netball players using wearable sensors

Abstract

Use of wearable wireless sensors (WWS) for classification of healthy female netball players is presented in this study. WWS comprised of wireless surface electromyography (EMG) sensors and 3Dimensional (3D) marker-based motion capture system for acquisition of lower extremity (LE) EMG data and 3DKinematics data respectively. Using WWS data obtained during ball interception (BI) task, subjects are classified based on their similarity-dissimilarity measure through hierarchical cluster analysis (HCA). By investigating existence of homogeneous subgroups (clusters) in LE features extracted, this work aimed to establish for the first time whether netball players exhibit identifiable and distinguishable EMG-3D Kinematic patterns during multiple trials of BI. BI is a key goal-oriented, often spontaneous, and multi-directional jump-landing task frequently performed by every player in a netball game. Thirteen professional subjects were recruited for this study with each asked to perform BI task in six trials in a semi-controlled game-play environment. EMG activity of eight LE muscles and 3D kinematics of the knee and ankle joints were recorded from each subject bilaterally during each BI trial. A total of sixty features (48 EMG and 12 3D-Kinematics) were extracted from the recorded raw data for analysis. Principal component analysis (PCA) was applied for dimensionality reduction of the total feature dataset, retaining only principal components that collectively explained more than 90% data variability. HCA was then used in clustering of the reduced datasets. Through inspection of the resulting dendrograms along with cophenetic correlation coefficients, 3 different clusters were confirmed. Based on HCA cluster-solutions, subjects were classified into three different classes (Class-1, Class-2, and Class-3) corresponding with respective clusters. Classification showed that majority (8 of the 13) subjects exhibited and maintained an identifiable LE biomechanical pattern 100% of the time (i.e for all six BI trials), while the remaining 5 subjects exhibited the same more than 66% of the time. Kruskal Walli's test showed that subgroups differed significantly (p<0.05) in their ranges of motion of the knee and ankle joints in sagittal and transverse planes, bilaterally. The integration of wearable wireless EMG sensors with motion capture system utilized in this research demonstrates that quantification of athletes' BI profiles based on their LE neuromuscular and 3D kinematics loadings is plausible. This allows trainers to make informed judgment on performance enhancement and injury prevention measures for BI task, both for individual athletes as well as for similar-groups as identified through HCA.