Supervised learning techniques and their ability to classify a change of direction task strategy using kinematic and kinetic features

This study examines the ability of commonly used supervised learning techniques to classify the execution of a maximum effort change of direction task into predefined movement pattern as well as the influence of fuzzy executions and the impact of selected features (e.g. peak knee flexion) towards classification accuracy. The experiment utilized kinematic and kinetic data from 323 male subjects with chronic athletic groin pain. All subjects undertook a biomechanical assessment and had been divided previously into 3 different movement strategies in an earlier paper. Examined supervised learning techniques were: a decision tree, an ensemble of decision trees, a discriminant analysis model, a naive Bayes classifier, a k-nearest-neighbour model, a multi-class model for support vector machines, a stepwise forward regression model, a neural network and a correlation approach. Performance (measured by comparing the predefined and classified movement pattern) was highest for the correlation approach (82% - CI 81-83%) and support vector machine (80% - CI 79-80%). The percentage of fuzzy observations within the data was between 16 and 25%. The most informative features for classification were: hip flexion and ankle rotation as well as ankle flexion moment, thorax [flexion and frontal sway], abduction angle in [hip and pelvis] and hip rotation. Findings of this study support the assumption that multiple patterns are used to execute a movement task and demonstrate that classification models can predict movement patterns with a high accuracy (~84%).