Biomechanics has entered an era of ’big data’, where human locomotion data is generated rapidly and in large quantities. This poses new challenges because conventional analysis methods are ineffective in the presence of the myriad of interacting variables that describe human locomotion in contemporary datasets. Consequently, there is a need for novel, more appropriate analysis methods that can help direct attention towards the ’right’ variables.

The purpose of this thesis was to apply layer-wise relevance propagation, a novel analysis method, within the context of human locomotion to isolate variables from large datasets that describe highly relevant movement characteristics that may inspire and direct future biomechanics research. Specifically, the challenges of separating unique / generic movement characteristics of runners, and forming functional groups based on runner-specific movement adaptations (induced by footwear interventions) were addressed. It was shown that (1) unique locomotion characteristics of novice runners were best described by variables belonging to the movement trajectories of the coronal and transverse plane during early stance within a dataset of lower extremity kinematics and ground reaction forces. Further, (2) kinematic variables that describe unique locomotion characteristics of highly trained runners were associated with movements of the spine and lower extremities during mid-stance and mid-swing, while generic locomotion characteristics were associated with sagittal plane movements of the spine during early and late stance within a dataset of full body kinematics. Finally, (3) groups of runners who adopt similar kinematic movement responses to a given footwear intervention were identified by clustering relevance patterns of subject-specific artificial neural network models.

Based on the presented findings it was concluded that layer-wise relevance propagation is a promising analysis method that can help direct a researcher’s focus towards those variables that are most relevant for black-box machine learning models such as artificial neural networks. It, therefore, addresses many of the emerging challenges that biomechanics research faces during the contemporary era of ’big data’.