Abstract
Machine learning (ML) techniques such as (deep) artificial neural networks(DNN) are solving very successfully a plethora of tasks and provide newpredictive models for complex physical, chemical, biological and socialsystems. However, in most cases this comes with the disadvantage of acting as ablack box, rarely providing information about what made them arrive at aparticular prediction. This black box aspect of ML techniques can beproblematic especially in medical diagnoses, so far hampering a clinicalacceptance. The present paper studies the uniqueness of individual gaitpatterns in clinical biomechanics using DNNs. By attributing portions of themodel predictions back to the input variables (ground reaction forces andfull-body joint angles), the Layer-Wise Relevance Propagation (LRP) techniquereliably demonstrates which variables at what time windows of the gait cycleare most relevant for the characterisation of gait patterns from a certainindividual. By measuring the timeresolved contribution of each input variableto the prediction of ML techniques such as DNNs, our method describes the firstgeneral framework that enables to understand and interpret non-linear MLmethods in (biomechanical) gait analysis and thereby supplies a powerful toolfor analysis, diagnosis and treatment of human gait.