Individuality decoded by running patterns: Movement characteristics that determine the uniqueness of human running

Abstract

Human gait is as unique to an individual as is their fingerprint. It remains unknown, however, what gait characteristics differentiate well between individuals that could define the uniqueness of human gait. The purpose of this work was to determine the gait characteristics that were most relevant for a neural network to identify individuals based on their running patterns. An artificial neural network was trained to recognize kinetic and kinematic movement trajectories of overground running from 50 healthy novice runners (males and females). Using layer-wise relevance propagation, the contribution of each variable to the classification result of the neural network was determined. It was found that gait characteristics of the coronal and transverse plane as well as medio-lateral ground reaction forces provided more information for subject identification than gait characteristics of the sagittal plane and ground reaction forces in vertical or anterior-posterior direction. Additionally, gait characteristics during the early stance were more relevant for gait recognition than those of the mid and late stance phase. It was concluded that the uniqueness of human gait is predominantly encoded in movements of the coronal and transverse plane during early stance.

Fig 1. Classification accuracies of the trained neural network on unseen data, stratified by participants and intervention groups.

Fig 2. Absolute relevance of each variable within a stride pattern averaged across all relevance patterns.

The top part (A) shows the summed contribution of relevance for each of the 100 time points of stance. In the center (B), darker colors indicate variables of high relevance, while lighter colors indicate variables of low relevance. In other words, to assign a stride pattern to the respective participant, the model relied more on variables with darker shades. Variables with lighter shades were less relevant for a correct classification of gait patterns. The right part of the figure (C) highlights the summed contribution of relevance of each direction of joint angle trajectories.

Fig 3. Average classification accuracy as a function of the number of highly relevant variables used for the classification.

Fig 4. Absolute relevance of the 200 variables with the highest relevance within a stride pattern averaged across all relevance patterns.