Towards personalized gait rehabilitation: How robustly can we identify personal gait signatures with machine learning?

Abstract

Personalizing gait rehabilitation requires a comprehensive understanding of the unique gait characteristics of an individual patient, i.e., personal gait signature. Utilizing machine learning to classify individuals based on their gait can help to identify gait signatures [1]. This work exemplifies how an explainable artificial intelligence method can identify the most important input features that characterize the personal gait signature. How robust can gait signatures be identified with machine learning and how sensitive are these signatures with respect to the amount of training data per person? We utilized subsets of the AIST Gait Database 2019 [2], the GaitRec dataset [3], and the Gutenberg Gait Database [4] containing bilateral ground reaction forces (GRFs) during level walking at a self-selected speed. Eight GRF samples from each of 2,092 individuals (1,410/680 male/female, 809/1,283 health control/gait disorder, 1,355/737 shod/barefoot) were used for a gait-based person classification with a (linear) support vector machine (SVM). Two randomly selected samples from each individual served as test data. Gait signatures were identified using relevance scores obtained with layer-wise relevance propagation [5]. To assess the robustness of the identified gait signatures, we compared the relevance scores using Pearson’s correlation coefficient between step-wise reduced training data, from k=6 to k=1 training samples per individual. For the baseline setup (k=6), the SVM achieved a test classification accuracy of 99.1% with 36 out of 4184 test samples being misclassified. The results for the setups with reduced training samples are visualized in Fig. 1. Fig. 1: Overview of the experimental results.Download : Download high-res image (210KB)Download : Download full-size image A reduction of training samples per individual causes a decrease in classification accuracy (e.g., by 17.7% in the case of one training sample per individual). The results show that at least five training samples per individual are necessary to achieve a classification accuracy of approximately 99% for over 2,000 individuals. A similar effect is observed for gait signatures, which also show a slight degradation in robustness as the number of training samples decreases. In some cases, a model trained with less data per individual learns a different gait signature than a model trained with more data. In the test sample with the lowest correlation (see Fig. 1E), we observe a significant deviation in relevance for some input features. However, only 114 test samples (2.7%) are below a moderate correlation of r=0.4 [6], indicating that gait signatures are quite robust, even when using one training sample per individual. This is supported by a strong median correlation of r=0.71 [6] (and the highest correlation of r=0.96) between the gait signatures. As automatically identified gait signatures seem to be robust, this approach has the potential to serve as a basis for tailoring interventions to each patient’s specific needs.