Multiscale SVD entropy for the analysis of gait dynamics

Abstract

This work explores the use of singular value decomposition (SVD) entropy to characterize the complexity of stride interval sequences from gait dynamics. The computation of the approximate (ApEn) and sample (SampEn) entropies require the specification of the length of the scrutinized patterns (m) and the pattern similarity tolerance (r). In contrast, the SVD entropy estimation does not require these parameters since the computations are based on the inspection of similarities among lagged vectors. In this regard, the present work aims to assess the SVD entropy to characterize gait dynamics. The gait maturation and the neurodegenerative disease (NDD) datasets from Physionet were considered. The SVDEn of the gait dynamics decreased with the age to normalized values of about 0.2. In contrast, the SampEn, which was used for comparison, decreased to values close to zero. For the gait NDD dataset, it was found that the gait dynamics of Parkinson, ASL and Huntington pathologies exhibit increased SVDEn relative to a control group, with normalized values of about 0.2 and 0.4, respectively. However, the SVDEn and the SampEn were unable to distinguish between the control and neurodisease groups (p < 0.05). Overall, the results indicated that the gait dynamics of the NDD are less regular than that of the control group. Although the SVDEn and other entropy metrics (e.g., SampEn) showed similar results, the SVDEn offers the advantage that the computation require the specification of only the scale parameter.