Probabilistic onset detection based on multiple strides

Abstract

Introduction: Due to the high variability of muscle activity collected using surface electromyography (sEMG) during gait analysis, most automatic algorithms for determining muscle on-/offset basedon these signals canbe tuned toperformwell for a single trial, but will then fail for others using the same settings. This is caused by a fundamental issue: most algorithms only use information of a single stride, which is subject to a high variability [1]. The aim of our study was to improve the usability of automatic sEMG onset detection using a probabilistic representation of muscle activation. Patients/materials and methods: Instead of a binary on-/off representation on a stride-by-stride basis, the concept probability of onset is introduced. This probability was computed in three steps: (1) determining themuscle activation stride-by-stride using two algorithms: thresholdmethod [2] and Staudesmethod [3] (Fig. A, using Staude’s method); and time-warping these signals to align the toe-off event for all strides (Fig. B); as well as determining the probability of onset by counting across all strides (Fig. C). The conventional and the proposedmethods were compared to the golden standard of manual determination by multiple experts. As performance figure, the number of correct classified points was used. Firstly, this was calculated on a strideby-stride base comparing the outcome of both methods with our golden standard. Secondly, a calculation on the averaged signals on a muscle-by-muscle base was made after a threshold for onset was put at 0.5. These results were also compared to the experts opinions on a stride-by-stride base.