The Performance of Onset Detection Methods for Surface Electromyographic Data

Abstract

The performance of five different algorithm-based onset detection methods for surface electromyography (SEMG) data was analyzed relative to visually determined onset times from three expert volunteers. It was hypothesized that at least one algorithm would out-perform the others in terms of degree of correctness.  Three-hundred data plots from three previous studies on motor control were selected as source data. The automated algorithms tested included: i) a forward-moving sliding window, considering window mean value and number of points above a threshold [12], ii) a backward-moving sliding window starting at the data peak, considering window mean value compared to a threshold (developed by authors), iii) a backward-moving sliding window starting at the highly smoothed data peak, considering window mean value and number of points below a threshold (developed by authors), iv) a forward-moving sliding window, considering window mean value and number of points above a threshold lasting for a minimum number of consecutive windows [13], and v) a system based on log likelihood ratios and maximum likelihood estimation [11]. Each method went through parameter optimization as part of the testing. The three expert volunteers determined onset times by visual inspection for the 300 data plots on two occasionsforeachoftwofiltermethods:a2ndorder Butterworth filter with 6 Hz cut-off frequency, and a 20 ms sliding window root-mean-square (RMS) filter. The algorithms were ranked based on i) a Z value, expressing the average of each data plot’s deviation from the visually determined onset distribution and ii) having no more than 10% erroneous results. The forward moving sliding window, considering window mean value and number of points above a threshold [12], resulted in the most accurate determination of onset time, with an RMS averaged Z value of 0.7738. However, the algorithms were comparable and several had unique advantages.