On the Reuse of Motor Unit Filters in High Density Surface Electromyograms with Different Signal-to-Noise Ratios

Abstract

We test the reuse of motor unit (MU) filters estimated by the Convolution Kernel Compensation (CKC) method from high-density surface electromyograms (HDEMGs) with different signal-to-noise ratios (SNRs). During the learning phase the MU filters are extracted from HDEMGs with four different SNRs, namely $\infty$ dB, 30 dB, 20 dB and 10 dB. The MU filters are then applied to HDEMG signals at the different SNRs, yielding the MU spike trains. We report mean precision and miss rate of MU firing identification. In order to test the sensitivity of MU filter learning to the length of the HDEMG signals, we repeated the experiment at 5 s and 15 s long learning sets of HDEMG signals.

The number of identified MUs decreased from about 12 MUs, when using MU filters learned on 15 s long HDEMG signals with SNR of $\infty$ dB, to about 3 MUs, when using filters learned on HDEMG signals with SNR of 10 dB, no matter how much noise was present in the MU filter application phase. However, if there was no or little noise present in the MU filter learning phase then a decrease in precision and an increase in miss rate was observed when MU filter was applied to the HDEMG signals with a lot of noise. The opposite was true when large level of noise was present during the MU filter learning, but no or little noise was present in the MU filter application phase.