On the Need for Spatial Whitening of High-Density Surface Electromyograms in Motor Unit Identification by Neural Networks

Abstract

AbstractIn recent years, many new decomposition methods for identification of motor unit (MU) firings from high-density surface electromyograms (HDEMG) have been developed, with recent attempts focused on the use of different neural networks (NN). In this study, we evaluated the need for HDEMG signal whitening in NN-based MU identification. For this purpose, we analyzed the learning efficiency of two different types of NN, namely dense NN and long short-term memory (LSTM) NN, on the same HDEMG signals, with and without spatial whitening applied to them. All the HDEMG signals used were simulated with advanced HDEMG simulator, providing a full control of MU firing patterns and MU characteristics in our test environment. Spatial whitening of HDEMG signals significantly improved the precision of MU identification, regardless of the type of NN tested. For dense NN, precision of identified MU increased from 32.2 ± 20.2% to 93.1 ± 8.7%, whereas miss rate decreased from 48.4 ± 23.9% to 12.0 ± 13.3% when whitening of HDEMG signals was employed. For LSTM NN the precision of MU identification increased from 59.7 ± 19.7% to 99.4 ± 2.0% whereas miss rate decreased from 43.1 ± 22.3% to 12.7 ± 9.7% with whitening.KeywordsNeural networksHigh-density surface electromyographyMotor unit identificationDense neural networksLong short-term memory neural networks