ART2 Neural Network for Surface EMG Decomposition

Abstract

Extraction of individual Motor Unit Action Potentials (MUAPs) from a surface ElectroMyoGram (EMG) is an essential but challenging task for clinical study and physiological investigation. This paper presents an automatic decomposition of surface EMGs using a self-organised ART2 neural network. In our approach, MUAP peaks are first detected using a Weighted Low-Pass Differential (WLPD) filter. A modified ART2 network is then utilised to classify MUAPs based on MUAP waveforms and firing time information. Individual MUAP trains are identified from real surface EMG signals recorded during weak contraction, and also from simulated surface EMGs. The firing statistics and the waveforms of individual MUAPs are then extracted. A number of computer tests on 50 simulated and real surface EMGs of limb muscles show that up to five MUAP trains can be effectively extracted, with their waveforms and firing parameters estimated. Being able to decompose real surface EMGs has essentially demonstrated the potential applications of our approach to the non-invasive diagnosis of neuromuscular disorders.