Comparison of AR parametric methods with subspace-based methods for EMG signal classification using stand-alone and merged neural network models

This research introduces an electromyogram (EMG) pattern classification of individual motor unit action potentials (MUPs) from intramuscular electromyographic signals. The presented technique automatically classifies EMG patterns into healthy, myopathic, or neurogenic categories. To extract a feature vector from the EMG signal, we use different autoregressive (AR) parametric methods and subspace-based methods. The proposal was validated using EMG recordings composed of 1200 EMG patterns obtained from 7 healthy, 7 myopathic, and 13 neurogenic-disordered people. A feedforward error backpropagation artificial neural network (FEBANN) and combined neural network (CNN) were used for classification, where the success rate was slightly higher in CNN. Among the different AR and subspace methods used in this study, the highest performance was obtained with the eigenvector method. The following rates were the results achieved by using the CNN. The correct classification rate for EMG patterns was 97% for healthy, 93% for myopathic, and 92% for neurogenic patterns. The obtained accuracy for EMG signal classification is approximately 94% for CNN. The rates for FEBANN were as follows: 97% for healthy patterns, 92% for myopathic patterns, and 91% for neurogenic patterns. The obtained accuracy was 93.3%. By directly using raw EMG signals, EMG classifications of healthy, myopathic, or neurogenic classes are automatically addressed