Census transform based feature extraction of EMG signals for neuromuscular disease classification

This paper aims to propose a systematic feature extraction technique for classification of Neuromuscular diseases. In this proposed scheme, modified CENTRIST (CENsus TRansform hISTogram) for 1-dimensional signals is introduced as the feature descriptor for the given Electromyography (EMG) signals from two different publically available databases. The proposed algorithm encodes discriminative structural properties of the EMG signal of different classes and is vigil in accounting abrupt transitions in the signal. K-nearest neighborhood (KNN) and Support Vector Machine (SVM) classifiers are employed for the purpose of classification. The crux of this paper is to formulate a feature extraction technique to distinguish the structural and topological properties of Normal, Myopathic, Neuropathic and Amyotrophic Lateral Sclerosis (ALS) EMG signals. The classification performance was experimentally verified for a 2-class problem and a 3-class problem using standard databases and the results revealed 95% classification accuracy.