Comparison of different feature extraction and machine learning algorithms for EMG signal classification

Abstract

Motivation and Objectives Biological signals could not be modeled and processed easily because of their complex, sophisticated structure and mathematical formulation. However, recently, the algorithms and programs developed for the processing signals without the need for mathematical formulation rapidly led to an improvement of biological signals [Hudson & Cohen, 2000]. Although the initial period of Artificial Intelligence studies designed for pattern recognition, nowadays this process has become more parallel with both image processing techniques and pattern recognition. The development of artificial intelligence techniques gave an opportunity to the experts to enter input data into the results of computers. In the mid-1980s, artificial neural networks models have been an alternative of artificial intelligence. Decision trees are also used extensively for classification throughout the history. Although decision trees are the simplest machine learning algorithm, they also give as close accurate as the neural networks. Materials and Methods In this study the EMG data that are collected from 25 subjects were analyzed. Two separate groups of myopathy and ALS patients and a control group are the participants of the research. Both females and males are included in the all groups. In the control group, 6 males and 4 females are selected between the ages of 21-37. Among these participants 6 of 10 subjects have the best appearance, the rest of them except one are in general good shape. In the control group subjects, no one has had any neuromuscular disorders. 2 females and 5 males of 7 subjects are included in the group of myopathy patients. The ages of them are limited from 19 to 63. All of them have the symptoms of myopathy15 that are concerned as clinical and electrophysiological ones. 4 males and females aged 35-67 years are chosen for the ALS group of. 5 of the patients within ALS group died after a while of the diagnosis of ALS. The reason of why medial vastus muscles and the brachial biceps are used in this research is that they are common for both ALS and myopathy groups of patients. The EMG signals were recorded under usual conditions for MUAP analysis. The recordings were made at low (just above threshold) voluntary and constant level of contraction and visual and audio feedback was used to monitor the signal quality. The EMG signals were recorded from five places in the muscle at three levels of insertion (deep, medium, low). For these measurements a standard concentric needle electrode was used. The high and low pass filters of the EMG amplifier were set at 2 Hz and 10 kHz. Results We preprocessed the EMG signals and used autoregressive method (AR) and discrete wavelet method (DWT) for feature extraction. Features are applied to various classification algorithms, namely Multilayer perceptron, C 4,5, CART, K-NN, Random forest tree. All the data are compared each other on the table try to find out the best classification and feature extraction methods. When AR-based feature extraction method was used, success rates of the algorithms are as follows: MLP 51.3%, CART 73.2%, C 4.5 73.2%, K-NN 79.3%, RFT 81.3%. When DWT-based feature extraction method was used, success rates of the algorithms are as follows: MLP 81.2%, CART 83.3%, C 4.5 83.2%, K-NN 87.5%, RFT 89.7%.