Investigating the use of multi-label classification methods for the purpose of classifying electromyographic signals

Abstract: The type of pattern recognition methods used for controlling modern prosthetics, referred to here as single-label classification methods, restricts users to a small amount of movements. One prominent reason for this is that the accuracy of these classification methods decreases as the number of allowed movements is increased. In this work a possible solution to this problem is presented by looking into the use of multi-label classification for classifying electromyographic signals. This was accomplished by going through the process of recording, processing, and classifying electromyographic data. In order to compare the performance of multi-label methods to that of single-label methods four classification methods from each category were selected. Both categories were then tested on their ability to classify finger flexion movements. The most commonly tested set of movements were the thumb, index, long, and ring finger movements in addition to all the possible combinations of these four fingers. The two categories were also tested on their ability to learn finger combination movements when only individual finger movements were used as training data. The results show that the tested single- and multi-label methods obtain similar classification accuracy when the training data consists of both individual finger movements and finger combination movements. The results also show that none of the tested single-label methods and only one of the tested multi-label methods, multi-label rbf neural networks, manages to learn finger combination movements when trained on only individual finger movements.