Optimization of features to classify upper-limb movements through sEMG signal processing

This paper presents the development of a computational intelligence method based on Regularized Logistic Regression to classify 17 distinct upper-limb movements through surface electromyography (sEMG) signal processing. The choose of the tuning parameters of the regularization and the generation of the different classification methods are presented. For the different models were used variations involving 12 sEMG channels and the Root Mean Square (RMS), Variance and Medium Frequency features with which we proposed to achieve a most proper combination of parameters to perform the movements classification. The tests involved 50 subjects, including 10 amputees, using the NinaPro database and also a database currently on development by the authors. The global mean accuracy rate considering all the subjects and the channel and features variations was 70,2% prior the definition of the best case scenario. Once defined the most proper features combination, the accuracy rate reached 87,1%, raising the rates of all movements accuracies performed for all databases