Non-negative Matrix Factorizations Based Spontaneous Electroencephalographic Signals Classification Using Back Propagation Feedback Neural Networks

Abstract

The paper proposes a new spontaneous EEG classification method for attention-related tasks. The algorithm was based on back propagation feedback neural network. Non-Negative Matrix Factorization (NMF) was used as a feature extraction tool. Six electrodes were selected from 32 international 10-20 electrode placement systems according to surface power distributing of EEG activity. Several experiments were carried out to decide an adaptive and robust structure of BP-ANN. The final structure of the NMF-ANN preserved the spatio-temporal characteristics of the signal. Simulation results showed that the averaged classification accuracy for designed three-level tasks can reach 98.4%, 86%, and 82.8%, which were better than other two reference methods.