Analyze EEG signals with extreme learning machine based on PMIS feature selection

An electroencephalogram (EEG)-based brain–computer interface (BCI) system can be used to determine the intents for a paralyzed user by analyzing the EEG signals recorded from his scalp. The key technology of BCI systems is feature extraction and pattern recognition algorithm. Many of the BCI research show that nonlinear classification algorithms perform better than the linear, but they are usually far slower than required. Besides, many of these methods employ less significant features and could make the classification process less efficient. In this paper, a method is provided to identify different patterns of EEG signals. A fast nonlinear classification algorithm extreme learning machine (ELM) is provided to identify two different EEG signals. After feature extraction, we adopt a nonlinear feature selection algorithm partial mutual information-based feature selection (PMIS) to eliminate the less significant features. The experimental result shows that (1) ELM can discriminate the two patterns of EEG signals with a relatively satisfactory accuracy and (2) PMIS feature selection can find out the important features effectively and improve the classification accuracy. The final accuracy is 91.5 % against the best existed result of 88.7 % for the tested data set. The classification performance may be further improved by adding complementary features from EEG. This approach may eventually lead to a reliable EEG-based BCI studies.