Classification of EEG signals for brain-computer interfaces using a Bayesian-Fuzzy Extreme Learning Machine

In brain-computer interface (BCI) applications, classification of motor imagery electroencephalogram (EEG) using Extreme Learning Machine (ELM) theory dates back to 2006. Even though, it is relatively new, advances in ELM-based classification have demonstrated to be a robust method-ology with strong generalization properties. In this study, a unified framework based on Bayesian and Fuzzy ELM theory referred to as Bayesian-Fuzzy Extreme Learning Machine (BFELM) is developed for EEG signals classification. The proposed methodology is a hybrid approach for the training of a class of Fuzzy Inference Systems (FISs) of Takagi-Sugeno-Kang (TSK). On the one hand, Fuzzy logic theory is applied to handle any bounded non-constant piecewise continuous membership functions (MFs). On the other hand, Bayesian ELM theory is used to calculate the consequent parameters of each fuzzy rule by estimating their likelihood while minimizing training error and improving associated model generalization. Performance comparison of BFELM with other existing ELM methods and Support Vector Machine (SVM) is implemented for the classification of EEG signals using two public data sets. The experimental results confirm the advantages of using a unified framework for an improved classification of EEG data associated with motor imagery (MI) in BCI applications.