Incremental Adaptive EEG Classification of Motor Imagery-based BCI

Generally, electroencephalogram (EEG) signals recorded from the brain computer interface (BCI) systems are very noisy and non-stationary, which may affect the online performance of classifiers established from the prior session heavily. In order to address such problems, the classifiers should also be capable of adapting the change of EEG automatically during the processing of evaluation. In this paper, we propose an incremental adaptive EEG classification scheme. In this scheme, an extended sequential adaptive fuzzy inference system (ESAFIS) is used to evolve its structure dynamically and adapts the classifier automatically online to address the non-stationarity of the EEG signals. ESAFIS is an evolving system, wherein the fuzzy rules are evolved based on the modified influence of the rule. This paper presents the classification of 2-class motor imagery EEG based on ESAFIS with adaptive strategy. Simulations are conducted based on two datasets: one is the BCI Competition IV dataset 2b and the other one is recorded from our own BCI experiments. Compared to other methods such as ELM and LDA, the simulation results demonstrate that the proposed scheme produces better classification results.