Hierarchical bayesian modeling for EEG-based brain-computer interfaces

Abstract

Brain-computer Interfaces (BCI) have been receiving much attention for its distinctive nature of utilizing signals from the brain for communication and control. This paper introduces a new attempt to process brain EEG signals for BCI by proposing a supervised hierarchical Bayesian model that jointly models the latent brain states and class labels of EEG. We assume a generative process of EEG and describe how the proposed method can model latent brain states and infer class labels given EEG; our method infers probability distribution over latent brain states of input EEG. We present evaluation of the proposed method by applying multi-class motor imagery classification tasks to the model. The results indicate that our approach yield competitive classification accuracy rate while simultaneously modeling underling brain states and conducting classification tasks. Discussion over the feasibility of our model as a mean for BCI is extensively addressed in the context. We believe that our method presents a potential in helping people with little or no voluntary muscular control by providing them with an effective BCI to communicate and control electrical devices.