Toward Generalized Brain Computer Interface Using Supervised Hierarchical Bayesian Model

Abstract

This paper introduces a new attempt to implement a generalized brain computer interface (BCI) algorithm across multiple subjects by applying a supervised hierarchical Bayesian model. The model is motivated by a type of probabilistic topic model named supervised latent Dirichlet allocation (sLDA). Our method preserves the advantageous attributes of sLDA to simultaneously construct feature space of latent brain states which incorporates various brain signal patterns and conducts classification tasks based on the generated feature space. The proposed supervised hierarchical Bayesian model automatically establishes the overall process without manual feature selection or optimization. Such attribute of the model makes it approapriate for obtaining a generalized BCI model for the classification across multiple subjects. Two class motor imagery classification problem is tested with public BCI data sets for the evaluation of the proposed method. The model is trained with EEG data from multiple subjects altogether, and tested with the data from the trained subjects as well as new subjects. The experiment results indicate the feasibility of the proposed model