A Multitask Bayesian Framework for the analysis of Motor Imagery EEG Data

Motor Imagery EEG patterns have been found to vary across subjects, datasets or even for the same subject across sessions. This affects considerable the performance of a BCI system requiring expensive and time-consuming individual calibration sessions to adapt to new subjects. Our work tackles the aforementioned problem using the Bayesian multi-task learning (MTL) framework to share common information across subjects, sessions and datasets. The proposed frame-work finds a balance between universal (subject-independent) classifiers and subject-specific classifiers by using data from other subjects (even datasets) and combine them to estimate the classifier parameters for the target subject. This combination is achieved by selecting appropriate prior and hyper-prior distributions for the Bayesian MTL framework. Experimental results in three widely used Motor Imagery datasets shown that the proposed MTL framework presents superior performance compared to other state-of-the art methods.