EEG spatial decoding with shrinkage optimized directed information assessment

This paper proposes an approach to infer neural interactions from EEG data using a James-Stein estimator of directed information called shrinkage optimized directed information assessment (SODA). SODA uses shrinkage regularization on empirical histograms to deal with the high dimensionality of multi-channel EEG signals and the small sizes of many real-world datasets. It is designed to make few a priori assumptions, and can handle both non-linear and non-Gaussian flows across electrode sites. The use of James-Stein shrinkage allows the SODA algorithm to achieve higher sensitivity to directed neural interactions for a given specificity. We augment this through a central limit theorem-based approach that can assess the statistical significance of each discovered interaction. When evaluated on brain computer interface EEG motor activity data the neural decoding obtained using SODA outperformed several state-of-the-art approaches including Granger causality, MI, unregularized directed information, and spatial coherence. Our results show that SODA localizes 30% more directed interactions in regions that are consistent with Brodmann functional areas of motor activity.