Online Tracking of Canonical Brain Network Activation and Behavioral Prediction Using Bayesian Filtering

Online inference of the activation of a group of brain regions or networks may prove useful for developing adaptive and closed-loop brain-machine (BMI) systems. In this paper, we applied a two-stage Block Sparse Bayesian Learning (BSBL-2S) method to electroencephalographic (EEG) data for online tracking of the relative evidence of activation of groups of cortical regions of interest (ROIs) comprising the nodes of two brain networks. Bayes factors (BF) are presented as measures of relative evidence for brain network activity. A second Kalman filter is used to predict behavior based on Bayes factors. The method is evaluated with simulated and real EEG data. EEG data was simulated by smoothly modulating the relative activations of canonical task positive network (TPN) and default mode networks (DMN). The calculated BFs (TPN/DMN) identified and tracked the changes in relative TPN and DMN network activation and accurately predicted simulated behavioral measures (reaction times to experiment events). The method was applied to EEG data from 8 subjects performing a lane-keeping simulated driving task in which the subjects were instructed to quickly respond to random perturbations of their vehicle's lane position. The estimated BFs showed significant linear and quadratic predictive relationships with response time (RT), with poorest task performance assocated with decreasing BF (increased DMN activation, relative to TPN). These results suggest that online estimation of BF for cortical network activation could be useful in the development of neuroscientifically grounded adaptive BMI systems.