Cluster-based algorithm for ROI analysis and cognitive state decoding using single-trial source MEG data

We present a cluster-based decoding algorithm for discovering regions of interest (ROIs) from EEG/MEG data in source space (or optimal cluster of sources) and predicting multiple conditions in a single experimental trial. Our algorithm automatically identifies contiguous brain regions that yield maximum mean test statistics from hypothesis tests over individual brain sources. We show that by utilizing these sources of interest it is possible to predict the experimental conditions via Bayesian decoding using single trials. Furthermore, we examine how these sources evolve spatially and temporally over the brain surface by running our algorithm at sequential time intervals. We demonstrate our method in a visually-guided wrist movement task where the goal is to infer the direction of wrist movement. MEG data were recorded using an Elekta Neuromag® device while subjects performed center-out wrist movements in 4 directions, and sources were estimated from raw data using the Minimum Norm Estimate method [1]. Our results demonstrate that the algorithm is a potential alternative to traditional ROI analysis using p-values, successfully discovers optimal clusters of sources relevant to the task, and achieves significant decoding accuracies in leave-one-out cross validation.