Bayesian estimation of ERP components from multicondition and multichannel EEG.

Extraction and separation of functionally different event-related potentials (ERPs) from electroencephalography (EEG) is a long-standing problem in cognitive neuroscience. In this paper, we propose a Bayesian spatio-temporal model for estimating ERP components from multichannel EEG recorded under multiple experimental conditions. The model isolates the spatially and temporally overlapping ERP components by utilizing their phase-locking structure and the inter-condition non-stationarity structure of their amplitudes and latencies. Critically, unlike in previous multilinear algorithms, the non-phase-locked background EEGs are modeled as spatially correlated and non-isotropic signals. A variational algorithm was developed for approximate Bayesian inference of the proposed model, with the effective number of ERP components automatically determined as a part of the algorithm. The utility of the algorithm is demonstrated with applications to synthetic data and the EEG data collected from 13 subjects during a face inversion experiment. The results show that our algorithm more accurately and reliably estimates the spatio-temporal patterns, amplitudes, and latencies of the underlying ERP components in comparison with several state-of-the-art algorithms.