Fast computation of scalp potentials in response to current sources within the brain using an artificial neural network

High-resolution EEG, in which data are recorded from tens or hundreds of electrodes simultaneously, has been widely applied to basic and clinical neuroscience. The data acquired can be utilized to computationally localize functional activity within the brain. In this approach the forward computation, which produces scalp potentials in response to assumed dipolar current sources within the brain, is a key problem. This paper investigates a novel computational algorithm utilizing a backpropagation artificial neural network (ANN). The ANN is pre-trained with thousands of numerically computed potential vector pairs. In each pair, one vector is computed by a fixed spherical head model and the other by a variable nonspherical head model. Once the training is complete, the ANN is able to map forward solutions between these two models, and the traditionally expensive forward computing based on a nonspherical head model of variable shape can be implemented at the low cost of the fixed spherical head model. Our experimental results show that this algorithm is highly accurate, and the improvement on computational speed is so great that real-time modeling of the high-resolution EEG can be realized