Estimation of single-trial evoked potential with RBFNN

Traditional technique used for estimating evoked potentials (EPs) is ensemble averaging, which not only requires repetitive experiments but also eliminates variations between trials. Trial-to-trial variability in EPs cannot be overlooked in clinical study, therefore EPs are expected to be estimated in single trial. Radial basis function neural network (RBFNN) has a powerful modeling capability in approximation. In this paper we present an EP estimation approach with this neural network. Using LMS approach, the network output will form a close approximation to the underlying EP signal. In the iterations of the self-learning process, pseudo data records are constructed by resampling the estimation error sequence and then adding it to the single-trial EP measurement and with these data sets, the estimation is re-computed. After convergence, the neural network output well estimates the underlying EP signals. This approach makes full use of single-trial record and tries to maximally approximate the underlying EP in a single trial. The experiment result shows that improved EP signal can be retrieved although the number of trials is reduced to 3.