An adaptive RBF neural network model for evoked potential estimation

A method for evoked potential estimation based on an adaptive radial basis function neural network (RBFNN) model is presented in this paper. During training, the number of hidden nodes (number of RBFs) and model parameters are adjusted to fit the target signal which is obtained by averaging. In order to reduce computational complexity and the influence of noise in estimating single-trial evoked potential (EP), the number of hidden nodes is also minimized in training. After training, both peak latency and amplitude, being distinctive features of an EP, are characterized by center and height of the corresponding RBF respectively. In EP estimation, an adaptive algorithm is employed to track the peaks from trial to trial by adapting the center and height of RBFs directly. The adaptive RBFNN is tested on a computer simulated data set and clinical EP recording. Our proposed algorithm is suitable for tracking EP waveform variations