Evoked potentials are usually embedded in the ongoing electroencephalogram with a very low signal-to-noise ratio. The neural network filtering technique which has the advantage of complex mapping is one of the applicable methods for evoked potentials estimation. The back-propagation algorithm based on second order statistics is commonly used to adapt neural network filters. However it is easily influenced by additive Gaussian noise. In this study, a neural network filter with a modified back-propagation algorithm for higher order statistics was proposed. With higher-order statistics technique, additive Gaussian noise is suppressed to improve the performance of evoked potentials estimation.