Real-time detection of signals in noise using normalized RBF neural network

The problem of denoising single-trial visual evoked potentials (VEP) is investigated in this contribution. The aim of single trial VEP's estimation is to recover the amplitude and the latency from the raw VEP without losing the individual properties of each epoch. This work is really meaningful to clinical practice. For this purpose, normalized radial basis function neural network (RBFNN) is developed to detect the single trial of VEP. The method is compared with two other nonlinear methods: the adaptive noise cancellation with RBFNN prefilter (ANC-RBFNN) and the RBFNN. The performances are compared with MSE and the ability to track peaks of the individual VEP. The proposed method provides convergent evidence that NRBFNN can effectively depress the noise and successfully detect each trial. Also the method provides a robust estimation for a wide range of VEP's. The recovery ability is better than ANC-RBFNN and RBFNN methods. Finally, both simulations and real VEP signal analysis are tested, showing the applicability and the effectiveness of the proposed algorithm.