Evoked Potential Signal Estimation Using Gaussian Radial Basis Function Network

Abstract

Gaussian radial basis function neural networks are used to capture the functional mapping of the evoked potential (EP) signal buried in the additive electroencephalographic noise. The kernel parameters are obtained from the signal edges detected using direct spatial correlation at several adjacent scales of its undecimated non-orthogonal wavelet transform (WT). The segment of the data, where the WT is highly correlated across Scales, is considered a component region and its width is employed as the variance of the Gaussian kernel placed at the center of the region. The weights of the kernels are computed using gradient descent algorithm. Results obtained for both simulated and real brainstem auditory EPs show the superior performance of the technique. Because the technique incorporates signal knowledge into the network design, the number of hidden nodes reduces, and a more accurate estimation of signal components ensues.