Denoising of Single-Trial Event-Related Potentials by Shrinkage and Phase Regularization of Analytic Wavelet Filterbank Coefficients

Abstract

Event-related potentials (ERP) provide reliable electrophysiological correlates of subsequent neural processing following sensory stimulation, offering insight into the activation patterns of participating neural structures which is of considerable value in both neuroscience research and clinical applications. 2D single-trial representations as ERP images have seen increased application in recent studies, accompanied by a rising number of approaches to improve their signal-to-noise ratio, which for the most part have been motivated from an image processing point of view (e.g., nonlocal operators, anisotropic diffusion filtering). In this paper, a brief overview of ERP image denoising prior art is given and a novel, fast denoising algorithm based on split amplitude and phase processing (i.e., phase-informed amplitude shrinkage and regularization of the phase structure) in analytic time-frequency representations of ERP single trials obtained using a perfect reconstruction wavelet filterbank is proposed. Furthermore, the performance of the proposed algorithm is subjected to a comparative evaluation using real-world chirp-evoked auditory ERP acquired from 20 normal hearing adults. Results suggest the suitability of the proposed method for a broad range of a posteriori ERP image denoising tasks, including those lacking a priori knowledge about the shape of potentially nonstationary traces in the ERP image due to, e.g., endogeneous states gradually changing during the experiment.