Low-probability event-detection and separation via statistical wavelet thresholding: an application to psychophysiological denoising

Abstract

Objectives: The aim of this paper is to introduce and test a general, wavelet-based method for the automatic removal of noise and artefact from psychophysiological data.

Methods: Statistical wavelet thresholding (SWT) performs blind source separation by transforming data to the wavelet domain, and subsequent filtering of wavelet coefficients based on a statistical framework. The observed wavelet coefficients are modelled using a Gaussian distribution, from which low-probability outliers are attenuated based on their z-scores.

Results: The technique was applied to both simulated and real event-related potentials (ERP) data. SWT applied to artificial data displayed increased signal-to-noise ratio (SNR) improvements as noise amplitude increased. ERP averages of filtered experimental data displayed a correlation of 0.93 with operator-filtered data, compared with a correlation of 0.56 for unfiltered data. The energy of operator-designated contaminated trials was attenuated by a factor of 7.46 relative to uncontaminated trials. SNR improvement was observed in simulated tests.

Conclusions: Variations of SWT may be useful in situations where one wishes to separate uncommon/uncharacteristic structures from time series data sets. For artefact removal applications, SWT appears to be a valid alternative to expert operator screening.