On-line EEG Denoising and Cleaning Using Correlated Sparse Signal Recovery and Active Learning

Abstract

We have developed two new methods that use sparse recovery and active learning techniques for near real-time artifact identification and removal in EEG recordings. The first algorithm, called Correlated Sparse Signal Recovery (CSSR) addresses the problem of structured sparse signal recovery when statistical rather than exact properties describing the structure of the signal are appropriate, as in the elimination of eye movement artifacts; such tasks cannot be done efficiently using structured models that assume a common sparsity profile of fixed groups of components. Our algorithm learns structured sparse coefficients in a Bayesian paradigm. Using it, we have successfully identified and subtracted eye movement (structured) artifacts in real EEG recordings resulting in minimal data loss. Our method outperforms ICA and standard sparse recovery algorithms by preserving both spectral and complexity properties of the denoised EEG. Our second method uses a new active selection algorithm that we call Output-based Active Selection (OAS). When applied to the task of detection of EEG epochs containing other non-structured artifacts from an ensemble of detectors, OAS boosts accuracy of the ensemble from 91% to 97.5% with only 10% active labels. Our methods can also be applied to real-time artifact removal in magnetoencephalography (MEG) and blood pressure signals.