On-line EEG Denoising using correlated sparse recovery
Manish Gupta, Scott A. Beckett, Elizabeth B. Klerman · 2016
We address the problem of structured sparse signal recovery when only certain statistical rather than exact properties describing the structure of the signal are available. Such problems arise in elimination of eye movement artifacts in waking EEG recordings; this task cannot be efficiently done using structured models that assume a common sparsity profile of fixed groups of components. We present an algorithm for learning structured sparse coefficients in a Bayesian paradigm. Using our algorithm, that we call Correlated Sparse Recovery (CSR), we have successfully removed eye movement artifact in real EEG recordings, with no data loss. Our method, which facilitates on-line application, outperforms ICA and standard sparse recovery algorithms in its ability to denoise EEG with minimal spectral distortion and maximal entropy preservation. As an artifact detection tool our method has 96% specificity and 97% sensitivity when compared with manual identification by an expert.