Systematic characterization of stochastic activity in non-invasively recorded neural signals

Scalp encephalograms (EEG) are often contaminated by various types of biological and non-biological noise that affect the performance of source localization, signal decoding and/or event estimation methods. The statistics and structure of EEG noise are usually unknown and time-varying. As these characteristics may vary substantially between subjects, as well as within subjects both in time and space, it may be difficult to select a unique noise model and/or to update its parameters. This study proposes a data-driven approach, based on the Empirical Model Decomposition and the autocorrelation function, for estimating and comparing the spatio-temporal statistical characteristics of EEG noise. The proposed approach was applied to a dataset of continuously recorded (over several hours) scalp EEGs from 3 patients with focal epilepsy. It is shown that the statistical characteristics of EEG noise vary substantially in time and space. Thus, adaptive signal processing methods may be most appropriate for denoising and/or estimation of data-driven and possibly subject-specific noise distributions, with the ultimate goal to improve source localization, waveform feature extraction and signal decoding.