Modelling Non-stationarities in EEG Data with Robust Principal Component Analysis

Abstract

Modelling non-stationarities is an ubiquitous problem in neuroscience. Robust models help understand the underlying cause of the change observed in neuroscientific signals to bring new insights of brain functioning. A common neuroscientific signal to study the behaviour of the brain is electro-encephalography (EEG) because it is little intrusive, relatively cheap and easy to acquire. However, this signal is known to be highly non-stationary. In this paper we propose a robust method to visualize non-stationarities present in neuroscientific data. This method is unaffected by noise sources that are uninteresting to the cause of change, and therefore helps to better understand the neurological sources responsible for the observed non-stationarity. This technique exploits a robust version of the principal component analysis and we apply it as illustration to EEG data acquired using a brain-computer interface, which allows users to control an application through their brain activity. Non-stationarities in EEG cause a drop of performance during the operation of the brain-computer interface. Here we demonstrate how the proposed method can help to understand and design methods to deal with non-stationarities.