Multiple modelling of EEG data to classify different mental states

Abstract

An electroencephalogram EEG gives the possibility of recording the electrical brain activity non-invasively from different locations on the scalp. The EEG data is known to be nonstationary due to continuously changing dynamics characterising the underlying mental states. This work investigates the applicability of the Autoregressive Switching Multiple Model (AR-SMM) framework for the automatic labelling of EEG data, taking into consideration both simulated data to mimic the nature of transient events and get more insight on the sensitivity of the framework to different mental state characteristics, as well as real EEG data presenting different challenges for the modelling of the data through AR-SMMs. Although Autoregressive (AR) models have been applied extensively for EEG data analysis, their combination with a switching framework that can handle better the abruptly changing dynamics of the nonstationary data has not been given much attention. In this work an existing model order identification criterion is modified and validated through Monte Carlo analysis on both univariate and multivariate data showing that it gives more accurate model order estimates. AR features estimated through an EM-based Kalman Smoother (EMKS) were then shown to give features that can reliably distinguish between left and right hand movements and where performance was insensitive to the model order estimate. The second part of this work focuses on the AR-SMM framework using lower bounding approximations for the identification of transitions between different candidate models. Applied to real EEG data the framework showed the capability of identifying multiple transient events in a unified framework which requires very little training data. A novel approach of learning new states in a semi-supervised manner also showed the possibility of using this framework as an analytical tool to obtain further insight on the dynamics of the EEG data.