Adaptive nonlinear multivariate brain connectivity analysis of motor imagery movements using graph theory

Abstract

Recent studies on motor imagery (MI)-based brain computer interaction (BCI) reported that the interaction of spatially separated brain areas in forms of functional or effective connectivity leads to a better insight of brain neural patterns during MI movements and can provide useful features for BCIs. However, existing studies suffer from unrealistic assumptions or technical weaknesses for processing brain signals, such as stationarity, linearity and bivariate analysis framework. Besides, volume conduction effect as a critical challenge in this area and the role of subcortical regions in connectivity analysis have not been considered and studied well. In this thesis, the neurophysiological connectivity patterns of healthy human brain during different MI movements are deeply investigated. At first, an adaptive nonlinear multivariate statespace model known as dual extended Kalman filter is proposed for connectivity pattern estimation. Several frequency domain functional and effective connectivity estimators are developed for nonlinear non-stationary signals. Evaluation results show superior parameter tracking performance and hence more accurate connectivity analysis by the proposed model. Secondly, source-space time-varying nonlinear multivariate brain connectivity during feet, left hand, right hand and tongue MI movements is investigated in a broad frequency range by using the developed connectivity estimators. Results reveal the similarities and the differences between MI tasks in terms of involved regions, density of interactions, distribution of interactions, functional connections and information flows. Finally, organizational principles of brain networks of MI movements measured by all considered connectivity estimators are extensively explored by graph theoretical approach where the local and global graph structures are quantified by computing different graph indexes. Results report statistical significant differences between and within the MI tasks by using the graph indexes extracted from the networks formed particularly by normalized partial directed coherence. This delivers promising distinctive features of the MI tasks for non-invasive BCI applications.