EEG-based effective and functional connectivity for differentiating and predicting altered states of consciousness

How does the brain sustain consciousness? In this thesis, and in the work leading up to it, we provide new computational evidence for the importance of the posterior hot zone on one hand, and for long-distance frontoparietal connectivity on the other, in explaining the contrast between loss of consciousness and in maintaining conscious responsiveness.

We adopt a factorial approach in our study, crossing two altered states of consciousness with two analytical methods for measuring changes in brain associated with these altered states. Specifically, we study healthy controls under propofol-anaesthesia and patients suffering from disorders of consciousness (DoC), employing functional and effective electroencephalographic (EEG) connectivity, thereby forming a 2-by-2 study design.

We first demonstrate the power of functional EEG connectivity for predicting anaesthetic states in the healthy brain, by building a single multivariate regression model combining phase-lag brain connectivity and behaviour- and power-based dependent measures. We show that baseline alpha- and beta-connectivity, as measured prior to an anaesthetic induction, can predict both behaviour- and power-based measures during the induction and peak unresponsiveness, specifically as measured from the posterior electrodes. Next, we study patients suffering from DoC and show that the alpha-band functional connectivity over the left hemisphere, and graph-theoretic network centrality on the right, significantly predict the patient's clinical diagnosis. Our findings suggest a dissociation between mean spectral connectivity and network properties.

Building on these findings, we then turn to dynamic causal modelling (DCM) to estimate modulations in effective brain connectivity due to anaesthesia, in and between the default mode network (DMN), the salience network (SAL), and the central executive network (CEN). Advancing current understanding of anaesthetic-induced LOC, we show evidence for a selective breakdown in the posterior hot zone and in medial feedforward frontoparietal connectivity within the DMN, and of parietal inter-network connectivity linking DMN and CEN. In a novel DCM-based out-of-sample cross-validation, we establish the predictive validity of our models, specifically highlighting frontoparietal connectivity as a generalisable predictor of states of consciousness. Importantly, we demonstrate a generalisation of this predictive power in an unseen dataset from the post-anaesthetic recovery state.

Finally, we again use DCM to investigate changes in the effective connectivity between DoC patients and healthy controls within the DMN. Specifically, we show that the key difference between healthy controls or conscious patients and completely unresponsive patients is a reduction in left-hemispheric backward frontoparietal connectivity. Finally, with out-of-sample cross-validation, we show that left-hemispheric frontoparietal connectivity can not only distinguish patient groups from each other, it can also generalise to an unseen data subset collected from seemingly unresponsive patients who show evidence of consciousness when assessed with functional neuroimaging. This suggests that effective EEG connectivity can be used to identify covertly aware patients who seem behaviourally unresponsive.

Overall, this thesis provides novel insights into the brain dynamics underlying transitions between altered states of consciousness and highlights the value of tracking these dynamics in a clinical context. DCM, though computationally more expensive, can accurately predict states of consciousness and provide causal explanations of the brain dynamics that cannot be inferred from functional connectivity alone. Functional connectivity, though correlational, is still an accurate predictive tool of altered states of consciousness. With clinically challenging, ambiguous cases like potentially covertly aware patients, we propose that the causal explanations and accurate predictions of DCM modelling could outweigh the computational complexity.