Characterization of the synaptic mechanisms underlying seizure onset with Dynamic Causal Modelling

• ¹ University of Ghent, Data Analysis, Belgium
• ² UCL Institute of Neurology, Department of Clinical and Experimental Epilepsy, United Kingdom
• ³ Universidade de Lisboa, Institute of systems and robotics, Instituto superior técnico,, Portugal
• ⁴ Laboratory for Clinical and Experimental Neurophysiology, Department of Clinical and Experimental Epilepsy, Belgium
• ⁵ University College London, The Wellcome Trust Centre for Neuroimaging, United Kingdom

In this work we used Dynamical Causal Modelling (DCM) to characterize the synaptic mechanisms that underlie epileptic seizure onset. The basic idea is to use electrophysiological measurements to estimate the underlying effective connectivity and synaptic parameters of small neuronal networks generating the observed responses. One can then characterise the changes in spectral activity in terms of a trajectory in parameter space – identifying key parameters or connections that cause the observed changes. Using intracranial electroencephalographic (iEEG) recordings from three seizures in one patient with refractory epilepsy, a two source network is built – one source lying within and the other one just outside the putative seizure onset zone. The activity recorded at two sources was used to infer the architecture of synaptic connections among neural populations assumed to generate seizure activity. We summarised ongoing activity before and after seizure onset using cross spectral density measurements taken from these two sources of activity. These measurements covered 10 seconds of activity immediately before and after the seizure onset and were summarised in terms of nine cross spectra (from overlapping two second segments) before seizure onset and nine after onset. The 18 cross spectra were then used to invert a dynamic causal model of cross spectral activity to estimate the underlying intrinsic and extrinsic connectivity. Our analyses were based upon a two-step Bayesian model comparison procedure. In the first step we identified the best model architecture – distinguishing between extrinsic connections from the primary ictal source to the secondary source (with reciprocal backward connections) – as opposed to the reverse architecture with backward connections from the primary to secondary source. To disambiguate these two architectures, we inverted all 18 time windows and pooled the evidences for the two alternative models over windows for the three seizures. The second stage of analysis focused on the changes in intrinsic and extrinsic connectivity between windows – and implicitly between pre-and per-ictal states. Using the best model from the first step, we allowed various combinations of intrinsic and extrinsic connections to change with each window by treating each window as a separate condition. This allowed us to quantify the trajectory of coupling parameters within and between pre-and ictal time points– while holding all other parameters at the same values (e.g., transmission delays and electrode gain parameters). We considered 16 models corresponding to a factorial model space with and without changes in: intrinsic connectivity in the primary source, intrinsic connectivity in the secondary source, forward connectivity and backward connectivity and we have pooled again the evidence of these models across the 3 seizures. The first step of the Bayesian model comparison of the two competing models with different extrinsic (forward and backward) connections suggested that we can be almost certain that the forward connections originates in the primary source. The winning model for the second step of the Bayesian model comparison was the one allowing changes in intrinsic connectivity in both the primary and the secondary source. These results suggest that the seizure onset appears to be mediated by an inhibition of superficial pyramidal cells in both sources and that the synaptic changes necessary to explain observed seizure show slow dissociable time courses over several seconds. In addition, knowing the ground truth, we have performed an analysis with simulated data with parameters based upon estimates from the empirical data. The results suggested that the trajectory of parameters can be recovered even under fairly realistic levels of sampling noise and biologically plausible values for the neuronal dynamics. Our work illustrates the use of dynamic causal modelling to provide biophysically informed characterisations of observed electrophysiological responses – and to illustrate the use of model comparison in testing different hypotheses about the pathophysiology of epilepsy.