Abstracts

Rationale: Communication between regions of the brain is a dynamic process that allows for different connections to accomplish different tasks. While the content of the interregional communications is complex, we hypothesize that the pattern of connectivity may be described in a lower dimensional state-space that contains information about the brain function. In epilepsy, seizures elicit changes in connectivity, whose pattern sheds insight into the nature of seizures and the epileptogenic zone (EZ). We investigated the temporal evolution of the brain connectivity before, during, and after seizure by applying network-based analysis on continuous multi-day subdural ECoG recordings from 12 drug-resistant epilepsy patients. Methods: We used a network-based approach to analyze the ECoG recordings from subdural and depth electrodes. For each patient, we investigated the role of the clinically annotated EZ by processing the ECoG data as follows. First, let us consider recordings from N>2 channels. By using a 3 s-long window, sliding every second, we computed a N x N matrix, A, whose generic (i,j)-th element is the coherence in the 14-25 Hz band (beta) between electrodes i, j, i.e.: A_ij = |P_ij|²/P_iiP_jj, where P_ij is the cross-power spectrum averaged over the frequency band, and P_ii and P_jj are the power spectrum for electrodes i and j averaged over the beta band, respectively. We computed the eigenvector centrality (EVC) to measure the connectivity of each electrode. For each electrode i, EVC(i) is defined as: EVC(i) = λ ∑_j=1^N A_ij EVC(j), where λ is the leading eigenvalue of the matrix A and EVC is the associated eigenvector of A. The leading eigenvectors of the connectivity matrices were calculated numerically at each second and then clustered via K-mean to uncover a finite set of connectivity states (i.e., network topologies) that the brain transitions between. Results: Across all patients, we found that (i) the network connectivity defines a finite set of brain states (2-3 during non-seizure periods, 2-7 during seizure), (ii) seizures are characterized by a significantly consistent progression of states (p-value < 0.005), (iii) the EZ is initially isolated from other regions near the seizure onset (signature of the EZ) and then becomes most connected in the network towards the end of seizure. A ROC analysis revealed that this particular state and the corresponding network structure can detect the EZ with high (i.e., above 0.9) specificity and sensitivity. Conclusions: To localize the EZ, clinicians inspect recordings from several intracranial electrodes, a time consuming process that may not result in definitive localization. We show that network-based statistics computed from all ECoG channels capture the brain dynamics as the seizures approach and reveal consistent temporal connectivity patterns near the EZ. This suggests that a state-space model can characterize the brain dynamics and help localize the EZ.