Abstracts

Rationale: High frequency oscillations (HFOs, ripples) have been shown to be an informative biomarker for epileptogenic zone (EZ) localization. Researchers have employed various algorithms to detect ripples in interictal and ictal states. These approaches are usually subjective and assume fixed features for all epileptic cases, while patients are biologically diverse and have dissimilar seizure types, etiology, and electrode implantations. An objective perspective can be realized by analyzing epileptogenic networks using graph measures, e.g. eigenvector centrality (EVC), to identify the EZ. While recent work has examined EVC in frequency bands below 90Hz (PNAS 2014; 111(49); E5321-30), here we evaluate for HFOs to 200HZ. The epileptogenic networks which evolve over time are modeled using a novel multilayer network with intra-slice connections between neighboring graphs. We hypothesize that during seizures, nodes (electrodes) inside the EZ should have a distinctive pattern of EVC. In other words, these electrodes demonstrate abnormal patterns in comparison to other nodes. Using low-rank approximation of EVC, we can find the principal components (PCs) of these patterns and detect segregated (potential EZ) nodes using unsupervised clustering. Methods: Ictal SEEG data from 9 seizure-free patients was used in this study (Table I). We pre-processed SEEG signals (N channels) and then computed the time-varying epileptogenic networks in two frequency bands, 80-140 Hz and 140-200 Hz, and T time points using lagged-coherence. Next, the connectivity matrices were normalized with respect to pre-ictal period and mapped to the interval [0 1] using an exponential transform.We constructed an NT×NT super-adjacency matrix of lagged coherence between contacts with coupling effects between layers. The diagonal blocks were the adjacency matrices in different time samples and off-diagonal terms were identity matrices, representing connectivity between nodes in adjacent layers (neighboring time points). The T largest eigenvectors of this matrix were summed and reshaped to an N×T matrix, representing the nodes’ centrality in time, termed the multilayer EVC (mlEVC). We quantized the mlEVC matrices, concatenated the ictal periods of all seizures, and applied SVD to find PCs, i.e. N×1 left-singular vectors. Finally, combinations of the first three PCs were used for unsupervised clustering and EZ detection. Results: Fig. 1. shows scatter plots of PCs. In most cases, we find a cluster of nodes with distinctive features which are identified as the possible EZ. In Table I, sensitivity and specificity of our method are calculated with averages of 34% and 96%, respectively. Except for subject 8, results represent substantial consistency between identified areas and resected regions, supporting our initial hypothesis. Conclusions: By using coupled multilayer networks, we can directly take seizure evolution and dynamics into account. Our findings indicate that mlEVC computed from these networks has potential for use in EZ detection. Results presented are consistent across seizure types, e.g. lesional and non-lesional cases and EZ in differing brain regions. Funding: R01 NS089212