The interictal suppression hypothesis is the dominant differentiator of seizure onset zones in focal epilepsy

Abstract

Successful surgical treatment of drug-resistant epilepsy traditionally relies on the identification of seizure onset zones (SOZs). Connectome-based analyses of electrographic data from stereo electroencephalography (SEEG) may empower improved detection of SOZs. Specifically, connectome-based analyses based on the interictal suppression hypothesis posit that when the patient is not having a seizure, SOZs are inhibited by non-SOZs through high inward connectivity and low outward connectivity. However, it is not clear whether there are other motifs that can better identify potential SOZs. Thus, we sought to use unsupervised machine learning to identify network motifs that elucidate SOZs and investigate if there is another motif that outperforms the ISH. Resting-state SEEG data from 81 patients with drug-resistant epilepsy undergoing a pre-surgical evaluation at Vanderbilt University Medical Center were collected. Directed connectivity matrices were computed using the alpha band (8-13 Hz). Principal component analysis (PCA) was performed on each patient's connectivity matrix. Each patient's components were analysed qualitatively to identify common patterns across patients. A quantitative definition was then used to identify the component that most closely matched the observed pattern in each patient. A motif characteristic of the interictal suppression hypothesis (high-inward and low-outward connectivity) was present in all individuals and found to be the most robust motif for identification of SOZs in 64/81 (79%) patients. This principal component demonstrated significant differences in SOZs compared to non-SOZs. While other motifs for identifying SOZs were present in other patients, they differed for each patient, suggesting that seizure networks are patient specific, but the ISH is present in nearly all networks. We discovered that a potentially suppressive motif based on the interictal suppression hypothesis was present in all patients, and it was the most robust motif for SOZs in 79% of patients. Each patient had additional motifs that further characterized SOZs, but these motifs were not common across all patients. This work has the potential to augment clinical identification of SOZs to improve epilepsy treatment.

Keywords: connectomics; drug-resistant focal epilepsy; interictal suppression hypothesis; machine learning; principal component analysis; stereo electroencephalography.

Figure 1

Data collection and principal component analysis schematic. All 81 patients were implanted with stereoelectroencephalography (SEEG) contacts. Each patient’s 2-min segment of resting-state SEEG data were fed into a principal component analysis (PCA) algorithm, providing coefficients and weights for all principal components. A whole-connectome connectivity analysis was then performed to find local and global connectivity motifs, which were then analysed with weights and coefficients. Based on the ranking of each principal component’s variance, weights and coefficients were used to reconstruct the original partial directed coherence data. Finally, a focused connectivity analysis was performed for common themes/features on reconstructed PCA data. Across the cohort, there was an average of 84.52 principal components per patient We then performed PCA-seeded connectivity analyses on this reconstructed data, allowing for an interpretation of the precise influence of each principal component on a patient’s larger connectivity profile.

Figure 2

Principal component coefficient visualization and feature identification. (A and B) Visualization of the top five principal component (PC) coefficients for two sample patients (ranked by explained variance in the PC analysis). (C) Canonical interictal suppression hypothesis (ISH) components are outlined with a red box. Coefficient values are computed both inwards and outwards. That is, bipole pairs are repeated. PC 1, which is marked by a red rectangle, best exemplifies canonical suppressive behaviour for both patients, with high inwards and low outwards connectivity at nodes (bipole pairs), which are seizure onset zones (SOZs). This process may be repeated for nodes in specific regions or at specific SOZ designations to discover lobar or SOZ-driven patterns, respectively. While the high inward and low outward coefficients are present globally, this motif is only highly weighted for SOZs.

Figure 3

Weight and variance analysis of interictal suppression hypothesis (ISH) components. (A) For each ISH component, explained variance is plotted between the principal component (PC) 1 and PC 2+ cohorts (P = 2.14 × 10⁻⁹). (B) The difference in weights at seizure onset zone (SOZ) and non-involved zone (NIZ) nodes is plotted between both cohorts (P = 0.00139). The difference in weights is defined as the average SOZ weight of a component minus the average NIZ weight of the same component. These analyses suggest that the PC 1 cohort had a greater difference between SOZs and NIZs. In the violin plots, each dot represents a data point, the white circle represents the median value and the grey bar represents the 25th and 75th percentiles of these data. Significance testing was performed with a paired t-test: ***P < 0.001.

Figure 4

Weight of the interictal suppression hypothesis component for seizure onset zone, propagative zone and non-involved zone between the principal component (PC) 1 and PC 2+ cohorts. (A) Within the PC 1 sub-cohort, weights were computed for seizure onset zones (SOZs), propagative zones (PZs) and non-involved zones (NIZs); one-way ANOVA, P = 2.57 × 10⁻⁹ with post hoc multiple pairwise t-test comparisons significant for all groups. (B) This analysis was repeated for the PC 2+ sub-cohort (one-way ANOVA, P = 0.0476 with post hoc multiple pairwise t-test comparisons significant for SOZ-PZ). Post hoc paired t-test significance: *P < 0.05, **P < 0.01, ***P < 0.001.

Figure 5

Alpha band reciprocal connectivity. (A) directed reciprocal alpha-band partial directed coherence (PDC) connectivity was evaluated for seizure onset zones (SOZs), propagative zones (PZs) and NIZs for the full 810-patient cohort (one-way ANOVA, P = 6.04 × 10⁻¹²), the principal component (PC) 1 cohort (one-way ANOVA, P = 6.10 × 10⁻⁸) and the PC 2+ cohort (one-way ANOVA, P = 3.74 × 10⁻⁶). (B) Using the same methods, reciprocal alpha-band connectivity was computed for each patient’s interictal suppression hypothesis (ISH) component in the full 81-patient cohort (one-way ANOVA, P = 2.29 × 10⁻⁹), the PC 1 cohort (one-way ANOVA, P = 2.69 × 10⁻⁷) and the PC 2+ cohort (one-way ANOVA, P = 2.78 × 10⁻²). (C) Using the same methods, reciprocal alpha-band connectivity was computed for all components other than the ISH component (i.e. excluding the ISH component) for the full 81-patient cohort (one-way ANOVA, P = 5.42 × 10⁻³), the PC 1 cohort (one-way ANOVA, P = 7.37 × 10⁻⁵) and the PC 2+ cohort (one-way ANOVA, P = 4.12 × 10⁻⁴). Significance levels for post hoc t-tests are shown: *P < 0.05, **P < 0.01, ***P < 0.001.