Identifying epileptogenic abnormalities through spatial clustering of MEG interictal band power

Abstract

Successful epilepsy surgery depends on localizing and resecting cerebral abnormalities and networks that generate seizures. Abnormalities, however, may be widely distributed across multiple discontiguous areas. We propose spatially constrained clusters as candidate areas for further investigation and potential resection. We quantified the spatial overlap between the abnormality cluster and subsequent resection, hypothesizing a greater overlap in seizure-free patients. Thirty-four individuals with refractory focal epilepsy underwent pre-surgical resting-state interictal magnetoencephalography (MEG) recording. Fourteen individuals were totally seizure-free (ILAE 1) after surgery and 20 continued to have some seizures post-operatively (ILAE 2+). Band power abnormality maps were derived using controls as a baseline. Patient abnormalities were spatially clustered using the k-means algorithm. The tissue within the cluster containing the most abnormal region was compared with the resection volume using the dice score. The proposed abnormality cluster overlapped with the resection in 71% of ILAE 1 patients. Conversely, an overlap only occurred in 15% of ILAE 2+ patients. This effect discriminated outcome groups well (AUC = 0.82). Our novel approach identifies clusters of spatially similar tissue with high abnormality. This is clinically valuable, providing (a) a data-driven framework to validate current hypotheses of the epileptogenic zone localization or (b) to guide further investigation.

Keywords: MEG; clustering; epilepsy; outcome; prediction; surgery.

FIGURE 1

Processing pipeline to identify abnormality clusters. A, First, patient abnormality maps were constructed using interictal band power MEG recordings and healthy recordings as a baseline. Second, patient abnormality maps are filtered to retain only the top N strongest abnormalities. Third, the strongest abnormalities are grouped into K clusters using k‐means clustering with four features (x, y, z coordinates of each ROI and the abnormality value). Each feature is scaled to minimize any bias in the clustering. Finally, the cluster containing the most abnormal region is selected, with the added constraint that regions reside in a single hemisphere to better reflect focal resections. The chosen hemisphere was decided based on which contained over half of the regions within the cluster. B, The overlap between the abnormality cluster and actual resection was compared using the dice score (DSC) which measures the ratio of overlap (2TP) relative to the union of the abnormality cluster and resection (2TP + FP + FN). FN, false negative; FP, false positive; TN, true negative; TP, true positive. The DSC varies from 0 to 1 with DSC = 0 corresponding to no overlap at all, and DSC = 1 corresponding to the perfect overlap (identical regions). Steps (A) and (B) are repeated to find the optimal values of N and K to maximize the DSC across the cohort whilst balancing model complexity. C, The DSC across the whole cohort is compared with differences between seizure‐free (ILAE 1) and not‐seizure‐free (ILAE 2+) quantified using the area under the receiver operator curve (AUC).

FIGURE 2

Overlap between the abnormality cluster and actual resection. A, Illustration of the overlap between the abnormality cluster and actual resection in an example seizure‐free patient. Visually, a high overlap exists between the actual resection (red) and the abnormality cluster (blue). This is apparent with a DSC = 0.53 indicating a good overlap. In this scenario, we hypothesized a good surgical outcome as both resection masks cover similar tissue. Conversely, (B) corresponds to an example poor outcome patient with no overlap between the abnormality cluster and the resection. A DSC = 0 would suggest a poor outcome for this patient as the abnormality cluster and the resection are in complete disagreement. Cohort‐wide comparisons (C) demonstrate that in seizure‐free patients (green) as hypothesized there is a higher overlap between the abnormality cluster and the subsequent resection in contrast to patients with poor surgical outcomes (red), AUC = 0.82, P = 0.0001.