Abstracts

Rationale: In epilepsy, EEG-fMRI investigates BOLD responses to discharges seen on EEG. However, scalp EEG only partially reflects underlying neuronal activity, particularly if the latter is deep or spatially restricted. We studied BOLD responses to ictal discharges, aiming to detect signal changes occurring earlier/later or with different duration than seen on scalp EEG. Methods: We considered EEG-fMRI data of 15 epileptic patients who had seizures during their scanning session (3T, TR=1.75s, 5mm voxels). After removal of gradient and ballistocardiogram artifacts, the ictal discharges were marked on the EEG. In a first analysis (LIN1), a standard linear approach identified voxels significantly correlated with the seizure blocks convolved with a canonical hemodynamic response function (HRF). A second analysis (LIN4) allowed for some variability by using 4 HRFs peaking at 3, 5, 7, and 9 seconds. A third approach (NLIN) used a canonical HRF shape, but also allowed the onset and duration of the seizure blocks to vary by up to 10 seconds from the values seen on the EEG, using non-linear regression to fit these parameters. The residuals were modeled as autoregressive noise, and statistical significance was computed by Monte-Carlo simulations. In all analyses, the voxel-wise significance level was set at p<0.001, with a minimum cluster size of 4 voxels to control for family-wise errors (p<0.05). Cluster volumes were compared across the 3 methods using Friedman’s test and post-hoc Tukey tests (p<0.05). Results: The volume of BOLD responses was significantly larger for the LIN4 (mean: 3009 voxels) and NLIN (mean: 2336 voxels) analyses than for LIN1 (mean: 1248 voxels), with no difference between LIN4 and NLIN. The NLIN analysis revealed that BOLD changes undetected by LIN1 had different onsets and durations than the EEG events. Volumes were also calculated only for the cluster nearest to the EEG seizure onset electrode, and thus presumably in the seizure onset zone. This cluster was significantly larger for LIN4 (mean: 1224 voxels) and NLIN (mean: 1376 voxels) than for LIN1 (mean: 540 voxels), with no difference between LIN4 and NLIN, indicating an increased sensitivity for these 2 methods (fig. 1). However, when excluding this cluster and considering only the remaining responses, which would be presumably distant from the seizure onset zone, the volume was significantly larger for LIN4 (mean: 1785 voxels) than for both LIN1 (mean: 708 voxels) and NLIN (mean: 960 voxels), with no significant difference between the latter 2 approaches. The LIN4 analysis may thus not be as specific as the NLIN method (fig. 2). Conclusions: The sensitivity of EEG-fMRI analyses with a single canonical HRF is affected by timing differences between ictal discharges and their scalp EEG correlates. Using 4 HRFs provides a more flexible model, but also increases detections distant from the presumed seizure onset zone. The non-linear method, which uses a canonical HRF but allows for variability in event onset and duration, may offer a better compromise between sensitivity and specificity. Funding: NSERC, CIHR MOP38079