Abstracts

Rationale: Several studies have investigated simultaneous EEG-fMRI in epileptic patients. Analysis using the general linear model (GLM) can often localize BOLD responses in areas that correlate well with other clinical data, but discordant activations may also occur. The present study investigated the use of independent component analysis (ICA) on EEG-fMRI data. ICA is an exploratory method that can separate BOLD changes due to independent spatial processes, without constraining the shape of the hemodynamic response function (HRF). Among the extracted processes, the identification of components corresponding to epileptic events may reveal previously undetected activation regions.Methods: A neurologist independently selected EEG-fMRI data from 10 epileptic patients (72-90 minutes, 1.5T, TR: 3s, 5mm voxel size). Using GLM methods, 5 patients showed BOLD changes to epileptic events that were concordant with clinical data (near an MRI lesion or in the lobe presumably generating EEG discharges), while the other 5 patients had discordant BOLD responses. ICA was used to extract independent spatial patterns with common time courses. To facilitate the identification of relevant components, a deconvolution method (Lu et al. Neuroimage 2006; 32(1):238-247) was applied to detect component time courses showing a significant response following epileptic events, without constraining the shape of the HRF. Spatial maps of these components were then compared to the results from the GLM methods.Results: In all 5 patients with a concordant GLM activation and 2 of the 5 patients with discordant GLM results, ICA could extract a component whose spatial map correlated well with other clinical data and whose time course was significantly related to the spike timings. The HRF fitted by the deconvolution method had a similar shape to the canonical HRF used in the GLM analysis. When additional components had a time course significantly related to the spike timings, their spatial map suggested that they represented large artifacts such as residual motion. Re-examination of the 2 patients with discordant GLM results showed that the regions of activation revealed by ICA were also present in the GLM statistical maps, but were below the threshold of significance of p=0.05, corrected using random field theory and taking into account spatial extent (see figure). The remaining 3 patients only had a maximum of 11 spikes during their EEG-fMRI acquisition. None of the components extracted by ICA had a time course significantly related to the spike timings.Conclusions: ICA can extract BOLD responses due to epileptic discharges from fMRI data without requiring a model of the HRF. By separating these responses from other independent sources of structured noise, ICA can also reveal plausible regions of activation that may have been missed by GLM methods. (NSERC CGSD, CIHR MOP-38079)