Abstracts

Rationale: Proper identification of the eloquent cortex is crucial for improving surgical outcomes in patients with drug-resistant epilepsy. However, non-invasive techniques to identify the eloquent cortex are considerably lacking. Phase-amplitude coupling (PAC), a form of cross-frequency coupling, involves interactions between the phase of a low-frequency oscillation and the amplitude of a high-frequency oscillation, which is particularly relevant in both healthy and pathological brain functioning. To address this critical need, we developed a novel algorithm for characterizing PAC during task versus spontaneous magnetoencephalography (MEG) recordings of epilepsy patients.

Methods: We analyzed a retrospective database of epilepsy patients at UNMC and selected thirty subjects with somatosensory (SEF-UL, SEF-UR) and spontaneous MEG recordings. Raw MEG data were preprocessed to eliminate electrical noise as well as cardiac and ocular artifacts. Epochs of 250 ms (first 50 ms as baseline) were created and averaged across trials, followed by source reconstruction using an LCMV beamformer. PAC was quantified using the mean vector length modulation index, with 1 to 12 Hz as the low-frequency range and 80 to 300 Hz as the high-frequency range. PAC was calculated on the time-courses of neural activity for 148 brain regions defined by the Destrieux atlas. Tensor decomposition using the Candecomp/Parafac algorithm with a rank of 3 was applied for dimensionality reduction. Density-based spatial clustering (DBSCAN) clustering with silhouette evaluation was employed to identify task-activated brain regions as outliers. A linear mixed effects (LME) model analyzed PAC as a response variable with expected regions as fixed effects and patients as a random variable. A binary support-vector machine (SVM) classifier was developed for somatosensory task to predict brain regions depending on PAC values.

Results: Clustering identified significant active brain regions among expected regions for SEF-UL (p=2.97 x 10^-9) and SEF-UR (p=4.6 x 10^-7, binomial test) with random distribution for spontaneous data. Highly significant p-values were observed for SEF-UL (1.11 x 10^-113) and SEF-UR (2.02 x 10^-129) for the active brain regions across LME models. The patient specific SVM classifier accurately predicts (96.13%) active brain regions with an area under the curve value of 0.92.

Conclusions: Our PAC algorithm demonstrated a robust capability to identify the somatosensory cortex with significantly altered evoked response in comparison to spontaneous data. The algorithm will be extended to identify the language, auditory, motor, and memory cortices and can be used as a biomarker of activated cortex. The findings of this work broaden our understanding of neural oscillations, as well as provide non-invasive detection techniques for pre-surgical mapping in epilepsy patients.

Funding: N.A