Authors :
Presenting Author: Hang Cao, MD – Xuanwu Hosp. / Penn
Nishant Sinha, PhD – University of Pennsylvania
Marc Jaskir, PhD – University of Pennsylvania
Daniel Zhou, MD – University of Pennsylvania
Sandhitsu Das, PhD – University of Pennsylvania
Joel Stein, MD, PhD – University of Pennsylvania
Guoguang Zhao, MD, PhD – Xuanwu Hospital, Capital Medical University
Kathryn Davis, MD – Center for Neuroengineering and Therapeutics and Penn Epilepsy Center, Department of Neurology, University of Pennsylvania
Rationale:
The Hurst exponent (H) captures scale-free temporal correlations in neural signals and deviates in both intracranial EEG (iEEG) and resting-state fMRI in drug-resistant epilepsy. Whether these deviations reflect convergent or modality-specific network properties related to seizure-onset zone localization and surgical outcome is uncertain. Clarifying this relationship may advance network-targeted surgery.
Methods:
We examined 65 adults (41 with paired iEEG + fMRI, 24 with iEEG only; 4,469 grey-matter electrodes). Reference distributions came from 106 atlas iEEG and 55 healthy fMRI controls. iEEG channels were labelled spared or resected using post-operative margins. After artefact rejection, common-average referencing, notch/band-pass filtering, and down-sampling to 200 Hz, H was estimated across six frequency bands with multifractal wavelet leaders and converted to ROI-level z-scores. 3-Tesla resting-state fMRI was processed using fMRIPrep and xcpEngine pipelines. Using Haar wavelet maximum-likelihood estimation, H was extracted at electrode-proximal voxels sampled by three schemes (nearest voxel, 3 mm radius, 5 mm radius) on smoothed and unsmoothed images, followed by voxel- and region-level z-normalization. Spatial autocorrelation was represented by embedding cortical-cortical geodesic and cortical-subcortical Euclidean distances into Matérn meshes within spatial GLMMs; residual dependence was assessed with Moran’s I. Fixed effects encoded modality, preprocessing, smoothing, averaging radius, normalization, and frequency band, whereas subject, electrode, and ROI were nested random effects. Likelihood-ratio tests, ΔAIC, and ΔBIC guided model selection, and multiple comparisons were controlled with the Benjamini–Hochberg procedure.
Results:
Cross-modal H correlations were weak and regionally heterogeneous. For voxel-wise fMRI H, an fMRI-only model with subject + electrode + ROI random effects yielded the best fit (ΔAIC = –72.5 vs. full; marginal R² = 0.304; conditional R² = 0.653). Adding iEEG predictors conferred negligible gain (ΔAIC = 0; R²ₘ = 0.291), whereas an iEEG-only model explained virtually no variance (R²ₘ≈0) and was decisively inferior (ΔAIC > 200, q< 0.001). In spatial outcome models, lower fMRI H predicted unfavorable prognosis (