Integrating multivariate resting-state fMRI features to localize epileptic networks in common childhood epilepsy

Accurate localization of epileptic activity remains challenging when interictal discharges are absent or sparse. Although resting-state functional MRI (rs-fMRI) is noninvasive, individual rs-fMRI metrics provide inconsistent and incomplete localization. We aimed to determine whether multivariate integration of rs-fMRI features could robustly identify syndrome-specific epileptic activity in childhood epilepsy.

Using simultaneous EEG-fMRI, we studied children with self-limited epilepsy with centrotemporal spikes (SeLECTS, n = 60) and childhood absence epilepsy (CAE, n = 30) alongside typically developing controls (n = 108). Forty-two rs-fMRI metrics spanning amplitude, connectivity, temporal dynamics, and directional measures were computed for each session. Individual abnormality maps were also generated relative to controls. Partial least squares (PLS) regression was then applied to identify a latent component (PLS1) that maximally covaried with syndrome-specific epileptic activation patterns derived from EEG-fMRI. Spatial correspondence, localization accuracy, non-discharge session sensitivity, and classification performance were evaluated.

PLS1 showed strong spatial correspondence with EEG-fMRI-defined epileptic activation patterns in both SeLECTS (rolandic cortex) and CAE (thalamocortical network) (both spin-test r = .68, p < .001). PLS1 showed superior localization performance compared with most single-metric rs-fMRI measures and reached EEG-fMRI-level localization accuracy. Notably, PLS1 detected graded, syndrome-specific abnormalities during non-discharge sessions and distinguished discharge, no-discharge, and control states, an effect that has not been observed with individual rs-fMRI metrics. In classification analyses, PLS1 differentiated CAE from SeLECTS with high accuracy (AUC = .79), performing comparably to the EEG-fMRI-based classification.

Multivariate integration of rs-fMRI features using a template-guided PLS framework enables sensitive and syndrome-specific detection of epileptic activity, even in the absence of overt discharges. This approach provides a clinically translatable strategy for noninvasive epilepsy network mapping.