Network-based biomarkers in pediatric drug-resistant epilepsy

Article information

Clin Exp Pediatr. 2026;69(8):636-638

Publication date (electronic) : 2026 July 24

doi : https://doi.org/10.3345/cep.2026.01774

Joseph Lim, MD, PhD¹, Jaeso Cho^,2,3, Hunmin Kim, MD, PhD^2,3

¹Department of Electrical and Computer Engineering, College of Engineering and Applied Science, University of Cincinnati, Cincinnati, OH, USA

²Department of Pediatrics, Seoul National University Bundang Hospital, Seongnam, Korea

³Department of Pediatrics, Seoul National University College of Medicine, Seoul, Korea

Corresponding author: Jaeso Cho, MD. Department of Pediatrics, Seoul National University Bundang Hospital, 82 Gumi-ro 173beon-gil, Bundang-gu, Seongnam 13620, Korea Email: jscho05@snubh.org

Received 2026 June 22; Revised 2026 July 13; Accepted 2026 July 14.

Key message

Approximately one-third of children with epilepsy develop drug-resistant epilepsy (DRE), defined by the International League Against Epilepsy as the failure of 2 appropriately chosen and used antiseizure medications (ASMs) to achieve sustained seizure freedom [1]. DRE may carry another concern: progressive disruption of the developing brain. Although approximately 85% of children undergoing epilepsy surgery show preoperative neuropsychological impairments in at least one cognitive domain [2], the neurobiological mechanisms linking them remain unclear.

Early life involves substantial white-matter growth and large-scale neural network integration. Their interruption may alter a child's neurodevelopmental trajectory [3]. However, conventional preoperative evaluations rely primarily on clinical surrogates such as seizure frequency, epilepsy duration, epilepsy onset age, and ASM burden. These variables reflect seizure exposure quantity but not how exposure may reshape brain network architecture.

The Jeong et al. [4] article published herein addresses this gap. Their diffusion-weighted imaging connectome (DWIC) framework uses seizure-associated structural abnormality biomarkers obtained from routine preoperative magnetic resonance imaging (MRI) scans to classify neurocognitive impairments in children with DRE.

Connectome-based epilepsy biomarkers have advanced considerably over the past decade. Larivière et al. indicated that structural and functional connectome measures can distinguish patients with DRE from healthy controls and track surgical outcomes [5]. Bernhardt et al. [6] showed that temporal lobe epilepsy is associated with network-level disorganization extending beyond the epileptogenic zone, including reduced global efficiency and altered hub topology. Stasenko et al. [3] described white-matter network disruption as an important contributor to neurobehavioral comorbidities in diverse epilepsy syndromes.

These prior studies were limited by mapping being largely agnostic of domain-specific neurocognitive function. Global graph metrics describe the overall brain network topology but do not indicate which cognitive systems are preferentially affected. Jeong et al.’s [4] approach differs: rather than interrogating the whole-brain network en masse, it uses 6 neurocognitive domain-specific networks (full-scale intelligence quotient [IQ], verbal IQ, nonverbal IQ, core language, expressive language, and receptive language) anchored to intracranial electroencephalography (iEEG)-confirmed seizure onset zones (SOZs). This individualized domain-resolved architecture offers a more granular alternative to population-level connectome summaries.

Jeong et al. [7] previously demonstrated that DRE in children with left hemispheric epilepsy is associated with a "crowding effect" quantified through a DWI connectome analysis. The current work extends that foundation by: (1) generalizing across lobar locations and lesion types; (2) formalizing the biomarker as a z-scored local efficiency deviation (ek = (1/M)ΣZm); and (3) deploying it within a supervised classification framework validated on an independent cohort. Table 1 compares connectome studies in pediatric and adult DRE.

Table 1.

Key connectome studies in pediatric versus adult DRE

Three methodological features merit attention in this present study. First, the iEEG-anchored seizure-affected network offers a useful solution to long-standing problems. Prior whole-brain connectome studies generally treated the epileptogenic zone as unknown or approximated it using lesion masks or scalp EEG. Jeong et al. [4] used iEEG recordings, considered the reference standard, to define SOZs and queried which nodes within each were structurally connected to them. This analysis reframes the mapping problem from describing the entire brain to characterizing the seizure source neighborhood, a more clinically interpretable target.

Second, the local efficiency deviation metric (ek) provides information that conventional metrics do not: patient-level and domain-specific quantification of the cumulative seizure burden on the brain's architecture. A child's seizure frequency indicates how often the network is stressed, while ek indicates how much structural change has accumulated in the networks serving a specific cognitive function. The large effect sizes (Cohen d ≈ 1.87–1.93) were maintained across subgroups defined by lesion type, SOZ location, seizure frequency, and seizure type, a degree of robustness that may distinguish this biomarker from prior well-performing metrics, primarily in homogeneous populations [6].

Third, partial least squares structural equation modeling (PL-SEM) moves beyond simple correlations to make inferences about pathways. Prior observational studies documented associations between epilepsy-related factors and cognitive outcomes [8] but could not readily disentangle the underlying pathways. Jeong et al. [4] demonstrated that including ek as a mediating imaging variable increases the coefficient of determination (r²) for neurocognitive impairment from 0.27 to 0.39 and raises the path effect size (f²) from 0.37 to 0.65. Although the design remains cross-sectional, this provides stronger structural evidence that seizure-associated white-matter disruptions contributes to neurocognitive impairment.

These strengths should be weighed against the limitations acknowledged by Jeong et al. [4]. The 6 neurocognitive networks are strongly intercorrelated, introducing information redundancy and an overfitting risk, and its cross-sectional design makes it difficult to separate seizure-related network propagation from the lesion's structural effect.

The practical significance of this study is as follows. First, diffusion tractography is now a routine component of preoperative MRI protocols at many epilepsy centers. The ek biomarker was derived entirely from clinicalgrade DWI scans, making its implementation feasible without requiring additional imaging. Second, the framework is individualized; because SOZs are derived from each child's own iEEG, the biomarker reflects individual seizure network topology rather than a group-averaged template. Third, a prediction tool that achieved 90%–98% accuracy in an independent cohort for general, verbal, and nonverbal cognitive domains could help inform surgical timing decisions.

Future research directions are as follows: Prospective longitudinal studies measuring ek before and after surgical resection would help test whether biomarker normalization tracks cognitive recovery analogous to postoperative language improvement studies using DWI tractography [9]. Combining the ek biomarker with complementary structural measures, such as regional brain volumetry stratified by genetic etiology [10], may help clarify which children are at greatest neurocognitive risk. The incorporation of advanced seizure severity scales could enrich the clinical correlates. Finally, extending the framework to resting-state functional MRI connectomes would allow functional validation of the structural abnormalities captured by ek and could indicate whether structural-level network inefficiency translates to disrupted functional dynamics.

By anchoring a DWI network analysis to iEEG-confirmed SOZs and resolving them across 6 neurocognitive domains, Jeong et al. [4] proposed a biomarker that may be more precise, individualized, and clinically interpretable than previous whole-brain metrics, results of which are supported by encouraging classification accuracy and PL-SEM rigor. Importantly, this study reframes DRE evaluations by counting seizures to map their network footprints. If validated in larger prospective cohorts, the DWIC framework could be a useful preoperative tool for identifying affected cognitive networks.