Presurgical EEG Network Analysis Predicts Seizure Outcomes after Corpus Callosotomy in Children with Drug-Resistant Epilepsy

Introduction: Corpus callosotomy (CC) is a palliative surgical procedure for drug-resistant epilepsy (DRE). Seizure reduction after CC is difficult to predict: around 45% of patients are exposed to the adverse events of this major procedure without deriving significant benefit. We aim to predict post-CC seizure outcome in children with DRE by leveraging presurgical scalp EEG network analysis and machine-learning, to ultimately enable enhanced patient selection.

Methods: We analyzed brief interictal scalp EEG data from 32 children with DRE who had CC (median age: 10.5 years). We estimated brain functional connectivity (FC) in five frequencies (delta, theta, beta, alpha, gamma) and computed: whole brain FC (WB-FC), within-hemisphere FC (WH-FC), between-hemisphere FC (BH-FC), and interhemispheric asymmetry (IHA). Each patient was classified based on their seizure-reduction outcome (Excellent:>90%, Intermediate:50-90%, Poor: < 50%). We tested the effect of each EEG network characteristic on patients’ outcome (ordinal logistic regression). The effect of clinical characteristics on the outcome was also similarly tested.
Finally, we tested the performance of different combinations of clinical and EEG features to predict excellent outcome via binary classification models (support-vector-machine, SVM; leave-one-out cross-validation).

Results: Presurgical gamma, alpha, and beta-network characteristics predicted post-CC seizure reduction: the greater the WB-FC, WH-FC, and BH-FC, the better the outcome (alpha: OR=2.0-2.3, beta: OR=1.8-1.9, gamma: OR=1.9-2.0, p< 0.05). High gamma IHA was also associated with better outcome (OR=2; p< 0.05).
The SVM classification model incorporating gamma EEG characteristics (plus age-at-surgery) achieved 91% accuracy in predicting outcome (area-under-ROC-curve=0.83). When we only used clinical characteristics (age-at-surgery, age at epilepsy-onset, spike-frequency, no. seizure types, epilepsy duration) classification accuracy was much lower (75%; area-under-ROC-curve=0.57).

Conclusion : Presurgical EEG network analysis may help predict post-CC seizure outcomes in children with DRE, augmenting the predictive value of conventional clinical data. Our data suggest that children with overconnected brain networks prior to surgery are most likely to experience excellent post-CC seizure reduction.