Addressing spatial bias in intracranial EEG functional connectivity analyses for epilepsy surgical planning

Objective. To determine the effect of epilepsy on intracranial electroencephalography (EEG) functional connectivity, and the ability of functional connectivity to localize the seizure onset zone (SOZ), controlling for spatial biases. Approach. We analyzed intracranial EEG data from patients with drug-resistant epilepsy admitted for pre-surgical planning. We calculated intracranial EEG functional networks and determined whether changes in functional connectivity lateralized the SOZ using a spatial subsampling method to control for spatial bias. We developed a ‘spatial null model’ to localize the SOZ electrode using only spatial sampling information, ignoring EEG data. We compared the performance of this spatial null model against models incorporating EEG functional connectivity and interictal spike rates. Main results. About 110 patients were included in the study, although the number of patients differed across analyses. Controlling for spatial sampling, the average connectivity was lower in the SOZ region relative to the same anatomic region in the contralateral hemisphere. A model using intra-hemispheric connectivity accurately lateralized the SOZ (average accuracy 75.5%). A spatial null model incorporating spatial sampling information alone achieved moderate accuracy in classifying SOZ electrodes (mean AUC = 0.70, 95% CI 0.63–0.77). A model incorporating intracranial EEG functional connectivity and spike rate data further outperformed this spatial null model (AUC 0.78, p = 0.002 compared to spatial null model). However, a model incorporating functional connectivity without spike rate data did not significantly outperform the null model (AUC 0.72, p = 0.38). Significance. Intracranial EEG functional connectivity is reduced in the SOZ region, and interictal data predict SOZ electrode localization and laterality, however a predictive model incorporating functional connectivity without interictal spike rates did not significantly outperform a spatial null model. We propose constructing a spatial null model to provide an estimate of the pre-implant hypothesis of the SOZ, and to serve as a benchmark for further machine learning algorithms in order to avoid overestimating model performance because of electrode sampling alone.