Predicting seizure onset zones from interictal intracranial EEG using functional connectivity and machine learning.

Functional connectivity (FC) analyses of intracranial EEG (iEEG) signals can potentially improve the mapping of epileptic networks in drug-resistant focal epilepsy. However, it remains unclear whether FC-based metrics provide additional value beyond established epilepsy biomarkers such as epileptic spikes and high-frequency oscillations (HFOs). Using interictal iEEG data from 26 patients, we estimated FC across eight frequency bands (4-290 Hz) using amplitude envelope correlation (AEC) and phase locking value (PLV). From the resulting FC-matrices, we estimated two graph metrics each to derive 32 FC-based features. We also extracted features related to spikes, HFOs, and power spectral densities (PSD). A trained support vector machine (SVM) classifier predicted seizure onset zones (SOZs) with an area under the ROC curve (AUC) of 0.91 for node-level 4-fold cross-validation (CV), 0.69 for patient-level 4-fold CV, and 0.73 for patient-level leave-one-out CV. Notably, gamma-band graph features from AECs outperformed spikes and HFOs in SOZ prediction when using an equivalent number of features. Our results strongly suggest that AEC-based features may provide more information about epileptogenicity compared to PLV-based features. Furthermore, machine learning provides a robust approach for identifying useful FC-based features and integrating information from putative biomarkers of epilepsy to better localize epileptogenic networks.