Multi-branch convolutional neural network and intracranial EEG high-frequency oscillations predict post-surgical seizure outcomes

Objective: Pathological High-Frequency Oscillations (HFOs) identify epileptogenic cortex, but their surgical utility is unproven. Current epilepsy surgery planning relies on a “gold standard” multidisciplinary consensus. We tested if a Convolutional Neural Network (CNN), leveraging HFO features, neuroanatomy, and surgical boundaries, could predict seizure freedom. Methods: HFOs were detected during NREM sleep EEG in 78 pre-surgical patients. A three-branch CNN was trained using SEEG contact inputs: stereotaxic coordinates, resection status, and 37 HFO features, utilizing known post-operative seizure outcome. Branches encoded spatial, electrophysiological, and surgical data. Outputs were concatenated and processed by fully connected layers; a final sigmoid layer predicted post-operative seizure freedom probability. Univariate HFO feature analysis employed two-way mixed-effect ANOVAs. Results: The HFO-informed CNN model distinguished seizure-free patients with 92% accuracy using fivefold cross-validation. Univariate analysis suggested that fast ripples, especially those superimposed on epileptiform spikes, are important HFO features for the model. Conclusions: A trained CNN model integrating HFO features, neuroanatomy, and surgical boundaries can accurately predict seizure freedom following “gold standard” surgical planning. Significance: This CNN model, using inter-ictal non-REM sleep recordings, can predict surgical success and allow counterfactual virtual resections to be iteratively tested by the CNN ML to potentially improve post-operative seizure outcome.