Machine-Learning-Based Localization of Cortical Hyperexcitability Zones from Background EEG Activity in Epilepsy

Abstract

Background rhythmic activity in routine EEG recordings of epilepsy patients contains extensive information about brain function under pathological conditions, far exceeding the duration of epileptiform and interictal discharges. However, clinical interpretation remains predominantly focused on detecting conspicuous pathological patterns, such as seizures and interictal events, which is labor-intensive and requires expert evaluation. Recent advances in rhythmic EEG analysis combined with machine learning (ML) have enabled reliable differentiation between healthy individuals and epilepsy patients. Building on this momentum, the present study introduces a novel ML framework for the automated analysis and localization of cortical hyperexcitability foci, using only background EEG oscillations in the absence of detectable interictal discharges or seizure events. Leveraging publicly available EEG data, we demonstrate that Random Forest and CatBoost algorithms can effectively predict the approximate localization of interictal discharge foci at the level of major cortical regions. In a cohort of 48 patients (782 one-minute background epochs, five localization classes), Random Forest achieved an accuracy of 0.92 with a macro F1-score of 0.90 under patient-wise cross-validation. These findings establish background EEG activity as a promising clinically relevant biomarker for focal epilepsy diagnosis and highlight the feasibility of developing automated, expert-independent localization tools, addressing a critical unmet need in clinical neurophysiology.