Transformer-based EEG Source Imaging Enables Robust Localization of Pathological High-Frequency Oscillations in Epilepsy

ObjectiveHigh-frequency oscillations (HFOs) are highly specific biomarkers of epileptogenic tissue, yet their noninvasive localization remains challenging due to their brief duration, low amplitude, and poor signal-to-noise ratio. Here, we introduce TH-DeepSIF, a transformer-based deep learning framework trained on biologically realistic neural mass model simulations, to robustly perform HFO source imaging from scalp EEG. MethodsTH-DeepSIF was evaluated in simulated single- and dual-source HFO scenarios, where it was tasked with recovering both the spatial location and temporal dynamics of HFO generators under increasing spatiotemporal complexity. We further validated TH-DeepSIF in 25 patients with drug-resistant epilepsy by comparing EEG source imaged HFO sources against surgical resection regions. ResultsTH-DeepSIF accurately recovered both the spatial location and temporal dynamics of simulated HFO generators, achieving low localization error and strong waveform correspondence with ground truth. TH-DeepSIF localization for pathological HFOs (pHFOs, or spike ripples) demonstrated strong agreement with surgical resection regions, achieving a median localization error of 11.9 mm and specificity of 0.896. Compared with all HFOs (aHFOs), pHFO-based source imaging showed significantly stronger spatial correspondence with resection regions, significantly smaller localization error, and higher precision, sensitivity, geometric mean, and F1 score. SignificanceThese findings demonstrate that TH-DeepSIF provides a robust, data-driven framework for noninvasive HFO source imaging with improved anatomical specificity and enhanced clinical utility for presurgical evaluation using scalp EEG. Moreover, they show that pathological HFOs (spike ripples )--rather than general HFOs--serve as robust EEG biomarkers for accurate localization of the epileptogenic zone. Key PointsO_LIA fully data-driven, parameter-free framework for noninvasive HFO source imaging using scalp EEG. C_LIO_LIPathological HFO source maps exhibit strong spatial concordance with both the surgical resection region and the seizure onset zone. C_LIO_LISource imaging based on pathological HFOs consistently outperforms imaging based on all detected HFOs. C_LI