Abstract

Background: Epilepsy is a heterogeneous syndrome. Personalised localisation of epileptogenic zone (EZ) is critical for diagnosis and treatment of drug-resistant focal epilepsy. Multichannel stereoelectroencephalography (SEEG) monitoring acquired over a period of two to three weeks was collected in different patients, resulting in comprehensive epileptogenic information and terabytes of high dimensional data. Consequently, there is a need for high-throughput data analytical methods to enable data-driven, personalised seizure detection and EZ localisation.

Methods: Here, a seizure detection and EZ localisation AI system — SEEGformer is proposed, by utilising SEEG data from 61 patients acquired across two centres and three cohorts capturing tens of thousands of abnormal discharges and around ten seizures per person on average. SEEGformer employs a parallel transformer architecture to analyse multiple representations of multichannel SEEG signals, including the real part, imaginary part, and amplitude of the analytic signal after Fourier transform. The MRI information was encoded in SEEGformer to construct the structural dependence of the brain areas. Inter-channel dependencies and interactions were captured for seizure detection. A cross-channel attention mechanism computed the epileptogenic risk score for each channel to localise EZ using ictal SEEG data. Each patient’s SEEG data was used to train and validate their individual-specific SEEGformer model.

Findings: In three clinical cohorts, SEEGformer achieved an average AUROC of 0.937 (95% CI, 0.922–0.950) for seizure detection and 0.798 (95% CI, 0.749–0.847) for EZ localisation. Localisation performance surpassed state-of the-art methods by over 5%. SEEGformer further revealed distinct phase synchronisation patterns in dynamically evolving epileptogenic zone networks, with a significance level of P < 0.0001.

Interpretation: Due to its high interpretability and visualisation capabilities, SEEGformer can enhance clinical decision-making by providing an objective, data-driven reference to optimise epileptogenic zone delineation and surgical strategy development. Currently, the improved SEEGformer is being developed to construct a dedicated SEEG atlas for epilepsy.

Fig. 1: Three intuitive cases of EZ localisation outcome based on SEEGformer. a Case 1: A 32-year old male with EZ that includes right hippocampus-amygdala. b Case 2: A 27-year old male with EZ that includes left hippocampus-amygdala. c Case 3: A 38-year old male with EZ that includes left temporal pole–lateral surface of temporal lobe.

Fig. 2: Public dataset evaluation and phase synchronisation analysis. a Seizure detection performance in external cohort 1 (16 patients). b ROC curves of different models in external cohort 1. c PR curves of different models in external cohort 1. d Seizure detection performance in external cohort 2 (18 patients). e ROC curves of different models in external cohort 2. f PR curves of different models in external cohort 2. g Scatter plotting of EZ localisation performance and seizure detection performance in internal cohort (27 patients × 4 strategies × 5 folds = 540 data points). h The PLV of each patient across different regions during pre-seizure stage in the 27-patient internal cohort (EZ: PLV within epileptogenic zone, No EZ: PLV of non-epileptogenic zone, Inter: PLV between epileptogenic zone and non-epileptogenic zone, Whole: PLV of all electrode sites). i Comparison of the PLV within epileptogenic zone in seizure stage with different clip-selection thresholds (0%, 50%, 75% and 90%) according to the attention output of our model (27 patients), *P < 0.01, **P < 0.001, ***P < 0.0001. P-values were obtained by paired-sample t-tests. j Comparison of the PLV of non-epileptogenic zone in seizure stage with different clip-selection thresholds (27 patients). k Comparison of the PLV between epileptogenic zone and non-epileptogenic zone in seizure stage with different clip-selection thresholds (27 patients). l Comparison of the PLV of whole brain in seizure stage with different clip-selection thresholds (27 patients).

DOI: 10.1016/j.ebiom.2025.106085