An artificial intelligence pipeline for a generalizable identification of seizure onset zone in focal epilepsy

In drug-resistant focal epilepsy, surgical treatment represents the best chance to reduce seizure frequency, and accurate localization of the Seizure Onset Zone (SOZ) is critical for pre-surgery planning. However, the high inter-patient heterogeneity of stereo-electroencephalography (sEEG) recordings makes the development of generalizable, patient-independent methods particularly challenging. This study proposes an automated pipeline for the localization of the SOZ in multiple patients using a unique set of parameters. Meaningful features are extracted from sEEG measures: signal features characterize the time domain, spectral features capture information from the frequency domain, network features model the functional relationships between brain regions. A Variational Autoencoder (VAE) produces a latent representation of the resting-state features, while a Gaussian Mixture Model (GMM) fits the encoded distribution and scores the likelihood of seizure features. A classifier decides which brain regions compose the SOZ based on the negative log-likelihood of seizure features, which is used as an indicator for anomaly detection. Hyperparameter optimization was performed, and the model performance was estimated using both Leave-One-Out Cross Validation (LOOCV) and evaluation of a test set of unseen patients. The proposed model achieved promising results during cross-validation, however a significant performance drop was observed on unseen test patients, reflecting the intrinsic heterogeneity of sEEG recordings across subjects. Results highlight that inter-patient variability remains the primary challenge in SOZ localization, however potential strategies are proposed to direct future work towards patient-adaptive approaches.