HFO-Guided Epileptogenic Zone Mapping With Deep Learning

Abstract

The precise localization of the epileptogenic zone in drug-resistant epilepsy remains a major clinical challenge, and high-frequency oscillations (HFOs) are considered one of the most promising biomarkers for addressing this challenge. However, the development of automated methods for HFO analysis has long been hindered by a lack of publicly available expert-annotated data, inter-center heterogeneity in recordings, and the absence of a unified framework for correctly comparing different approaches. This study conducted a controlled comparison of supervised, self-supervised, and semi-supervised approaches to the classification of pathological HFOs on the multi-center Omni-iEEG dataset within a unified experimental framework using CNN and Transformer encoders, with evaluation performed both at the event level and at the downstream level of pathological channels. The results showed that supervised learning provides a strong baseline, the best performance at the event level is achieved with self-supervised pretraining followed by fine-tuning, and semi-supervised learning allows for competitive results even with a limited amount of labeled data. However, improvements in classification quality at the event level are only partially transferred to downstream clinical evaluation, indicating the need for further development not only of the models but also of the clinical problem formulation itself.