Machine learning detection of epileptic seizure onset zone from iEEG.

Abstract

Accurate identification of seizure onset zones (SOZ) is essential for the surgical treatment of epilepsy. This narrative review examines recent advances in machine learning approaches for SOZ localization using intracranial electroencephalography (iEEG) data. Existing studies are analyzed while addressing key questions: What machine learning techniques are used for SOZ localization? How effective are these methods? What are the limitations, and what solutions can drive further progress in the field? This narrative review examined peer-reviewed studies that employed machine learning techniques for SOZ localization using iEEG data. The selected studies were analyzed to identify trends in machine learning applications, performance metrics, benefits, and challenges associated with SOZ identification. The review highlights the increasing adoption of machine learning for SOZ localization, mostly with supervised approaches. Particularly support vector machine (SVM) using high frequency oscillation (HFO) biomarker feature being the most prevalent. High accuracy and sensitivity, especially in studies with smaller sample sizes are reported. However, patient-wise validation reveals limited generalizability. Additionally, ambiguity in SOZ definition and the scarcity of open-access iEEG datasets continue to hinder progress and reproducibility in the field. Machine learning offers significant potential for advancing SOZ localization. Development of more robust algorithms, integration of multimodal data, and greater model interpretability, can improve model reliability, ensure consistency, and enhance real-world applicability, thereby transforming the future of SOZ localization.