Classification of epilepsy seizure types in pediatrics based on Turkish EEG reports

Abstract

This study focuses on the binary classification of pediatric epilepsy seizure types as focal or generalized using Turkish electroencephalography (EEG) reports, leveraging natural language processing (NLP) and machine learning methodologies. A novel dataset comprising 130 Turkish EEG reports was developed and publicly released, addressing the scarcity of resources in this domain. The study employed various text representation models, including TF-IDF, FastText, ElectraTR, XLM, and BERTurk, along with classifiers such as Logistic Regression, Support Vector Machines, and CatBoost. The highest classification performance was achieved using BERTurk embeddings combined with Logistic Regression, yielding an accuracy of 96.6 %. This work is significant for being the first to explore focal versus generalized seizure classification from text-based EEG reports in Turkish. It underscores the critical role of contextual embeddings in handling morphologically rich languages and demonstrates the potential of NLP techniques in advancing pediatric epilepsy diagnostics. The findings pave the way for automating diagnostic processes and improving efficiency in clinical settings. Future research aims to expand the dataset, incorporate EEG signal data, and refine the models for broader applicability.