CNN-Based Transformer and Vision Transformer Models for The Prediction of Epileptic Seizure Using Electroencephalogram in GAERS Model

Abstract

Epilepsy is a prevalent neurological disease defined by unprovoked seizures and loss of awareness. The patient’s quality of life is negatively impacted, creating a physical and psychological burden. Therefore, it is essential to predict seizures at the right time to take appropriate treatment measures. While most studies focus on human Electroencephalography (EEG) datasets, animal experiments remain essential for advancing our understanding of seizure prediction.Validated genetic models such as the Genetic Absence Epilepsy Rats from Strasbourg (GAERS) play a crucial role in exploring predictability and identifying biomarkers for controlled drug delivery and electrical stimulation studies. However, the number of prediction studies using GAERS remains limited, increasing the scientific relevance of our work. In the last decade, researchers have designed EEG-based prediction algorithms with transformer architectures. In this work, we used two transformer models: a CNN-based transformer and a vision transformer. We obtained better results from the CNN-based transformer compared to the vision transformer. The highest sensitivity achieved was 89.53% for the CNN-based transformer and 80.94% for the vision transformer, depending on time windows of 1 to 7 seconds before seizure onset. Short prediction intervals provide valuable insights into seizure dynamics and pharmacological responses, highly relevant for translating preclinical findings into clinical practice. Even short preictal intervals can offer meaningful information for guiding patient-specific treatments. Our findings suggest that the proposed method can support the development of clinically oriented therapeutic strategies in epilepsy, including controlled drug delivery systems and electrical stimulation-based interventions.