Discrimination of seizure topographical images utilizing transfer learning approaches

Abstract Epilepsy is one of the most common neonatal neurological disorders, causing significant nervous system dysfunction. Early detection of seizures is critical for timely clinical intervention and reducing long-term neurological consequences. This study proposes an automated framework utilizing deep transfer learning architectures to classify neonatal seizures using topographic brainpower maps extracted from EEG signals. We utilized a publicly accessible dataset comprising recordings from 79 newborns, analyzed across three temporal window scenarios (1, 5, and 8 s). Three deep transfer learning models, GoogLeNet, ResNet-18, and ShuffleNet, were rigorously evaluated. While initial results approached the optimal accuracy across all temporal windows, a more rigorous subject-wise protocol was implemented to ensure clinical generalizability, where the models were tested on entirely unseen patients. Under this strict validation, ResNet-18 emerged as the superior architecture, achieving a remarkable peak subject-wise accuracy of 97.1% in the 8-second window scenario. Our findings confirm that the proposed framework is highly robust against inter-patient variability, representing a reliable and methodologically sound strategy for automated neonatal intensive care monitoring.