SeizureFed: Federated Learning for Pediatric Seizure Detection through EEG Using E-ResNet

Ensuring rapid and efficient in-clinic treatment, maintaining patient confidentiality, and developing automated pediatric seizure detection algorithms have been important objectives in epilepsy research. In recent decades, EEG has become a significant tool for neurologists to diagnose neurological disorders, particularly for detecting seizures. Seizures are unpredictable and can occur suddenly, making timely detection crucial. Given the sensitive nature of neurological data, it is essential to ensure data privacy. However, constructing an effective model requires a substantial amount of data, presenting a significant challenge. Thus, the primary objective of this research is to develop a precise and automated model for identifying pediatric seizures utilizing EEG data while preserving patient confidentiality. This study presents a novel framework called E-ResNet, employing a modified Convolutional Neural Network (CNN) to classify seizure activity by ensembling ResNet-50 and ResNet-18. Additionally, federated learning for pediatric seizure detection is suggested to encourage a collaborative and privacy-aware methodology that can be employed in healthcare settings with numerous data sources. Utilizing the publicly available CHB-MIT database, the proposed model achieved an accuracy of 99%, 92.3% F1-score and an AUC of 91%. The proposed framework is an effective architecture for pediatric seizure detection that incorporates E-ResNet alongside CNN and federated learning techniques.