Multiclass Epileptic Seizure Detection Using Few-Shot Learning with Federated Learning Extension

Epilepsy has been one of the most prevalent neurological disorders, affecting over 50 million people worldwide. Reliable and timely seizure detection has been crucial for improving patients' quality of life and guiding appropriate treatments. While deep learning has shown significant promise in this domain, its effectiveness has depended on large, well-labeled, and centralized datasets. However, medical data has often been scarce and distributed across multiple institutions, with privacy regulations making data sharing infeasible. To address these challenges, we have explored both centralized and federated few-shot learning (FSL) for multiclass epileptic seizure detection, enabling effective model training under data constraints while preserving patient privacy.
We have evaluated three classifier architectures—a similarity-based classifier, a fully connected neural network (FCNN), and a transformer—under both centralized and federated settings. Additionally, we have analyzed how varying the number of support set samples influences model performance. Our results have shown that the FCNN consistently achieved the highest performance across different sample sizes in both settings. However, transformers occasionally demonstrated stronger discriminative capabilities between seizure types. Interestingly, we discovered that increasing the support set size did not always lead to performance improvements, highlighting the complexities of few-shot learning in seizure detection.
In the federated setting, we have successfully implemented a privacy-preserving FSL approach and identified key challenges, such as the impact of class imbalance across institutions on model generalization. Our findings have revealed that standard federated learning algorithms were not well-suited for scenarios with limited data. Despite these limitations, centralized FSL has achieved strong performance, while federated FSL has remained a promising solution for EEG-based seizure detection in privacy-sensitive environments.
This study has introduced a novel approach to centralized FSL and laid a foundation for federated FSL. Our results have demonstrated meaningful progress in privacy-preserving seizure detection with limited data, paving the way for future advancements in both centralized and federated FSL for epileptic seizure classification.