A privacy-preserving federated spatiotemporal dynamic graph neural network framework for epileptic seizure prediction

Abstract

Abstract Epilepsy ranks among the most prevalent and debilitating neurological disorders, globally affecting an estimated 50 million individuals across all age groups and socioeconomic backgrounds [WHO, 2019]. In recent years, research on its prediction methods has made significant progress, driven by advancements in artificial intelligence technology. Epilepsy prediction models based on electroencephalogram (EEG) signals and deep learning have become an important research direction in the field of neuroscience, and related research results have shown an exponential growth trend. However, there are still several key problems that need to be solved in existing research: First, the mainstream model architecture are focused on traditional neural network framework, exclusively trained in centralized settings that are incompatible with the privacy regulations and data-sharing constraints governing real-world clinical environments; second, the deep learning approaches, including recent Graph Neural Network (GNN) models, using static graph modeling methods, which ignore the dynamic network topological evolution characteristics of EEG signals in the time-varying process, fails to effectively explore the high-order nonlinear correlation characteristics contained in the topological structure of brain functional networks; To address this limitation, this paper proposes a patient-dependent privacy-preserving Federated Learning framework that integrates an epilepsy prediction model DygonNet based on spatiotemporal dynamic graph neural network, a local learning model at each federated client, deployed within a cloud-based simulation environment. In the proposed architecture, the CHB-MIT, SWEC-ETHZ and the TJU-HH iEEG datasets are treated as three independent federated clients representing distinct clinical sites, each performing federated training on their private EEG data, while a central cloud server simulated on Google Colab Pro aggregates the model updates using the Federated Averaging (FedAvg) and FedProx algorithms without ever accessing raw patient recordings. The model defines the dynamic graph structure of EEG signals, innovatively introduces the Transformer model and dynamic graph neural network into the field of epilepsy prediction to fully learn the temporal and spatial characteristics of EEG signals, and proposes a hierarchical graph pooling mechanism based on the attention mechanism in a Federated environment. Experiments show that the model shows excellent epilepsy prediction performance on both public and private datasets.