Early Prediction of Epileptic Seizures Based on Multifractal Analysis and Optimized Graph Neural Networks Using Scalp EEG Data

Abstract

Early prediction of epileptic seizures remains an active research area in scalp electroencephalography (EEG) analysis. Current methods focused on this topic often rely on handcrafted features that insufficiently capture the multiscale nonlinear dynamics of preictal activity, process EEG channels independently without modeling brain connectivity, or demand computationally expensive manual hyperparameter tuning. To address these limitations, this study presents a framework integrating multifractal analysis (MFA), graph neural networks (GNNs), and a differential evolution algorithm (DEA) to classify non-overlapping one-minute EEG segments as preictal (within 60 min before onset) or reference (interictal) states. Five complementary MFA techniques extract nonlinear descriptors from each segment, which serve as node attributes in a graph of the 21 EEG channels. Unlike conventional approaches, this graph explicitly encodes neuroanatomical proximity to capture spatial brain dynamics and inter-channel topological dependencies. This graph-based representation allows the GNN to learn from the relational structure of the brain network, capturing spatiotemporal interactions critical for early prediction, while the DEA systematically optimizes the GNN architecture, eliminating subjective manual tuning. Under a segment-level validation protocol on the CHB-MIT database, the framework achieved 96.38% accuracy, 94.36% sensitivity, 98.41% specificity, and an AUC-ROC of 99.37%. These results demonstrate the effectiveness of combining multifractal descriptors, graph-based processing, and evolutionary optimization for segment-level discrimination. Nevertheless, patient-wise validation remains essential for clinical translation, and this work is consequently presented as a proof-of-concept demonstration.