EPI-SPA Net: hypergraph-based deep learning framework for seizure detection and propagation analysis in EEG signals
Epilepsy is a chronic neurological disorder characterized by recurring seizures brought on by aberrant brain electrical activity. However, existing methods fail to effectively capture complex spatio-temporal dependencies in noisy electroencephalogram (EEG) signals. These limitations reduce the robustness and reliability of automated seizure detection. To overcome this, a novel EPI-SPA Net has been proposed for automated seizure detection using EEG signals. The Empirical Wavelet Transform (EWT) is employed to suppress noise and artifacts while preserving seizure-related frequency components. A Spatio-Temporal Hypergraph Neural Network (STH-NNet) is introduced to extract discriminative EEG features by jointly learning higher-order spatial connectivity and long-range temporal dependencies. The extracted features are then further processed using a Capsule Network to classify each 4-s EEG window as seizure or non-seizure. A temporal convolutional segmentation network (TCSN) converts the window scores into seizure events with onset and offset times. Finally, an exploratory lobe-level propagation analysis measures the order and delay of lobe recruitment directly from the EEG, and a Graph Sample and Aggregate (GraphSAGE) network with a Directional Conformer estimates this order. The proposed EPI-SPA Net achieves an overall accuracy of 93.1%, a precision of 96.0%, a recall of 55.7%, an F1-score of 70.5%, and a Cohen's κ of 0.669. Lobe recruitment orders were more consistent within patients than between patients (Kendall τ 0.371 vs. 0.008, p < 0.001). Thus, the proposed EPI-SPA Net combines automated seizure detection with an exploratory analysis of seizure propagation from scalp EEG, which may support the clinical assessment of epileptic seizures.
Authors
- R. Manjupriya
- A. Anny Leema
Institutions
- Vellore Institute of Technology University (IN)
Publication Details
- Journal
- Frontiers in Artificial Intelligence
- Published
- 2026-11-16
- DOI
- https://doi.org/10.3389/frai.2026.1962680
- Primary Topic
- EEG and Brain-Computer Interfaces
- Type
- article
- Field-Weighted Citation Impact
- 0.00