Background Epilepsy affects approximately 50 million individuals worldwide, with 30% experiencing drug-resistant seizures despite optimal pharmacological management. Recent computational neuroscience advances have identified chimera states—spatiotemporal patterns where synchronized and desynchronized neural dynamics coexist—as potential biomarkers preceding seizures by 15–90 minutes. However, clinical translation faces critical challenges: (1) existing detection methods require extensive manual parameter optimization limiting scalability, (2) machine learning approaches show 20–35% accuracy degradation when applied to new patients, and (3) deep learning models lack the interpretability required for clinical validation. This paper seeks to answer the question: Can integrating physics-based constraints from Kuramoto oscillator theory with graph neural networks enable automated, robust, and interpretable chimera-based seizure prediction that generalizes across patients?. Methods We developed HP-GNN (Hybrid Physics-Informed Graph Neural Network), a novel architecture integrating data-driven learning with Kuramoto oscillator dynamics. The framework transforms multi-channel EEG into dynamic hypergraphs capturing higher-order neural interactions through: (1) adaptive hypergraph construction using Phase Locking Values with threshold τ = 0.65 for 3-clique detection, (2) three-layer hypergraph convolutions (64 → 128 → 256 dimensions), (3) Mamba state space networks achieving linear O(T) complexity, (4) physics-informed regularization with Kuramoto dynamics (weight λ₁ = 0.03), and (5) multi-task prediction heads. We employed two-stage training: self-supervised pre-training on 844 hours of continuous EEG, followed by supervised fine- tuning. Evaluation used 4-fold cross-validation on CHB-MIT (22 pediatric patients, 182 seizures) with external validation on IEEG.org (16 adults, 87 seizures). Results HP-GNN achieved 84.7% chimera detection accuracy (95% CI: 82.3–87.1%), representing 9.2% improvement over Delay Differential Analysis (75.5%, p < 0.001). Seizure prediction demonstrated 89.3% sensitivity with 68.2% maintained at 90-minute horizons, achieving 0.48 false positives per hour. Cross-patient generalization reached 79.8%, improving 14.6% over graph baselines. Physics constraints reduced training requirements by 35% (achieving 80% accuracy with 260 vs 400 patient- hours). Zero-shot transfer from scalp to intracranial recordings achieved 71.3% accuracy. GNNExplainer identified critical electrodes with κ = 0.68 agreement with neurologists. Learned parameters showed biological plausibility: synchronized components at 2.3 ± 0.5 Hz (delta), desynchronized at 9.1 ± 1.3 Hz (alpha). Conclusions Integrating physics-based constraints with graph neural networks enables robust seizure prediction addressing key deployment barriers. The combination of improved performance, cross- patient generalization, data efficiency, and clinical interpretability positions HP-GNN as a promising foundation for clinical seizure forecasting systems.
Amiri et al. (Thu,) studied this question.