Predicting epileptic seizure dynamics is a complex computational challenge due to the highly nonlinear and variable nature of brain activity. Traditional machine learning models excel at capturing statistical patterns but often struggle with generalization, while physics-based models provide principled descriptions of neural dynamics but lack adaptability to patient-specific data. In this work, we propose a hybrid physics-informed machine learning framework that integrates the Kuramoto coupled-oscillator model with neural ordinary differential equations (ODEs) to improve the predictive modeling of seizure dynamics from multi-channel EEG data. By leveraging the strengths of both approaches, our method enhances robustness, interpretability, and generalizability across different seizure patterns. Using the Temple University Seizure Corpus (TUSZ), we demonstrate that our model outperforms both purely data-driven and physics-based baselines in predicting future EEG signals across different types of epileptic seizures. This suggests that our approach can improve seizure prediction while contributing to the computational modeling of neural activity. By integrating data-driven learning with physics-based principles, our method provides a promising framework for further exploration in seizure forecasting and neuroscience applications.Clinical relevance— Epilepsy affects millions worldwide, with many patients requiring continuous monitoring due to drug-resistant seizures. Reliable seizure prediction enables timely interventions, improving patient safety and quality of life. Our hybrid model enhances robustness and adaptability to individual seizure patterns, making it a promising approach for more effective, patient-specific monitoring and early warning systems.