Hybrid Optuna and Q Learning for Adaptive Hyperparameter Optimization in EEG Seizure Classification

Abstract Epileptic seizure classification using electroencephalogram (EEG) signals plays an important role in automated neurological diagnosis and patient monitoring. Although deep learning models have demonstrated promising performance for EEG-based seizure classification, their effectiveness strongly depends on efficient hyperparameter optimization and stable convergence behavior. Motivated by limitations in existing Optuna-optimized GRU–HiBERT–MLP framework, this paper proposes an adaptive hyperparameter optimization that combines Optuna and Q Learning for EEG seizure classification. The proposed framework integrates Bayesian optimization with reinforcement learning–based adaptive exploration to improve hyperparameter tuning efficiency convergence stability and overall classification performance. A GRU–HiBERT–MLP architecture is employed for preictal and ictal EEG classification using wavelet-transformed features extracted from the CHB-MIT scalp EEG dataset. Experimental results demonstrate that the proposed framework achieved a classification accuracy of 93.10% and a ROC–AUC score of 0.9774, outperforming standalone Optuna and Q-learning optimization approaches. Furthermore, convergence analysis demonstrated more efficient hyperparameter search and greater convergence stability, resulting in improved classification performance compared with standalone optimization approaches. SHAP and LIME explainability identify influential features and explain the model predictions for clinical EEG analysis. By integrating adaptive optimization and explainable artificial intelligence, the proposed system serves as an intelligent decision-support framework for epilepsy diagnosis and monitoring.