EEG Seizure Prediction via Ensemble Learning and Feature Selection

Abstract

Accurate predictive modelling in EEG-based seizure forecasting is challenging. It is constrained by significant inter-patient variability and limited patient-specific data. This leads to poor generalisation in conventional pooled learning approaches. This study proposes a patient-specific seizure prediction framework that preserves feature heterogeneity while improving predictive performance under constrained data conditions. The proposed approach integrates a hierarchical filter-wrapper feature selection pipeline with a hybrid ensemble model. EEG signals are segmented into seizure and non-seizure states using temporal windowing, and features are extracted using a scalable time-series framework. Model evaluation is conducted using patient-wise temporal cross-validation. This helps to preserve chronological structure and prevent data leakage. Experimental results demonstrate that patient-specific modelling significantly outperforms the merged dataset. The patient-specific model achieves an average F1 Score of 0.973 (ROC AUC of 0.952). Ablation analysis confirms that the ensemble architecture results in performance gains, while feature selection improves model efficiency and stability. Moreover, SHAP-based interpretability states distinct patient-specific EEG biomarkers. The results firmly support the necessity for patient-specific modelling approaches. The findings also highlight the importance of protecting the personalisation of features in EEG-based prediction tasks. The findings further establish that our solution of patient-specific models provides a robust, scalable, and interpretable solution. Moreover, the proposed individualised model solution exhibits excellent results in heterogeneous and limited data clinical situations.