Ensemble Models for Epileptic Seizure Recognition: a Comparative and Explainability Analysis

Epileptic seizure detection is a critical task in neurophysiological analysis. This paper presents a robust pipeline using ensemble learning techniques, specifically Random Forest and Stacking Classifier, to achieve high accuracy in seizure recognition. The SMOTE method is used to address class imbalance. Further, feature scaling and SHAP (SHapley Additive exPlanations) analysis enhance the understanding of feature importance and model interpretation. Extensive experimentation on the Epileptic Seizure Recognition dataset demonstrates the superiority of ensemble methods, achieving 97.8 % accuracy with the Stacking Classifier. Comparative results, feature importance analysis, PCA visualizations, and misclassified sample analysis are provided to offer a comprehensive understanding of model performance. Future work includes integrating real-time prediction frameworks for clinical applications.