Enhancing seizure detection with hybrid XGBoost and recurrent neural networks

Abstract

Epileptic seizures are sudden and unpredictable, posing serious health risks and significantly affecting the quality of life of patients. An accurate and timely prediction system can help mitigate these risks by enabling preventive measures and improving patient safety. This study investigates machine learning and deep learning algorithms for seizure prediction, comparing their effectiveness on a large EEG dataset of epileptic patients. Signal processing techniques were applied to enhance data quality, and all models were trained on the same dataset for binary classification. Sixteen models were evaluated, including traditional classifiers such as Logistic Regression, K-Nearest Neighbors, Decision Trees, ensemble methods that include Random Forest, Gradient Boosting, and advanced techniques such as Extreme Gradient Boosting, Support Vector Machines, Gated Recurrent Units, and Long Short-Term Memory networks. Performance was assessed using multiple evaluation metrics on both training and validation datasets. While simpler models showed varied accuracy, ensemble and deep learning models performed significantly better, with hybrid approaches demonstrating strong generalization. Results show that whereas ensemble and deep learning models far exceeded simpler models, their accuracy varied. AUC of 0.995 and accuracy of 98.2% on validation data and 0.994 AUC with 96.8% accuracy on test data were obtained by the proposed hybrid Model integrating XGBoost with RNN-based architectures (LSTM and GRU). High recall (96.2%) shown by the Model guarantees minimal false negatives and is important for clinical uses. Furthermore, EEG signal preprocessing methods improved data quality, raising classification accuracy. This Model can be implemented for real-time monitoring using wearable devices, enabling continuous patient observation and remote healthcare applications.