<![CDATA[Epilepsy is a neurological disorder marked by recurrent seizures, which can greatly reduce patients' quality of life. Early and accurate seizure prediction is essential for effective clinical intervention and patient safety. This study proposes and evaluates a seizure prediction system using EEG signals processed through machine learning techniques combined with optimized feature extraction methods. The research contribution is the comprehensive comparative analysis of classifier-feature pairs for identifying the most effective configuration for seizure prediction tasks. Three classifiers—Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost)—were systematically compared, each combined with precisely engineered feature extraction methods, including Common Spatial Pattern (CSP), Discrete Wavelet Transform (DWT), statistical features, and frequency domain features. EEG data from seven patients, totaling approximately 68 hours with 40 seizure events, were obtained from the Children's Hospital Boston database. The results demonstrate that XGBoost with CSP features achieved the highest overall accuracy at 88% and specificity at 88%, while XGBoost with DWT features reached the highest sensitivity at 87%. Additional metrics including F1-score (0.85) and AUC-ROC (0.91) confirmed XGBoost's superior performance. Comparison with five recent studies showed our approach offers a 3-5% improvement in accuracy and sensitivity. These findings highlight the critical impact of both classifier selection and feature engineering in improving EEG-based seizure prediction, with implications for developing real-time monitoring systems despite challenges in clinical implementation due to inter-patient variability.]]>
