Towards Intelligent Epilepsy Detection

Objectives: This study aims to develop a robust and intelligent diagnostic system for automated epileptic seizure detection using Electroencephalography (EEG) signals. The system addresses key limitations of existing methods, particularly patient-specific variability and reduced clinical performance, and provides interpretability via feature importance analysis. Methods: EEG data from the Bonn University dataset were preprocessed using a second-order Butterworth filter to remove artifacts such as power line noise, followed by normalization. A sliding window technique expanded the dataset from 500 to 5500 segments. From each segment, statistical features (mean, standard deviation, skewness, kurtosis) and frequency-domain features (power spectral density of delta, theta, alpha, beta bands) were extracted. A Random Forest classifier was trained on these features, with hyperparameters optimized via GridSearchCV. The dataset was divided into 80% training and 20% testing. Results: The initial Random Forest model achieved 95% accuracy on the test set, with strong recall scores (97% for class Z, 96% for class S) and high precision (1.0 for class O). After tuning (n_estimators=200, max_depth=20), test accuracy remained at 95%, while cross-validation accuracy reached 92%. Feature importance analysis revealed that delta and theta band power, along with statistical metrics, were the most discriminative features. Conclusion: The proposed framework offers a reliable, interpretable, and efficient solution for seizure detection. Its performance and feature transparency support clinical relevance. Future work will focus on real-time implementation via wearable EEG devices, use of explainable AI techniques, and extending the system to diagnose other neurological conditions.