Compact Propagation and Morphology-Based EEG Features for Real-Time Seizure Detection Using Machine Learning

Accurate and efficient seizure detection remains a major challenge for portable EEG-based monitoring systems, where computational and power limitations restrict the use of deep learning approaches. We introduce a compact set of fourteen propagation-based EEG features that quantify spike directionality, propagation velocity, and spatial coherence across channels. These physiologically interpretable features were evaluated using a gradient-boosted XGBoost classifier on the TUH Seizure Corpus under a leave-one-subject-out cross-validation framework. The proposed model achieved 97.8% accuracy, 98.7% specificity, 93.6% sensitivity, and a weighted F1 score of 97.8% using 2 s epochs and eight electrodes. The framework remained robust across generalized, focal, and complex partial seizure types and maintained consistent short-window performance. The compact 14-feature representation enables efficient, accurate, and interpretable seizure detection with strong potential for real-time wearable EEG-based applications. The proposed gradient-boosting approach demonstrates that shallow, interpretable architectures can achieve performance comparable to deep learning methods while offering improved computational efficiency, making them promising for low-power embedded implementations.