Does a multi-domain feature engineering approach with classical machine learning accurately detect seizures in paediatric EEG recordings?
A patient-independent machine learning approach using 15 multi-domain EEG features achieves over 90% accuracy for seizure detection in paediatric patients, offering a computationally efficient method suitable for near real-time monitoring.
Seizure detection from scalp EEG recordings requires features capable of capturing the rapidly evolving patterns of cerebral activity. We present a strictly patient-independent approach that investigates 1-second EEG windows using nested cross-validation to prevent data leakage within the paediatric CHB-MIT dataset. In this approach, we combine features extracted from temporal, spectral, wavelet, and spatial (CSP) domains. Feature selection, spatial filtering, and hyperparameter optimization are performed only within training folds to avoid leakage. A compact feature set of just 15 features enables traditional classifiers to achieve remarkable performance: a K-Nearest Neighbours classifier achieves 90.6% accuracy, whereas a Random Forest achieves 90.5% accuracy with an AUC of 0.927. The proposed system is interpretable, highly computationally efficient, and suitable for near real-time clinical monitoring.
Ghosh et al. (2026) studied this question.