Unsupervised Clustering of Pre-Ictal EEG in Children: A Reproducible and Lightweight CPU-Based Workflow

Eyad Talal Attar^*

• Electrical and Computer Engineering Department, Faculty of Engineering, King Abdulaziz University, Jeddah, Saudi Arabia

Background: Early identification of seizures in children is important for safety, intervention success, and quality of life improvement, because many children are unable to reliably communicate sensed preictal warning features. Recognition of pre-ictal EEG microstates is a path toward wearable and bedside monitors that may deliver actionable alerts to caregivers. However, most existing approaches remain constrained by manual labels, expert calibration, or computationally expensive models with limited clinical utility. Methods: The study developed an unsupervised clustering pipeline for pediatric pre-ictal EEG using PCA, UMAP, and K-Means, without the need for manual annotations or GPU resources. The CPU-based and open-source design makes the workflow accessible and potentially adaptable for future real-time neurodiagnostic applications. Results: PCA retained > 95% variance, confirming stable feature extraction. ICA reduced blink and line-noise artefacts by 85% and 34%, respectively, improving signal quality. Optimal cluster number (k=4) was identified via Elbow and Silhouette methods, revealing distinct and physiologically meaningful EEG microstates preceding seizure onset. UMAP embeddings showed well-separated clusters with a high initial Silhouette Score (0.779), indicating robust internal structure. Noise removal improved interpretability without compromising cluster validity. Conclusions: The unsupervised nature of the study approach provides experimental evidence for the demarcation of a number of distinct pre-ictal states. These are associated with changes in cortical excitability and network synchrony, consistent with the predicted dynamics of a model of epilepsy. This study should be regarded as a proof-of-concept that advances methodological aspects of unsupervised EEG clustering within this specific dataset. The findings are hypothesis-generating rather than conclusive, providing a preliminary platform for exploring automatic pre-ictal state monitoring without expert input.