Detection of pre-ictal epileptic events using a self-attention based neural network from raw Neonatal EEG data

Epileptic seizures can occur unpredictably, making real-time monitoring and early warning systems critical, especially in neonatal patients, where timely intervention can significantly improve outcomes. Neonatal seizures are often subtle and difficult to detect, increasing the need for automated, early prediction methods to aid clinical decision-making. While machine learning models have been widely used for seizure detection, their application in preemptive seizure warning remains underexplored. In this study, we propose a self-attention-based neural network that processes raw EEG data to detect pre-ictal signals, enabling early seizure prediction. A key challenge in using attention mechanisms for EEG analysis is the computational burden of handling high-frequency, long-duration signals. To address this, we introduce a second-wise summary statistics-based embedding that significantly reduces the input sequence length while retaining essential features. We validate our model using a publicly available dataset of 79 neonatal patients with physician-annotated EEG recordings. Our classifier achieves a maximum accuracy of 91 percent in distinguishing pre-ictal and ictal events from non-ictal signals. Notably, we evaluate our model on completely unseen patients, demonstrating its potential for real-world applicability in neonatal seizure prediction. This study provides a proof-of-concept for a preemptive seizure warning system, paving the way for AI-driven neonatal epilepsy management and broader clinical applications in seizure detection.