EEG-Based Seizure Prediction Approaches Within Clinically Relevant Pre-seizure Windows Using Scalp EEG Datasets: A Systematic Review

This systematic review synthesized evidence on EEG-based seizure prediction within clinically relevant pre-seizure windows, focusing on deep learning and machine learning models evaluated on public datasets. The review aimed to summarize predictive performance, including sensitivity, receiver operating characteristic curve findings when reported, and false-alarm outcomes for scalp EEG seizure prediction systems. It also examined methodological heterogeneity across studies, including the preictal window, defined as the period before seizure onset used for prediction; the seizure prediction horizon, defined as the minimum warning interval before seizure onset; and the seizure occurrence period, defined as the interval during which a seizure is expected after an alarm. Additional objectives were to compare validation strategies, including patient-wise generalization, and to identify design features most consistently associated with clinically actionable performance. The included evidence showed substantial variability in reported performance. It was limited by heterogeneous study designs, frequent reliance on patient-specific internal validation, limited external validation, inconsistent reporting of false alarms, and high or unclear risk of bias. Although scalp EEG seizure prediction models have shown potential to achieve moderate-to-high sensitivity in selected settings, confidence in their real-world generalizability remains limited. Overall, the findings suggest that clinical translation will require more standardized prediction windows, clearer reporting of alarm burdens, stronger validation frameworks, and prospective evaluations under real-world conditions.