Real-time EEG-based epileptic seizure prediction using artificial intelligence: A systematic review.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epilepsy impacts 50 million globally, with AI seizure prediction facing a significant bench-to-bedside gap.
• Current AI translation is hindered by limited generalizability and lack of standardized real-world deployment evaluation.

Purpose of the Study:

• Evaluate the operational readiness of AI models for real-time epileptic seizure prediction.
• Examine AI architectural performance, real-time responsiveness, interpretability, and multimodal integration for clinical deployment.

Main Methods:

• Systematic review adhering to PRISMA 2020 guidelines, searching major scientific databases (PubMed, Scopus, IEEE Xplore, ScienceDirect) from Jan 2017 to May 2025.
• Included 23 studies on EEG-based real-time prediction, assessing risk of bias using ROBINS-I, and standardizing false-alarm rate (FAR) to events per hour.

Main Results:

• Deep learning models, particularly CNNs and hybrids, demonstrated high sensitivity (up to 99.81%).
• Over-reliance on the CHB-MIT dataset (74% of studies) and predominantly patient-specific validation (PS) were noted.
• Limited reporting on prediction horizon, end-to-end latency, energy metrics, and lack of patient-independent (PI) or external validation (EXT) were critical findings.

Conclusions:

• High accuracy in AI seizure prediction is insufficient for clinical translation due to a lack of prospective validation and standardized deployment metrics.
• Proposed operational framework emphasizes standardized evaluation (PS/PI/EXT), defined latency budgets, and prospective FAR reporting for clinical viability.