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

Epilepsy affects approximately 50 million people worldwide. While AI-driven seizure prediction shows technical promise, the transition from experimental prototypes to clinical tools is hindered by a significant 'bench-to-bedside' gap Current translation remains constrained by limited generalisability and the absence of standardised evaluation for real-world deployment.

This systematic review evaluates the operational readiness of artificial intelligence (AI) models for real-time epileptic seizure prediction and examines architectural performance alongside deployment-critical factors such as real-time responsiveness, interpretability, and multimodal integration.

Adhering to PRISMA 2020 guidelines we searched PubMed, Scopus, IEEE Xplore, and ScienceDirect (1 Jan 2017-31 May 2025), supplemented by targeted citation tracking (up to Feb 2026); 23 studies met eligibility criteria. Inclusion required EEG-based real-time prediction with model/dataset/validation details and sufficient methodological information for ROBINS-I risk-of-bias assessment. Quality and bias were assessed independently using the ROBINS-I tool. Two reviewers screened records and extracted data; false-alarm rate (FAR) was standardised to events per hour.

Deep learning architectures consistently outperformed conventional machine learning. CNNs and hybrids reported sensitivity up to 99.81%; however, a critical finding is the over-reliance on the pediatric CHB-MIT dataset, which accounts for 74% of included studies. Validation was predominantly patient-specific (PS); only three studies adopted patient-independent (PI) schemes, and none reported cross-dataset external validation (EXT). Reporting of prediction horizon was limited, and end-to-end latency or energy metrics were rarely provided.

Despite high accuracy, the lack of prospective validation and standardised deployment metrics remains a major barrier to clinical translation. We propose an operational framework requiring standardised evaluation (PS/PI/EXT), explicit end-to-end latency budgets, and prospective FAR reporting per 24 h, to ensure clinical viability and patient outcomes.