Real-Time EEG-Based Epileptic Seizure Prediction Using Artificial Intelligence: A Systematic Review

Epilepsy affects approximately 50 million people worldwide, and seizures remain difficult to predict in onset, severity, and duration. Real-time seizure prediction may enable proactive intervention and improve patient safety and quality of life. Despite the development of high-performing algorithms, translation remains limited by predictive accuracy, interpretability, and generalisability. This systematic review evaluates artificial intelligence (AI) models for real-time epileptic seizure prediction and assesses machine learning and deep learning performance alongside real-time responsiveness, interpretability, and multimodal integration. We searched PubMed, Scopus, IEEE Xplore, and ScienceDirect (1 Jan 2017–31 May 2025); 23 studies met eligibility. Inclusion required EEG-based real-time prediction with model/dataset/validation details and sufficient methods for ROBINS-I risk-of-bias assessment. Two reviewers independently screened, extracted, and cross-checked data. False-alarm rate (FAR) was standardised to /h. Deep learning consistently outperformed conventional machine learning. CNNs and hybrid CNN–recurrent architectures reported up to 99.01% accuracy, 99.81% sensitivity, and 97.70% specificity. Validation was predominantly patient-specific; three studies adopted patient-independent schemes, and none reported cross-dataset external validation. Reporting of prediction horizon was limited, and latency/energy metrics were rarely provided. Despite promising accuracy, the lack of real-world validation and the under-reporting of deployment metrics hinder clinical translation. We recommend standardised evaluation (PS/PI/EXT), an end-to- end latency budget (pre-processing, model runtime, I/O, alert handling), and prospective FAR per 24 h, with prospective and cross-dataset validation to enhance reliability and patient outcomes. We synthesised 23 real-time EEG seizure-prediction studies (2017–2025). Deep learning peaks: Acc 99.01%, Sens 99.81%, Spec 97.70% (not one model). Validation was patient-specific; no cross-dataset external validation reported. Latency and energy metrics were rarely reported; FAR harmonised to /h. Report PS/PI/EXT, end-to-end latency budget, and FAR per 24 h.