Machine learning for prediction of sudden cardiac death and sudden cardiac arrest across clinical and wearable settings: a systematic review

Introduction. SCD causes approximately 4–5 million deaths annually. Current guidelines and risk scores have only modestly increased survival rates. These could be improved thanks to predictive models using ML. Material and methods. This review, registered in PROSPERO and following PRISMA 2020 standards, evaluates ML-based models for SCA prediction. Two reviewers independently screened studies retrieved by a four-element search strategy and extracted performance metrics. Quality was assessed via PROBAST and GRADE frameworks to ensure reliable evidence across clinical settings and prediction horizons. Results. Across included studies, three categories were identified by prediction horizon: long-term risk prediction (n = 7), short-term in-hospital prediction (n = 8), and imminent prediction (n = 2). Long-term models ranged from small ECG/HRV studies (accuracy ~95–99%, likely inflated) to large registry-based models achieving AUROC 0.80–0.94, several with external validation. Short-term in-hospital models demonstrated consistently strong performance (AUROC ~0.85–0.95) and frequently outperformed conventional early-warning scores, with several including external validation. Imminent-prediction models forecast ventricular fibrillation minutes to over an hour before onset with high accuracy but remain small, proof-of-concept studies with limited real-world validation. Conclusions. ML-based models may help address current gaps in risk stratification, particularly in emergency and ambulatory settings, enabling prediction with promising reported performance in controlled settings. However, most remain experimental and lack external validation or prospective designs. Overreliance on small, curated datasets further limits generalizability. These findings require cautious interpretation before routine clinical implementation.