Deep learning for sudden cardiac arrest risk prediction: a meta-analysis of discriminative performance
D Koeckerling, J Barker, H Aggour, A Birdi, L Pastika, A El Medany, J B Guichard, K Patlatzoglou, J Tonko, G R Khattak, N Peters, S Engelhardt, N Frey, A Varnarva, F S NgAbstract
Background
Sudden cardiac arrest is a leading cause of premature mortality, and current preventive strategies, largely based on left ventricular ejection fraction, incompletely capture individual risk. Deep learning offers potential for improved risk prediction but reported performance and clinical readiness remain uncertain.
Methods
We conducted a PRISMA-compliant systematic review, meta-analysis and meta-regression. Medline, Embase, CINAHL, CENTRAL, Emcare, Scopus and Web of Science were searched through August 2025 for English-language studies using deep learning to predict sudden cardiac arrest or sustained malignant ventricular arrhythmia outcomes. Data extraction and AI reporting quality assessment were performed in duplicate. Discrimination was summarised using the area under the receiver operating characteristic curve (AUC). AUCs were pooled under random-effects models fitted using restricted maximum likelihood estimation. Prespecified subgroup analyses examined unselected cohorts, structural heart failure cohorts and genetic disease cohorts. Study-level moderators were evaluated using random-effects meta-regression.
Results
Twenty-seven studies reporting 28 eligible deep learning models were included, representing over 1 million participants, predominantly from retrospective, multicentre cohorts in high-income health systems. Model testing was conducted on a combined 116,288 participants, with the remainder being used for model development and training. Quantitative synthesis was possible for 25 studies, yielding a pooled AUC of 0.83 (95% CI: 0.78 to 0.86) with high heterogeneity (I² 97%). Discriminative performance was highest in unselected cohorts (AUC 0.85, 95% CI: 0.81 to 0.89) and genetic disease studies (AUC 0.84, 95% CI: 0.77 to 0.90), and comparatively lower in structural heart disease cohorts (AUC 0.73, 95% CI: 0.62 to 0.81). Longer predictive time horizon was associated with reduced discrimination in multivariable meta-regression (beta -0.95, p=0.036). AI reporting quality was highly variable, and no study was completely reported; key deficits included limited description of model development and validation, seldom external validation, sparse data and code sharing, and absent fairness assessment.
Conclusions
Deep learning models demonstrate promising performance metrics for sudden cardiac arrest risk prediction, yet discriminatory capacity was reduced with longer predictive time horizons and in populations with structural heart disease. Beyond discriminatory performance, substantial heterogeneity, incomplete reporting and infrequent external validation, fairness evaluation and regulatory readiness currently limit clinical translation.Graphical abstract