Clinical readiness and methodological rigor of AI models for ventricular arrhythmia and sudden cardiac death prediction: A systematic review with PROBAST+AI appraisal.

Abstract

Sudden cardiac death (SCD) remains a major cause of mortality, and risk stratification based on left ventricular ejection fraction (LVEF) has significant limitations. Artificial intelligence (AI) has emerged as a potential approach for modeling complex arrhythmic risk. We conducted a systematic review with prespecified anchor subgroup reporting (without pooling) to evaluate the performance of AI models in predicting SCD and malignant ventricular arrhythmias. We searched MEDLINE/PubMed, Embase, and Scopus from inception to 3 October 2025. Anchor subgroup reporting was restricted to studies reporting area under the receiver-operating characteristic curve (AUROC) with 95% confidence intervals (CIs) within predefined modality/horizon subgroups. Studies without CIs or relying on cross-validation variance were summarized narratively as supplementary evidence. Thirteen studies met the inclusion criteria; five reported AUROC with 95% CIs within prespecified modality/horizon subgroups and were included as anchor subgroup estimates (k=1 per subgroup). Methodological quality and reporting were limited: 54% of studies (7/13) were classified as high risk of bias according to PROBAST+AI, driven by cross-validation-only designs, unit-of-analysis issues, limited external validation, and sparse calibration reporting. Among anchor estimates, deep learning on single-lead ambulatory ECG showed high discrimination for near-term ventricular tachyarrhythmias (AUROC 0.948, 95% CI 0.926-0.967), while 12-lead ECG models achieved good performance for long-term SCD (AUROC 0.820, 95% CI 0.794-0.847). The multimodal anchor estimate (AUROC 0.84) was numerically higher than the CMR radiomics anchor (AUROC 0.70), but this represents an indirect between-study contrast. AI-ECG models show high reported discrimination in near-term risk windows, but cross-study comparisons are indirect, and methodological heterogeneity remains a barrier to translation into clinical practice.