Deep learning-based electrocardiography for sudden cardiac death risk prediction: a systematic review and meta-analysis of validation studies
J Gohri, A J Jolly, S R Rao, A S Y Singh Yadav, S K Kalra, A M MaheshwariAbstract
Background
Sudden cardiac death (SCD) remains a leading cause of mortality in heart failure populations. Traditional ECG parameters and guideline criteria (LVEF <35%) identify only ~50% of high-risk patients. AI-derived ECG features may capture subclinical arrhythmic substrate. Evidence synthesis on ECG-AI SCD prediction needed.
Methods
MEDLINE, Embase, clinicaltrials.gov (2018–January 2026) searched for prospective cohort studies using machine learning/deep learning on ECG signals to predict SCD. Inclusion: N≥100, ≥1-year follow-up, reported discrimination metrics (AUC/C-statistic) for SCD outcomes, comparison to conventional ECG risk scores. Exclusion: case-control designs, synthetic ECG data. Primary outcome: AUC for SCD prediction. Random-effects meta-analysis; subgroup analyses by patient population (HF vs population-based), ECG-AI architecture (CNN, LSTM, Vision Transformer).
Results
Thirty-two studies (n=18,500 patients; 1,247 SCD events; median 2.8-year follow-up) included. ECG-AI models achieved AUROC 0.75–0.90 (median 0.85) for SCD prediction, significantly outperforming conventional ECG risk scores (AUROC 0.62, p<0.001) and guideline-directed ICD criteria. In HF subgroup: ECG-AI combined with clinical variables achieved AUROC 0.90 (95% CI 0.86–0.94) versus conventional criteria 0.68 (p<0.001). Three-year SCD discrimination improved from 28.7% (conventional high-risk) to 36.1% (ECG-AI high-risk) (p for trend <0.001). External validation: minimal performance degradation (median ΔC-statistic −0.04).
Conclusions
ECG-based AI models significantly enhance SCD risk identification with improved discrimination over guideline criteria (AUROC 0.85 vs 0.62). Integration into ICD implantation protocols may improve primary prevention strategies. Prospective randomized trials assessing ICD therapy guided by ECG-AI warranted.