Multi-model AI-ECG ensemble with rule-based consensus quantifies predictive uncertainty and identifies enriched high-risk subgroups for 1-year lethal ventricular arrhythmia

Background Artificial intelligence electrocardiography (AI-ECG) enables scalable prediction of ventricular arrhythmia (VA) (1). Whether independent representation-learning paradigms capture complementary arrhythmic risk signals and improve time-dependent discrimination remains uncertain. Methods Three independently developed AI-ECG models were evaluated in a USA held-out cohort using a 30-day landmark and one ECG per patient. Models were: esCAPE, an embedded survival contrastive model trained on 5M multicontinental ECGs; CELA, contrastive alignment between 12-lead ECG signals and 12SL reports trained on 10M USA ECGs; and SurvResNet, a discrete-time survival ResNet trained on 1M USA ECGs (2). Time-dependent discrimination from 3 months to 10 years was assessed using inverse probability of censoring weighted AUC(t). Equal-weight 2- and 3-model ensembles were constructed on the risk scale. The primary endpoint was 1-year VA. Model-specific IPCW-Youden thresholds were used to derive a 3-model Consensus Confidence Index (CCI; 0–3). Results Among 50,563 patients event-free at 30 days (64 events within 1 year; 392 total thereafter), the 3-model ensemble achieved the highest time-dependent discrimination at every evaluated horizon from 3 months to 10 years. At 1 year, AUC(t) was 0.804 for esCAPE, 0.789 for CELA, and 0.818 for SurvResNet, increasing to 0.825 with the 3-model ensemble. Across horizons, ensemble gains over the strongest individual model ranged from 0.002 to 0.009 and were sustained through 10 years (0.775 vs 0.766). Model agreement provided powerful risk stratification. The Consensus Confidence Index (CCI) distribution was 52%, 23%, 14%, and 11% for scores 0–3. Relative to CCI 0, hazard ratios were 1.72 for CCI 1, 3.24 for CCI 2, and 12.01 for CCI 3 (all p < 0.001), demonstrating a strong monotonic gradient of risk. In known 1-year status analysis, CCI ≥ 2 yielded LR+ 3.28 and LR− 0.27, while CCI = 3 yielded LR+ 5.70, identifying a markedly enriched high-risk subgroup. Conclusion Independent AI-ECG representation-learning paradigms capture complementary arrhythmic risk information. Equal-weight multi-model consensus improves time-dependent discrimination across both short and long horizons and provides intrinsic uncertainty quantification through model agreement. High concordance between models identifies a substantially enriched subgroup that may be suitable for risk-guided intervention.