Investigators screened 512 records, reviewed 27 full texts, and included 13 studies that met prespecified criteria. The review covered publications from 2016 through 2025. Included cohorts originated from the United States, Canada, Taiwan, Switzerland, South Korea, the United Kingdom, and Brazil. Sample sizes varied widely — from several hundred ECGs to cohorts exceeding two million recordings — and the aggregated patient exposure approximated 5.9 million individuals based on study-reported samples. Two reviewers independently screened and extracted data; risk of bias assessments used PROBAST and are described in the source’s appendices.
Eleven studies reported on all-cause mortality, split across short-term (30-day to 1-year) and long-term (>1 year) horizons. Short-term models demonstrated AUCs commonly in the 0.80–0.89 range in internal or held-out validations. For example, Sun et al. reported an AUC of 0.843 for 30-day mortality and 0.812 for 1-year mortality on an internal holdout. Raghunath et al. reported a 1-year AUC of 0.88 across a very large Geisinger dataset; the model retained discrimination even among ECGs labeled clinically “Normal.” Tsai and Lin reported AUCs or C-indices approaching 0.89 for 1-year outcomes in large hospital-based cohorts. Long-term predictions also returned substantive discrimination: Sau reported a c-index of 0.775 for all-cause death over a mean follow-up of 5.46 years, and Tsai reported a C-index of 0.858 for 10-year all-cause mortality in internal validation.
These studies uniformly used raw waveform inputs rather than conventional human-read features. Investigators reported that AI models often outperformed traditional clinical risk scores in direct comparisons, though the review cautioned that heterogeneity in baseline populations complicates direct equivalence.
Four studies focused on cardiovascular mortality as an endpoint. Lin’s ResNet-18 survival model achieved a c-index of 0.891 for 1-year cardiovascular death using a large Chang Gung dataset. Sau’s BIDMC-derived model reported a C-index of 0.832 for cardiovascular death over approximately 5.5 years. Hughes’ SEER model showed internal AUC 0.83 with external AUCs of 0.78 and 0.83 in validation cohorts. Tsai’s survival deep learning model identified high-risk groups with markedly elevated hazard ratios for cardiovascular death in both internal and external validation. Investigators therefore described consistent discrimination across cardiovascular mortality studies, while also noting variability in follow-up windows and endpoint definitions.
Eight included studies examined MACE either as composite outcomes or as individual components. Four studies reported composite MACE definitions; Ouyang’s PreOpNet included mortality, cardiac arrest, myocardial infarction, heart block, and pulmonary oedema in its composite. Other studies reported individual components such as myocardial infarction, heart failure hospitalization, and stroke. Reported discrimination for composite and component endpoints varied by cohort and outcome definition. The review stressed that inconsistency in composite definitions and inclusion criteria limited direct comparison across studies and prevented pooled estimates.
Twelve of the 13 studies used deep learning architectures, most commonly convolutional neural networks and ResNet variants trained on raw ECG voltage-time data. One study compared a ResNet-based approach against an XGBoost method. Validation strategies ranged from internal holdout sets to external validation across institutions; a minority of studies reported multi-centre external validation. Investigators reported that heterogeneity in population case-mix, ECG input types (12-lead vs single-lead), endpoint definitions, follow-up durations, and overlapping regional datasets made statistical pooling inappropriate. The review authors therefore performed a narrative synthesis and explicitly avoided meta-analytic aggregation to prevent biased double-counting.
The authors applied PROBAST to assess bias and reported concerns tied to retrospective designs, variable endpoint definitions, and inconsistent external validation. Several large studies used overlapping regional data, which complicates cross-study comparisons. The review flagged the absence of standardized prognostic endpoints and inconsistent reporting of model calibration as key barriers to clinical translation. Investigators concluded that while AI-ECG models show consistent discrimination for mortality and cardiovascular events across diverse cohorts, prospective validation, harmonized outcome definitions, and broader external testing are essential before routine clinical implementation. Specific unanswered questions documented in the source include the need for prospective impact studies, standardized MACE composites, and assessments of calibration and decision-analytic utility in real-world clinical workflows.