JMIR Publications

ABSTRACT

Background:

Cardiovascular disease kills approximately 20 million people each year, yet identifying who is at highest risk early enough to act remains difficult in routine practice. The 12-lead electrocardiogram (ECG) is inexpensive, universally available, and completed in minutes, but conventional physician interpretation captures only part of its prognostic signal. Artificial intelligence (AI)–enabled ECG analysis (AI-ECG) has shown promise for predicting cardiovascular outcomes, yet published estimates of its accuracy remain fragmented across different populations, AI architectures, and outcome definitions.

Objective:

We aimed to quantify the pooled prognostic discrimination of AI-ECG models for all-cause mortality, cardiovascular mortality, and composite major adverse cardiovascular events (MACE), and to identify key sources of between-study heterogeneity.

Methods:

We conducted a systematic review and meta-analysis of studies evaluating AI-based ECG analysis for cardiovascular prognostication (PubMed, Embase, Cochrane CENTRAL, Web of Science, IEEE Xplore; January 2015–June 2026). Eligible studies applied AI-based ECG models to predict all-cause mortality, cardiovascular mortality, or MACE in adults with at least six months of follow-up. Risk of bias was assessed using the PROBAST+AI tool. Pooled AUROC and hazard ratios (HRs) were estimated using logit-transformed and DerSimonian–Laird random-effects models, respectively (PROSPERO: CRD420261430251).

Results:

Twelve studies comprising more than 3.7 million patients across test and validation cohorts were included, published between 2020 and 2025 across seven countries. The pooled AUROC for all-cause mortality was 0.843 (95% CI 0.800–0.878; I²=99.9%; 7 studies). For cardiovascular mortality, the pooled AUROC was 0.854 (95% CI 0.796–0.897; I²=99.9%; 3 studies), and for MACE, 0.815 (95% CI 0.724–0.880; I²=96.2%; 2 studies). High-risk AI-ECG classification was associated with a pooled HR of 2.46 (95% CI 1.56–3.89; I²=97.3%; 5 studies) for long-term mortality. Sensitivity analysis restricted to externally validated cohorts yielded a pooled AUROC of 0.758 (95% CI 0.672–0.827). Nine of twelve studies were classified as low overall risk of bias. Substantial between-study heterogeneity was observed across all analyses.

Conclusions:

AI-ECG models discriminate cardiovascular risk with clinically meaningful accuracy across diverse populations and model architectures (pooled AUROC 0.843 for all-cause mortality; pooled HR 2.46 for high-risk classification). Performance consistently attenuates in external validation (pooled AUROC 0.758), and between-study heterogeneity is substantial, indicating that local validation is necessary before deploying any AI-ECG model in a new clinical setting. Clinical Trial: PROSPERO: CRD420261430251