This review adheres to the PRISMA statement and CHecklist for critical Appraisal and data extraction for systematic Reviews of prediction Modelling Studies (CHARMS) (15, 16) (Supplementary Table 1). Furthermore, this study has been registered with PROSPERO (CRD420251012231). A systematic search of the literature published from March 2005 to March 2025 was conducted in the PubMed, IEEE Xplore, Medline, and Embase databases. The search terms used were (“Artificial Intelligence” OR “Machine Learning” OR “Deep Learning”) AND (“Electrocardiogram” OR “ECG”) AND (“Heart Failure” OR “congestive heart failure” OR “cardiac insufficiency” OR “Cardiac Failure”) AND (“Prediction” OR “Early Detection”). The inclusion criteria were: (1) Artificial intelligence models developed solely based on 12-lead ECG data for the prediction of heart failure; (2) Reporting of discriminative performance metrics (such as AUROC, c-statistic) and 95% confidence intervals (CIs); (3) Original research published in English and subjected to peer review. The exclusion criteria encompassed commentary, Conference Paper, reviews, non-English literature, models integrating non-ECG variables (such as laboratory indicators), and studies that did not provide complete performance metrics.

Two independent reviewers extracted data using a standardized form that captured study characteristics, including author, year, cohort, and sample size, as well as the AI model architecture and performance metrics. Discrepancies were resolved through consensus or third-party adjudication. The risk of bias and quality were assessed using the PROBAST + AI tool, which evaluates domains such as participants and data sources, predictors, outcomes, and analysis methodology (17). Studies were classified as having low, unclear, or high risk of bias based on predefined criteria. Cochran's Q test and the I²statistic were employed to quantify heterogeneity. A sensitivity analysis was conducted by sequentially removing individual studies based on a random effects model.

In a single study, we evaluated the c-statistic/AUROC of the model. When a study reported multiple cohorts and presented data for each cohort separately, we evaluated the model performance of each cohort in the study individually. Funnel plots were created to examine publication bias. We analyzed the discriminative ability through summary measures of the AUROC/c-statistic and the corresponding 95% CI. When the 95% confidence interval is not reported, we calculated it using the method described by Debray et al. (18). We calculated the 95% prediction interval (PI) to describe the degree of inter-study heterogeneity and indicate the possible range of the predicted model performance in new validations. Based on previous literature, summary AUROC/c-statistics were predefined as inadequate (<0.60), sufficient (0.60–0.70), acceptable (0.70–0.80), and excellent (>0.80) (19). We performed a meta-analysis using the metafor package in R (R Foundation for Statistical Computing 4.5.0). Our primary analysis evaluated the overall discriminative ability of all models for which cohorts reported AUROC/c-statistic data. In the secondary analysis, we compared the AI-ECG model with traditional FHS-HF/PCP-HF evaluation models. One researcher rated the certainty of the evidence for the primary results, while another researcher conducted a review.

A systematic literature search conducted across PubMed, IEEE Xplore, Medline, and Embase identified 154 initial records (Supplementary Table 2). After the removal of duplicates, 112 unique studies underwent title and abstract screening. Of these, 20 articles were excluded due to irrelevance and unavailability of full texts. The remaining 92 full-text articles were assessed for eligibility. A total of 87 studies were excluded for the following reasons: 52 were classified as reviews, commentaries, or conference papers; 16 studies did not integrate HF prediction models; 12 prediction models were not based on artificial intelligence or 12-lead ECGs; and two studies included non-ECG variables alongside ethnicity, gender, and age. Ultimately, five studies met all inclusion criteria and were included in the systematic review and meta-analysis (Figure 1). Applying the PICOTS framework to clarify the intended objectives or purposes of predictive model evaluation (Supplementary Table 3).

The analysis included five studies comprising 11 independent cohorts, totaling a sample size of 1,728,134 participants. The cohorts were predominantly from the United States (7 out of 11), with additional representation from Asia (1 out of 11, TSGH), Europe (1 out of 11, UKB), and South America (1 out of 11, ELSA-Brasil in Brazil). Sample sizes varied significantly across the cohorts, ranging from 6,736 participants in the MESA cohort to 539,934 in the TSGH cohort. The mean age of participants ranged from 51 years to 62.2 years, with female representation varying between 49.6% and 56.6%. Follow-up durations varied from 3.1 months to 6.8 years (Table 1).

All studies employed deep learning as the core model architecture. Two studies (12, 21) utilized the same ARIC cohort (n = 14,613) but reported distinct analyses. The outcome definitions varied; most studies relied on clinical HF diagnoses using ICD codes, while the YNHHS cohort (13) incorporated echocardiographic criteria, specifically left ventricular ejection fraction (LVEF) less than 50%. Exclusion criteria predominantly focused on pre-existing HF, missing ECGs, or incomplete follow-up data. Notably, none of the included studies evaluated clinical utility or cost-effectiveness, and all lacked external validation in geographically diverse settings. The largest cohort (14) derived data from electronic health records, which may introduce unmeasured confounders, whereas smaller cohorts, such as MESA, emphasized community-based populations with detailed phenotyping.

Although all included studies employed deep learning architectures, there was considerable diversity in model design, input representation, and validation strategies, which may influence performance and generalizability (Table 2). Input formats varied between raw signal time-series and image-based ECG representations, which may affect feature extraction robustness. Preprocessing also differed, though most studies applied noise filtering and signal normalization. Only two studies (13, 21) included external validation cohorts, highlighting a general limitation in geographic and clinical generalizability.

Most AI-ECG models (80%) demonstrate a low overall risk of bias. The unclear risk primarily stems from ambiguous reporting of result intervals and a lack of procedures for handling missing data (Figure 2; Supplementary Table 4).

Significant heterogeneity was observed across studies (I² = 89%, Cochran's Q = 98.7, p < 0.001). Funnel plots were symmetrical but exhibited additional horizontal scatter (Figures 3), which is consistent with the presence of between-study heterogeneity. This heterogeneity was driven by three primary factors: (1) variability in outcome definitions (e.g., clinical HF vs. HF with LVEF < 50%); (2) diversity in data sources (hospital-based vs. community cohorts with differing baseline risk profiles); and (3) methodological inconsistencies in ECG preprocessing and model hyperparameter optimization. To assess the robustness of the pooled AUROC [0.76 (0.74–0.78)], we conducted a leave-one-out sensitivity analysis. After excluding any single study, the pooled estimates showed minimal changes (range: 0.75–0.77). When the Dhingra (13) (ELSA-Brasil) cohort with the highest AUROC was excluded, the pooled estimate slightly decreased to 0.75 [0.73–0.77]. Upon excluding the UKB cohort with the lowest precision, the confidence interval narrowed to 0.76 [0.74–0.78]. The sensitivity analysis confirmed the robustness of the pooled results.

The definition of HF, the primary outcome, varied across the included studies. Most studies defined HF based on clinical diagnoses recorded in hospital administrative databases using ICD codes. In contrast, one study (13), YNHHS cohort) applied a more stringent, imaging-based criterion, defining HF as a LVEF of less than 50%. This methodological distinction between broadly defined, clinically coded HF and a more specific, phenotypically defined HF was identified a priori as a major potential source of clinical heterogeneity.

The pooled AUROC across all cohorts was 0.76 (95% CI: 0.74–0.78; p < 0.001), indicating moderate-to-good discriminatory ability for HF prediction (Figure 4). Subgroup analyses revealed nuanced variations: among ethnic subgroups, the highest AUROC was observed in Asian populations [0.78 (0.77–0.79), (20)], while Hispanic and non-Hispanic White cohorts showed comparable performance [0.79 (0.79–0.80) and 0.77 (0.77–0.78), respectively], suggesting minimal ethnic disparity. Undifferentiated cohorts (e.g., ARIC and MESA) exhibited consistent AUROCs (0.756–0.77) despite differences in sample size and HF prevalence (3.5%–5.5%). Geographic variability highlighted challenges in low-incidence settings [e.g., UKB: AUROC 0.769 (0.670–0.867), HF incidence 0.1%], whereas smaller cohorts with limited events (e.g., ELSA-Brasil: n = 31) achieved higher point estimates [AUROC 0.810 (0.714–0.907)] but with wider confidence intervals, underscoring the impact of sample size and event frequency on precision.

The head-to-head comparison between AI-ECG and traditional FHS-HF/PCP-HF risk scores across 5 cohorts from 3 studies demonstrated a pooled AUROC of 0.757 (95% CI: 0.730–0.782) for AI-ECG models vs. 0.742 (95% CI: 0.692–0.787) for traditional models, with overlapping confidence intervals and a non-significant test for subgroup differences (P = 0.575) indicating no consistent statistical superiority for the AI-ECG approach (Figure 5). Despite this overall equivalence, population-specific advantages emerged, as the AI-ECG model showed significantly superior performance in the large Asian TSGH cohort (AUROC 0.75 vs. 0.61, P < 0.001) and numerically higher discrimination in the MESA cohort, whereas performance remained comparable in the ARIC and other SUMC sub-cohorts. The substantial 95% prediction intervals for both AI-ECG (0.649–0.838) and traditional models (0.612–0.838) reflect considerable uncertainty in their relative performance in new settings. From a clinical perspective, the observed absolute AUROC difference of approximately 0.015, while not statistically significant at the overall level, may still hold relevance in population-wide screening contexts where even modest improvements in discrimination can meaningfully reclassify risk categories for a substantial number of individuals, potentially enabling more targeted preventive interventions.

Subgroup analysis based on heart failure (HF) definition revealed consistent performance of AI-ECG models across different diagnostic criteria. The diagnostic code-defined group (7 cohorts, n = 951,654) demonstrated a pooled AUROC of 0.765 (95% CI: 0.743–0.786), indicating robust discriminative ability (Figure 6). Substantial heterogeneity was observed (I² = 89%), attributable to variations in population characteristics, study design, and healthcare settings. The imaging-defined group (LVEF < 50%, 1 cohort, n = 231,285) showed comparable performance with an AUROC of 0.810 (0.800–0.820). The complete overlap between the prediction interval of the diagnostic code group (0.692–0.826) and the point estimate of the imaging group suggests no statistically significant difference in model performance between definition types, though this interpretation is limited by the single study in the imaging subgroup.

These findings indicate that AI-ECG models maintain predictive utility regardless of HF definition methodology. The diagnostic code-based approach offers practical advantages for large-scale screening applications, while imaging-based definitions may provide more objective endpoint adjudication. The observed heterogeneity underscores the need for standardized outcome definitions and validation protocols in future multicenter studies. Further research should prioritize inclusion of additional imaging-defined cohorts to strengthen comparative analyses and evaluate definition-specific impacts on model calibration and clinical utility.

This systematic review and meta-analysis demonstrates that AI-ECG models possess moderate to good discriminatory ability for predicting heart failure (HF), with a pooled AUROC of 0.76 (95% CI: 0.74–0.78). This performance is comparable to established clinical risk scores such as the FHS-HF model (AUROC 0.742), suggesting that AI-ECG holds promise as a non-invasive and cost-effective tool. Its operational simplicity—leveraging ubiquitously available ECG data without the need for invasive blood tests or complex clinical parameters—positions it as an ideal candidate for rapid point-of-care screening and global cardiovascular risk reduction strategies, particularly in resource-limited areas.

Critically, the potential of AI-ECG is underscored by its consistent performance across ethnic groups in our subgroup analyses (AUROC range: 0.77–0.79). Notably, while the technology did not demonstrate universal superiority over traditional models, it showed enhanced performance in specific cohorts, such as the large Asian TSGH population. This variability suggests that AI-ECG may offer the greatest incremental value in populations with specific baseline risk profiles or where traditional risk factors are not fully captured. Therefore, AI-ECG should be viewed not as a blanket replacement for established scores, but as a complementary tool with particular utility in certain contexts.

Significant heterogeneity was observed across the included studies (I² = 89%), which can be attributed to several methodological factors. First, inconsistent definitions of heart failure outcomes represent a principal source of variability. While the majority of studies relied on ICD codes extracted from electronic health records, others adopted imaging-confirmed criteria such as left ventricular ejection fraction (LVEF). This fundamental discrepancy in endpoint ascertainment not only limits the direct comparability of model performance but also introduces potential misclassification bias, particularly concerning heart failure with preserved ejection fraction (HFpEF).

Additionally, the diversity in ECG data sources and processing methodologies further contributes to the observed heterogeneity. The performance and generalizability of AI-based ECG models are intrinsically tied to the quality and consistency of the input data. In real-world settings, ECG signals are acquired from highly heterogeneous sources, including native digital recordings from standard 12-lead machines, digitized scans of paper printouts, or abbreviated waveforms from wearable monitoring devices. Consequently, disparities in acquisition hardware, sampling frequency, filter settings, and data format may introduce technical confounders that substantially influence model output.

Finally, cohort-specific performance variations underscore this inconsistency. For instance, larger cohorts derived from real-world electronic health record data demonstrated marginally lower AUROC values, potentially attributable to unmeasured confounders or data noise. Notably, the performance in specific subgroups, such as young Black women in one study (20), showed a significant decline. Collectively, these factors highlight the persistent challenges in methodological standardization and the necessity for unified protocols in data acquisition, model validation, and outcome reporting to improve reproducibility.

Our analysis has several key limitations that chart a course for future work, including the absence of clinical utility assessments, the inherent “black-box” nature of deep learning models (22), and potential publication bias. To address these gaps, an urgent priority is the design and execution of large-scale, prospective, multi-center trials to determine whether AI-ECG-guided risk stratification actually reduces heart failure incidence or improves patient outcomes compared to standard care—a prerequisite for widespread adoption. Beyond establishing efficacy, future efforts must extend into implementation science, focusing on the seamless integration of these algorithms into clinical workflows. This entails developing interoperable systems that embed AI-ECG analysis within electronic health records and provide intuitive decision support to clinicians at the point of care. Furthermore, the sustainable implementation of AI-ECG hinges on demonstrating its long-term value and equity. This mandates rigorous health economic analyses, particularly in low-resource settings where its operational advantages could be most impactful, coupled with proactive measures to ensure fairness and transparency. Independent third-party validation and periodic auditing are indispensable to mitigate performance disparities across ethnic, and socioeconomic groups and to prevent the exacerbation of existing health inequities. Adopting privacy-preserving frameworks, such as federated learning, could further facilitate continuous model refinement across diverse institutions.

In conclusion, AI-ECG models demonstrate moderate-to-good discriminatory ability for predicting heart failure, with performance comparable to traditional risk models. The technology shows particular promise as a non-invasive, cost-effective screening tool, especially in environments where traditional risk factor collection is challenging. Future research should prioritize robust prospective validation, seamless clinical integration, and a steadfast commitment to equitable deployment. Ultimately, the goal is to identify the specific patient populations and clinical scenarios where AI-ECG provides the greatest incremental value, thereby enabling its transition from a research tool to a clinically actionable solution for personalized HF prevention.