This systematic review and meta-analysis followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [16].

We included studies that (1) applied an AI or ML algorithm; (2) used ECG data (single-lead or 12-lead) as the input; (3) focused on detecting or predicting cardiac dysfunction related to impaired diastolic function, including HFpEF, LVDD, or objectively confirmed elevated filling pressures; and (4) reported sufficient data to construct a 2 × 2 contingency table. We found a total of nine studies that were available for inclusion in this analysis.

PubMed, Cochrane, CTG, and Google Scholar databases were systematically searched to identify original literature that evaluated the diagnostic accuracy of AI algorithms using ECG data for people with HFpEF. The full search strategies for all databases are provided in Supplementary materials.

Two review authors independently performed the literature search and title/abstract screening against predefined criteria, with full texts reviewed for eligibility, and the articles were excluded that were not associated with the research topic. Nine studies met the inclusion criteria and were incorporated into both qualitative and quantitative synthesis. Results of the literature search are shown in Figure 1.

For each study, we extracted the first author, publication year, study design, population characteristics, AI/ML model type, diagnostic endpoint, and reported performance metrics. For the meta-analysis, we obtained or derived true positives, false positives, true negatives, and false negatives from data reported in the original publications, including confusion matrices, diagnostic accuracy tables, or reported sensitivity, specificity, and sample size. No AI models were trained or revalidated by the authors; all analyses were based exclusively on published study results.

Quality of studies was independently assessed using the QUADAS-2 tool [17], which evaluates four domains: Patient Selection, Index Test, Reference Standard, and Flow/Timing. Each was classified as low, high, or some risk of bias. Two reviewers independently performed the data extraction and quality assessment. Disagreements were resolved through discussion and independent assessment by another researcher to reach a consensus. A total of two disagreements (out of 2 assessments) were identified, all of which were resolved through consensus discussion.

Our primary outcomes were the pooled sensitivity and specificity of AI-ECG models. Sensitivity was calculated as TP/(TP + FN), representing the proportion of true disease cases correctly identified. Specificity was calculated as TN/(TN + FP), representing the proportion of non-diseased cases correctly classified. We used bivariate random-effects for meta-analysis, allowing joint estimation of sensitivity and specificity while accounting for their correlation. This approach produced a summary operating point and a hierarchical summary receiver operating characteristic (SROC) curve with 95% confidence intervals (CIs) and prediction intervals.

Heterogeneity was quantified using the I² statistic, with values above 75% considered substantial. To explore potential sources of heterogeneity, we conducted subgroup analyses by (1) clinical endpoint (HFpEF, LVDD, or increased filling pressure) and (2) AI model type (DL vs. classical ML). All analyses were performed in R.

HFpEF, LVDD, and elevated filling pressures were analyzed together as diagnostic endpoints reflecting diastolic dysfunction, given their shared pathophysiological basis and overlapping diagnostic criteria.

The study selection process is depicted in the PRISMA flow diagram (Figure 1). Nine studies were ultimately included. They varied a lot in their designs, study populations, and methodological approaches (Table 1). A wide range of AI models were evaluated, including DL architectures such as convolutional neural networks (CNNs) and classical algorithms like Random Forest and XGBoost. ECG inputs ranged from standard 12-lead recordings to new single-lead data from smartphones and wearable patches. Clinical endpoints included the spectrum of cardiac dysfunction, including HFpEF, LVDD, and invasively confirmed elevated filling pressures.

The QUADAS-2 evaluation is summarized in Figure 2. All nine included studies were judged to have a low risk of bias in the reference standard domain, as each employed appropriate diagnostic comparators such as guideline-based echocardiography or invasive hemodynamic assessment. In contrast, patient selection was the most frequent source of concern. Several studies relied on retrospective designs, thus limiting the representativeness of their findings. Additional concerns were noted in the flow and timing domain when subsets of participants were excluded from final analyses, as a result, introducing selection bias.

Forest plots of individual study estimates for sensitivity and specificity are shown in Figures 3 and 4. Reported sensitivities varied widely, ranging from 0.67 to 0.99 while specificities ranged from 0.60 to 0.98.

The bivariate random-effects meta-analysis produced a pooled sensitivity of 0.82 (95% CI: 0.70–0.90) and a pooled specificity of 0.83 (95% CI: 0.74–0.89). Both outcomes demonstrated extremely high and statistically significant heterogeneity (I² > 96%, p < 0.0001).

To synthesize these findings and account for the correlation between sensitivity and specificity, a bivariate random-effects meta-analysis was performed. The resulting SROC curve is presented in Figure 5. The summary operating point for the AI-ECG models yielded a pooled sensitivity of 0.82 (95% CI: 0.70–0.90) and a pooled specificity of 0.83 (95% CI: 0.74–0.89). The 95% confidence region around this summary point indicates the precision of the mean estimate, while the much wider 95% prediction region illustrates the substantial heterogeneity between studies and the likely range of performance for a future study.

To investigate the cause of this heterogeneity, we conducted pre-specified subgroup analyses (Figures 6 and 7). We examined study-level specificity estimates according to AI model type (classical ML vs. DL). While individual studies showed variation in specificity, statistical comparison did not show a significant difference between model types (p = 0.60). High heterogeneity persisted in both groups. Likewise, subgrouping by clinical endpoint—diastolic dysfunction, HFpEF, or increased filling pressure—did not reveal significant between-group differences in specificity (p = 0.52). Importantly, heterogeneity remained very high across all subgroup comparisons (I² > 77%), indicating that neither model type nor target condition accounted for the variability. Subgroup analyses for sensitivity yielded similar findings.

Based on the available evidence, AI-enhanced ECG models appear most suitable as screening or triage tools rather than standalone diagnostic tests. Their high specificity and negative predictive value suggest potential utility in ruling out cardiac dysfunction in low- to intermediate-risk populations, thereby optimizing referral for echocardiography and improving resource allocation within clinical workflows.

This meta-analysis has several strengths: a comprehensive synthesis of an emerging literature, a formal evaluation of study quality using QUADAS-2, and the application of a robust bivariate meta-analytic model. Importantly, our exploration of heterogeneity, though inconclusive, highlights the methodological complexity underpinning this research field.

However, the findings must be interpreted with caution. The unexplained heterogeneity weakens the precision of the pooled estimates, which may represent little more than an average across very different contexts. The limited number of studies also reduced the power of subgroup comparisons. Also, the included studies were largely retrospective, and our dataset was based on a defined pool of studies rather than an exhaustive search, further limiting generalizability.

AI-enhanced ECG systems show high pooled specificity for detecting cardiac dysfunction and thus represent a promising, cost-effective, and accessible screening option. However, the field is marked by profound heterogeneity that cannot be explained by endpoint choice or model type alone, and model performance is clearly context-dependent. Large-scale, prospective studies across diverse populations, along with better methodological standardization, are important for developing robust, generalizable AI-ECG tools that can be confidently integrated into clinical practice.