The diagnostic accuracy of artificial intelligence enhanced electrocardiography for the detection of cardiac dysfunction

Background: Heart failure (HF) remains a growing global health problem, with nearly half of all cases attributed to HF with preserved ejection fraction (HFpEF) and its precursor, left ventricular diastolic dysfunction (LVDD). Although echocardiography is the diagnostic gold standard, its high cost and limited availability restrict its use for large-scale screening. In contrast, the electrocardiogram (ECG) is inexpensive and widely accessible. Recent advances in artificial intelligence (AI) have created opportunities to leverage ECG data for the early detection of cardiac dysfunction. The objective of this study was to systematically review and meta-analyze the diagnostic performance of AI-based ECG models for detecting cardiac dysfunction. Methods: The QUADAS-2 tool was used to assess the risk of bias. Pooled sensitivity and specificity were estimated using a bivariate random-effects model, with heterogeneity quantified using the I2 statistic. Pre-specified subgroup analyses were conducted according to clinical endpoint and AI model type. Results: Following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, nine eligible studies evaluating AI algorithms applied to ECG data for the detection of HFpEF were identified. Considerable methodological and population heterogeneity was observed across studies. Risk of bias was generally low for reference standards, although concerns were noted in patient selection. The pooled specificity of AI-ECG models was high at 0.83 [95% confidence interval (CI): 0.74–0.89], while pooled sensitivity was 0.82 (95% CI: 0.70–0.90). Both estimates demonstrated extremely high heterogeneity (I2 > 96%). Subgroup analyses by endpoint and model type did not explain this variability. Discussion: AI-enhanced ECG models show good diagnostic accuracy, specifically in ruling out cardiac dysfunction due to their high specificity. However, the high and unexplained heterogeneity across these studies limits the immediate generalizability of the results. Large, prospective validation studies across diverse populations are essential before these models can be confidently adopted into routine clinical practice.