Exploring heterogeneity in the diagnostic accuracy of artificial-intelligence-enhanced ECG (AI ECG) for predicting left ventricular dysfunction: a meta-analysis

Abstract Objective This meta-analysis aimed to evaluate the diagnostic performance of AI-based electrocardiography (ECG) in predicting left ventricular (LV) dysfunction, with a specific focus on exploring sources of heterogeneity across studies. Methods After conducting a comprehensive literature search across multiple databases and performing a systematic study evaluation, a total of nine studies were included in the meta-analysis. The total sample comprised 145,855 individuals, with 7,957 diagnosed with LV dysfunction and 137,898 classified as non-diseased, yielding an overall prevalence of 5%. The primary outcomes assessed were pooled sensitivity, specificity, diagnostic odds ratio (DOR), and likelihood ratios (LR+ and LR-). To evaluate heterogeneity, we applied multiple metrics, including variance in logit sensitivity and specificity, median odds ratios (MOR), the bivariate I² statistic, and the area of the 95% prediction ellipse. Results The pooled sensitivity of AI ECG for detecting LV dysfunction was 0.835 (95% CI: 0.764–0.888), and the pooled specificity was 0.894 (95% CI: 0.813–0.942). The DOR was 42.794 (95% CI: 23.791–76.976), indicating strong discriminatory ability. The likelihood ratios (LR+ = 7.883, 95% CI: 4.512–13.774; LR- = 0.184, 95% CI: 0.131–0.260) confirm that AI ECG is effective both for confirming and excluding LV dysfunction. However, significant heterogeneity was detected. The bivariate I² statistic was 0.952, indicating high variability across studies. Variances in logit sensitivity (0.418) and logit specificity (1.016) reflected moderate variation, and the MOR for sensitivity (1.852) and specificity (2.615) suggested notable differences in diagnostic performance between studies. The Summary Receiver Operating Characteristic (SROC) curve (Figure 1) illustrates the overall diagnostic accuracy of AI ECG. The area of the 95% prediction ellipse was 0.242, indicating a degree of consistency in the sensitivity and specificity estimates, despite heterogeneity. Conclusion AI ECG demonstrates strong overall diagnostic accuracy for detecting LV dysfunction, with high sensitivity and specificity. However, significant heterogeneity (bivariate I² = 0.952) indicates that differences in study populations, AI model characteristics, and clinical settings may influence diagnostic performance. Further research is required to standardize AI ECG applications and enhance reproducibility across diverse patient cohorts. Figure 1: The SROC curve illustrates the diagnostic performance of AI-enhanced ECG for detecting LV dysfunction. The solid curve represents the trade-off between sensitivity and specificity, with each dot corresponding to an individual study. The summary point shows the pooled sensitivity and specificity. The confidence ellipse indicates the uncertainty surrounding the pooled estimate, while the prediction ellipse reflects the expected variability across future studies.