AI-Enabled ECG and echocardiography for pulmonary hypertension: a meta-analysis

We assessed the diagnostic performance of AI models for pulmonary hypertension, including evidence from external validation studies. Successful clinical translation may ultimately support earlier diagnosis. We searched PubMed, Embase, and Web of Science from inception to January 15, 2026 for studies evaluating artificial intelligence models applied to electrocardiography (ECG) or echocardiography for pulmonary hypertension detection. Echocardiography studies using right heart catheterization (RHC) to ascertain PH formed the primary evidence base; control ascertainment was not uniformly catheterization-based. ECG studies used non-RHC or mixed RHC and echocardiographic reference standards, so their pooled analysis was exploratory. Study-level leave-one-out analyses omitted all estimates from each study together. Risk of bias was assessed using PROBAST + AI, and evidence certainty was graded using GRADE. Pooled sensitivity, specificity, and area under the curve were calculated using bivariate random-effects models. Fifteen studies contributed 506,264 participants, summed after within-study deduplication. In the primary analysis of studies using RHC-based PH ascertainment, echocardiography-based artificial intelligence yielded a pooled sensitivity of 0.92 (95% CI: 0.86–0.95), specificity of 0.79 (95% CI: 0.64–0.89), and area under the curve of 0.93 (95% CI: 0.83–0.97). Because two ECG studies used mixed RHC and echocardiographic definitions and three used non-RHC definitions, ECG data were pooled across all eligible studies irrespective of reference standard and were explicitly treated as a secondary, exploratory analysis. This analysis yielded a pooled sensitivity of 0.81 (95% CI: 0.79–0.82), specificity of 0.75 (95% CI: 0.68–0.81), and area under the curve of 0.84 (95% CI: 0.81–0.87). Given the absence of a uniformly RHC-verified ECG dataset and the heterogeneity of reference standards, these ECG estimates should not be considered definitive performance benchmarks. The certainty of evidence for the pooled diagnostic estimates ranged from very low to moderate according to GRADE. Only two studies directly compared artificial intelligence with clinician interpretation; these data were summarized descriptively and were not interpreted as demonstrating equivalence. ECG- and echocardiography-based artificial intelligence models showed encouraging diagnostic performance in retrospective datasets, with the most robust clinically relevant evidence derived from analyses with RHC-based PH ascertainment where sufficient studies were available. However, the overall certainty of evidence, assessed with GRADE, was very low to moderate, heterogeneity was substantial, and non-RHC reference standards may introduce misclassification. The available ECG data did not support a uniformly RHC-verified standalone model, and current data do not establish superiority or equivalence to clinician interpretation, nor do they confirm screening effectiveness in real-world settings. Prospective multicentre validation using standardised catheterisation-based definitions, prespecified thresholds, and implementation-focused outcomes is required before clinical translation. CRD42024590680.