Artificial intelligence-enabled electrocardiography for estimating left ventricular filling pressure

Introduction Elevated left ventricular filling pressure (LVFP) is a hallmark of heart failure with preserved ejection fraction (HFpEF), yet its non-invasive assessment remains challenging. AI-enabled ECG (AI-ECG) has shown potential for detecting cardiac dysfunction, but its utility for identifying elevated LVFP has not been fully established. Purpose To evaluate the diagnostic performance of AI-ECG for detecting elevated LVFP compared with established clinical scoring systems, H2FPEF and ABA scores. Methods We conducted a retrospective study using 12-lead ECG–echocardiography pairs obtained within 14 days from two tertiary centres. Elevated LVFP was defined according to the 2025 American Society of Echocardiography (ASE) algorithm for estimation of mean left atrial pressure. We fine-tuned an ECG foundation model (pretrained on >10 million ECGs) using a multi-task loss that combined elevated LVFP classification with regression of the five echocardiographic parameters underlying the 2025 ASE algorithm. The model was evaluated in internal and external test cohorts and compared against H2FPEF and ABA scores. Model discrimination, calibration, and clinical utility and classification metrics were assessed, and the incremental value of AI-ECG over the comparator scores was also evaluated. Statistical testing was performed for comparative performance and incremental value analyses. Results The internal cohort included 183,027 patients (266,342 ECG–echocardiography pairs; elevated LVFP prevalence 17.7%). AI-ECG outperformed both ABA and H2FPEF in the internal cohort (AUROC 0.870 vs 0.641 and 0.729, respectively; both p < 0.001) and maintained similar superiority in the external cohort (72,308 patients; 96,132 pairs; AUROC 0.866 vs 0.697 and 0.729; both p < 0.001). AI-ECG also showed better calibration with lower Brier scores than both comparator scores in the internal (0.099 vs 0.219 and 0.136) and external cohorts (0.091 vs 0.175 and 0.127), and showed greater net benefit on decision curve analysis. At a probability cutoff of 0.5, AI-ECG achieved high specificity and PPV in both cohorts, while incremental value analyses consistently favored AI-ECG over ABA and H2FPEF (Figure 1,2). Conclusion AI-ECG demonstrates superior performance for detecting elevated LVFP at rest compared with established scoring systems, with robust generalizability across internal and external cohorts. These findings support the potential role of AI-ECG as a scalable non-invasive tool for identifying elevated LVFP.