Artificial Intelligence-Enhanced Electrocardiography for the Diagnosis of Heart Failure With Preserved Ejection Fraction: A Systematic Review and Meta-Analysis.

Abstract

Background: Accounting for approximately 50% of heart failure, heart failure with preserved ejection fraction (HFpEF) is becoming increasingly common as populations age and multimorbidity grows. Diagnosis remains challenging, requiring multimodal testing with echocardiography, biomarkers and sometimes invasive haemodynamics. Artificial intelligence applied to the electrocardiogram (AI-ECG) offers a low-cost, scalable means of detecting HFpEF by extracting patterns beyond human interpretation.

Methods: We conducted a systematic review and meta-analysis in accordance with PRISMA, registered prospectively with PROSPERO. PubMed, Embase, Web of Science, and IEEE Xplore were searched to 1 August 2025 for studies evaluating AI/ML models applied to ECGs for the diagnosis of HFpEF or left ventricular diastolic dysfunction (LVDD). Eligible studies reported diagnostic performance compared with a recognized reference standard. Risk of bias was assessed with QUADAS-AI. AUROC values were pooled using a logit transformation and random-effects model, with results back-transformed for interpretability.

Results: Ten studies (2021-2025) met inclusion criteria, encompassing > 270,000 participants across diverse populations. Seven studies provided sufficient data for pooling, contributing 11 independent cohorts. The pooled AUROC was 0.84 (95% CI 0.78-0.88), indicating good discriminatory ability, though heterogeneity was extreme (I² = 100%). Three additional studies reported diagnostic metrics without AUROC variance and were synthesized narratively. Risk of bias was moderate to high in several domains, driven by selective cohorts, inconsistent reference standards, and incomplete reporting.

Conclusions: AI-ECG shows promise for the detection of HFpEF, but the current evidence base is predominantly retrospective, methodologically heterogeneous, and limited by variable reference standards and insufficient external validation. No prospective, outcome-based studies have yet established its clinical utility, and real-world implementation remains untested.