Evaluation of an Echocardiography-Based Artificial Intelligence Algorithm for Early HFpEF Detection: A HeartShare/AMP-HF Cohort Study

Early heart failure with preserved ejection fraction (HFpEF) is difficult to identify because of nonspecific symptoms and subtle echocardiographic abnormalities. EchoGo HF, an FDA-cleared artificial intelligence (AI) model, estimates HFpEF probability from a single apical 4-chamber echocardiographic clip, but has been validated only in patients with clinically adjudicated HFpEF.

This study aims to evaluate whether AI-HFpEF can detect early HFpEF in community-dwelling adults without a clinical heart failure (HF) diagnosis and predict incident clinical HFpEF in patients within a health system.

The authors included 2 community cohorts (the MESA [Multi-Ethnic Study of Atherosclerosis] Early HF study and the DHMS [Dallas Hearts and Minds Study]). Early HFpEF was defined by impaired exercise capacity or health-related quality of life with objective evidence of elevated cardiac filling pressures in patients without clinically diagnosed HF. The AI-HFpEF model was applied to resting echocardiograms, and the diagnostic performance was assessed. Multivariable models examined associations between AI-HFpEF output and early HFpEF phenotype. Separately, we included adults in the UTSW (University of Texas Southwestern) Health System with structurally normal echocardiograms and no prior HF (2018-2024) and assessed the association of AI-HFpEF output with incident HFpEF (adjudicated using a validated registry) using Cox regression.

In MESA (N = 2,544; 8.3% early HFpEF), AI-HFpEF outperformed H₂FPEF for detecting early HFpEF (area under the receiver-operating characteristic curve [AUROC]: 0.77 vs 0.70). AI-HFpEF diagnostic positivity was associated with higher odds of early HFpEF (OR: 3.26; P < 0.001), lower 6-minute walk distance, and lower Kansas City Cardiomyopathy Questionnaire-12-Clinical Summary Score. Similarly, in DHMS (N = 624; 18.6% early HFpEF), AI-HFpEF outperformed H₂FPEF (AUROC: 0.83 vs 0.77), and diagnostic positivity was associated with early HFpEF (OR: 6.76; P < 0.001), shorter exercise duration, and lower peak oxygen uptake. In the UTSW cohort (N = 6,876; median follow-up = 33 months; incident HFpEF = 3.4%), an AI-HFpEF diagnostic positive result was associated with >2-fold higher incident HFpEF risk (aHR: 2.08; P < 0.001). This association persisted across all conventional diastolic dysfunction grades.

AI-HFpEF identified the early HFpEF phenotype in community-dwelling adults and independently predicted incident clinical HFpEF among patients in a health system.