Optimal Approach to Implementing an Echocardiography-based Artificial Intelligence Algorithm for Diagnosing Hemodynamically Confirmed HFpEF in a Referral Population

BACKGROUND: Diagnosing heart failure with preserved ejection fraction (HFpEF) in outpatients remains challenging. An AI model (EchoGo HF) that analyzes resting apical 4-chamber echocardiographic clips shows promise, but its diagnostic performance relative to gold-standard exercise hemodynamic testing is unknown. Additionally, the optimal approach to integrating the AI-HFpEF model with clinical risk scores has not been established. METHODS AND RESULTS: This study included outpatients referred for suspected HFpEF, who underwent exercise hemodynamic testing. HFpEF was defined by elevated filling pressures on exercise right-heart catheterization, pulmonary capillary wedge pressure [PCWP] ≥15 mmHg at rest, ≥25 mmHg during exercise, or PCWP/CO ≥2 mmHg/L/min), or resting/exercise echocardiography (E/e' >14). AI-HFpEF was applied to resting echocardiograms, and diagnostic performance (area under the receiver operating characteristic [AUROC] and F1) was compared with the H₂FPEF score. Associations with disease severity (PCWP, peak exercise oxygen uptake [VO₂peak]) were evaluated using multivariable regression. Additionally, the performance of a combined model (H₂FPEF + AI-HFpEF) was assessed using AUROC. Among 280 patients, 222 (79%) had hemodynamically confirmed HFpEF. AI-HFpEF classified 57% (n = 159/280) as HFpEF, 38% (n = 106/280) as no HFpEF and 5% (15/280) as uncertain. AI-HFpEF-detected HFpEF independently predicted elevated PCWP (OR [95% CI] = 2.48[1.17, 5.22]; P = 0.017), and lower VO₂peak (β = -1.38 [-2.64, -0.13] mL/kg/min; P = 0.031). The AI-HFpEF model had good model discrimination (AUROC = 0.702) with an F1 score of 76.8%, a sensitivity of 67.3% (false-negative rate = 32.7%) and a specificity of 68.5% (false-positive rate = 31.5%). Combining H₂FPEF with AI-HFpEF significantly improved performance to detect hemodynamically confirmed HFpEF vs H₂FPEF score alone (0.724 vs 0.656; DeLong P = 0.021). Incorporating the AI-HFpEF model with H₂FPEF score led to greater improvement in detecting hemodynamically confirmed HFpEF among patients with low/intermediate than those with high H₂FPEF scores (AUROC = 0.746 vs 0.535). A 2-step strategy, combining upfront H₂FPEF score assessment with selective AI-HFpEF assessment in participants with low/intermediate H₂FPEF, identified a larger proportion of patients with hemodynamically confirmed HFpEF (78%) vs H₂FPEF score alone (48%). This strategy resulted in a net reclassification improvement of 15.5% and a relative integrated discrimination improvement of 94%. CONCLUSIONS: AI-HFpEF is independently associated with markers of disease severity, accurately identifies hemodynamically confirmed HFpEF, and improves diagnostic performance when integrated with H₂FPEF, particularly at low/intermediate scores.