An FDA-approved AI tool detects heart failure with preserved ejection fraction (HFpEF) from echocardiographic images (apical 4-chamber), without other clinical data. In this external validation (47 patients with unexplained dyspnoea who underwent both echo and right heart catheterisation), the tool was compared with the invasive standard (wedge pressure ≥15 mmHg at rest or ≥25 mmHg with exercise). The AI had high specificity (88%) and positive predictive value (86%) but low sensitivity (30%); using the PCWP/cardiac-output slope, specificity and PPV reached 100%. AI-positive patients had higher filling pressures and pulmonary resistance. Among patients with indeterminate H2FPEF scores, the AI identified 80% of invasively confirmed HFpEF. The tool is thus best for identifying patients with clearly abnormal haemodynamics, not for ruling out HFpEF.
To externally validate an FDA-approved artificial intelligence (AI) tool for detecting heart failure with preserved ejection fraction (HFpEF) using echocardiographic video clips, comparing its performance to invasive haemodynamic criteria in a real-world referral cohort with unexplained dyspnoea.
We retrospectively analysed 47 patients who underwent transthoracic echocardiography (TTE) and right heart catheterization (RHC), including 28 with both rest and exercise haemodynamics. The AI model evaluated apical 4-chamber video data and classified outputs as HFpEF, no HFpEF, or non-diagnostic, without any other clinical data. The primary outcome was invasively defined HFpEF: pulmonary capillary wedge pressure (PCWP) ≥15 mmHg at rest or ≥25 mmHg with exercise. Secondary analysis used a PCWP/cardiac output (CO) slope >2 mmHg/L/min. We assessed sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).
Among patients with exercise RHC (n = 28), 71% met haemodynamic HFpEF criteria. The AI tool demonstrated sensitivity of 30%, specificity of 88%, PPV of 86%, and NPV of 33%. Using the PCWP/CO slope (n = 25), specificity and PPV were 100%, sensitivity 27%, and NPV 16%. AI-positive patients had significantly higher resting PCWP (20 vs 15 mmHg, P = .029), mean PA pressure (29 vs 24 mmHg, P = .02), and PVR (2.1 vs 1.3 WU, P = .002). In patients with indeterminate H2FPEF scores (n = 25), the AI correctly identified 80% of those with invasively confirmed HFpEF. Model performance was consistent across TTE-RHC intervals <365 and <90 days.
This AI model demonstrated high specificity and positive predictive value (PPV) for detecting HFpEF, reliably identifying patients with more advanced haemodynamic abnormalities. Its performance remained robust across variable intervals between transthoracic echocardiography (TTE) and right heart catheterization (RHC), and in patients with indeterminate clinical scores. Due to limited sensitivity, the tool is best utilized to enrich identification of patients with clearly abnormal haemodynamics rather than to exclude HFpEF, particularly in early or borderline cases. While broader use as a screening tool is promising, prospective validation studies are necessary to confirm its utility in general populations.
