DOI: 10.1161/circ.148.suppl_1.18252 ISSN: 0009-7322

Abstract 18252: Comparison of an Artificial Intelligence Heart Failure Detection Model and Clinical Prediction Models in Patients With Heart Failure With Preserved Ejection Fraction (HFPEF)

Ashley P Akerman, Nora Al-Roub, Constance Angell-James, Rasheed Thompson, Lorenzo Bosque, Will Hawkes, Hania Piotrowska, Paul Leeson, gary woodward, Ross Upton, Jordan B Strom
  • Physiology (medical)
  • Cardiology and Cardiovascular Medicine

Background: An artificial intelligence (AI) model to diagnose heart failure (HF) with preserved ejection fraction (HFpEF) based on deep learning of echocardiographic (TTE) images could improve under-recognition of disease. However, the diagnostic accuracy in high-risk patients, compared to existing clinical prediction models, remains uncertain.

Methods: In this retrospective case-control study, we identified patients who received routine TTE at Beth Israel Deaconess Medical Center, Harvard Medical School. Cases were defined by presence of ICD-10 codes for HF (I50.X), LVEF≥50%, and grade II or III diastolic dysfunction. Controls were defined by absence of HF within 1 year of the index TTE. The discriminatory performance of an AI HFpEF model (EchoGo Heart Failure, Ultromics Ltd.) was compared to the clinical H2FPEF score.

Results: Both cases (n=213, 75±12 years, 53% female, 29.4±7.1 kg/m2) and controls (n=226, 75±13 years, 54% female, 27.0±5.6 kg/m2) were high-risk for HFpEF, with high rates of hypertension, atrial fibrillation, diabetes, coronary artery, kidney, and pulmonary disease. The H2FPEF score demonstrated high sensitivity (mean [95% CI]; 97.3% [94.3, 100%]) and modest specificity (38.1% [25.9, 55.9%]). The AI HFpEF model had similar sensitivity (88.8% [84.0, 93.4%]) and slightly improved specificity (55.6% [47.7, 64.0%]). In the 265 patients (60.4%) with non-diagnostic H2FPEF scores, the AI HFpEF model successfully classified 144 patients as probable HFpEF ( Figure ). The H2FPEF score identified 31 (54.5%) whereas the AI HFpEF model identified 45 (78.9%) HFpEF patients who were hospitalized within 1 year of the TTE.

Conclusions: An AI model to detect HFpEF improves upon identification of HFpEF amongst high-risk individuals, thus potentially improving the capacity to identify patients who might benefit from treatment.

More from our Archive