Predicting Mortality and Hospitalization in Heart Failure With Preserved Ejection Fraction by Using Machine Learning.

Background: Few studies have incorporated echocardiography and laboratory data to predict clinical outcomes in heart failure with preserved ejection fraction (HFpEF).

Objectives: This study aimed to use machine learning to find predictors of heart failure (HF) hospitalization and cardiovascular (CV) death in HFpEF.

Methods: From the Chang Gung Research Database in Taiwan, 6,092 HFpEF patients (2,898 derivation, 3,194 validation) identified between 2008 and 2017 were followed until 2019. A random survival forest model, using 58 variables, was developed to predict the composite outcome of HF hospitalization and CV death.

Results: During 2.9-year follow-up, 37.7% of derivation and 36.0% of validation cohort patients experienced HF hospitalization or CV death. The study identified 15 predictive indicators, including age ≥65 years, B-type natriuretic peptide level ≥600 pg/mL, left atrium size ≥46 mm, atrial fibrillation, frequency of HF hospitalization within 3 years, body mass index <30 kg/m, moderate or severe mitral regurgitation, left ventricular (LV) posterior wall thickness of <10 or ≥13 mm, dysnatremia, LV end-diastolic dimension of <40 or ≥56 mm, uric acid level ≥7 mg/dL, triglyceride level of <70 or ≥200 mg/dL, blood urea nitrogen level ≥20 mg/dL, interventricular septum thickness of <11 or ≥20 mm, and glycated hemoglobin (HbA) level of <6% or ≥8%. The random survival forest model demonstrated robust external generalizability with an 86.9% area under curve in validation.

Conclusions: Machine learning identified 15 predictors of HF hospitalization and CV death in HFpEF patients, helping doctors identify high-risk individuals for tailored treatment.