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

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning in Medicine

Background:

• Limited studies integrate echocardiography and laboratory data for predicting clinical outcomes in heart failure with preserved ejection fraction (HFpEF).
• Predicting adverse events in HFpEF remains challenging due to the heterogeneity of the condition.

Purpose of the Study:

• To employ machine learning techniques to identify significant predictors of heart failure (HF) hospitalization and cardiovascular (CV) death in HFpEF patients.
• To develop a robust predictive model for clinical outcomes in HFpEF using a large patient cohort.

Main Methods:

• Utilized the Chang Gung Research Database in Taiwan, analyzing 6,092 HFpEF patients (2,898 derivation, 3,194 validation) from 2008-2017, with follow-up until 2019.
• Developed a random survival forest model incorporating 58 variables to predict the composite outcome of HF hospitalization and CV death.
• Assessed model generalizability using an independent validation cohort.

Main Results:

• A total of 15 predictive indicators were identified, including age, B-type natriuretic peptide, left atrium size, atrial fibrillation, prior HF hospitalization frequency, BMI, mitral regurgitation severity, left ventricular wall thickness, dysnatremia, LV end-diastolic dimension, uric acid, triglycerides, blood urea nitrogen, interventricular septum thickness, and HbA1c.
• The random survival forest model demonstrated strong predictive performance with an 86.9% area under the curve in the validation cohort.
• Significant proportions of both derivation (37.7%) and validation (36.0%) cohorts experienced the composite outcome during a 2.9-year follow-up.

Conclusions:

• Machine learning effectively identified 15 crucial predictors for HF hospitalization and CV death in HFpEF.
• These findings facilitate the identification of high-risk HFpEF patients, enabling tailored and personalized treatment approaches.
• The developed model shows robust external validity, supporting its clinical utility.