Abstract 14400: Predicting In-Hospital Mortality Among Patients Admitted With Heart Failure: A Machine Learning Approach

Abstract

Introduction: Patients hospitalized with heart failure (HF) experience high in-hospital mortality. Machine learning predictive models have the potential to identify heart failure patients at risk of adverse outcomes. Methods: Using the electronic medical record (EMR) at a large academic medical center, we identified 49,457 patients who were admitted with heart failure from 2011 to 2022. 71% were admitted with HF with reduced ejection fraction (HFrEF, EF≤40%) and 29% with HF with mildly reduced or preserved ejection fraction (HFmrEF/HFpEF, EF >40%). Using patient demographics, comorbidities, admission medications, labs, and vitals throughout the hospitalization, we constructed random forest machine learning models to predict in-hospital mortality. We compared model performance to that of logistic regression utilizing the same variables. Results: The median age (25th-75th percentile) of the cohort was 72 (59-83) years, and 45% of the group was female. 71% patients were admitted with HFrEF and 29% patients were admitted with HFmrEF/HFpEF. For the overall cohort, our machine learning model was able to predict in-hospital mortality with a sensitivity (SN) of 91%, a specificity (SP) of 85%, and an area under the curve (AUC) of 0.95 (Figure 1). For those with HFrEF, our model demonstrated a SN of 91%, SP of 87% and AUC of 0.95, and for those with HFmrEF/HFpEF, a SN of 87%, a SP of 89%, and an AUC of 0.94. For all three cohorts, our models outperformed logistic regression. The top predictors of in-hospital mortality were BNP, BUN, and Systolic BP for those admitted with HFrEF, and sex, creatinine, and diastolic BP for those with HFmrEF/HFpEF. Conclusions: Machine learning models built from EMR data have the ability to accurately predict in-hospital mortality among patients admitted for HF. Such prediction tools, when used in the appropriate context, may help flag hospitalized heart failure patients at high risk for in-hospital mortality.