Predictive Modeling of Heart Failure Readmissions

PurposeFederal programs to mitigate hospital readmission of patients with heart failure (HF) monetarily encourage hospitals through the use of penalties. The limited performance of predictive models have created potential challenges of implementation and unintended consequences, with criticisms about its unintended consequences and the low performance of its predictive models. We study sought to refine existing predictive models of readmission using heart failure (HF) data from a large multi-payer national dataset. MethodsThe Premier healthcare database, a nationally representative all-payor dataset, was utilized to examine over 300 variables from HF patients (2016-2023) including demographics, comorbidities, cardiac diagnoses, provider characteristics, medications, and lab values, defined using diagnosis-related group and ICD-10 codes. Outcomes from patients with primary and secondary HF diagnoses included 30-day all-cause readmissions and 30-day HF-related readmissions. Data were divided into training (60%), validation (20%), and testing (20%) sets. We evaluated logistic regression, random forest, neural networks, modified neural networks, support vector machines, naive Bayesian decision trees, and XGBoost models, comparing them based on accuracy (AUC), precision, recall, and F-score. ResultsOf 722,974 HF patients examined, 12.0% and 11.3% experienced all-cause and HF-related 30-day readmissions, respectively. Mean age was 71 years and 48% were female. A total of 68,649 patients readmitted with a primary HF diagnosis for homogeneity (2021-2023) was thoroughly analyzed using multiple contemporary Bayesian and non-Bayesian models. This subset was 47% female with a mean age of 72 years. The XGBoost model performed best, with an AUC of 0.63 for all-cause and 0.62 for HF-related readmissions. The key predictors of readmissions were age and chronic non-cardiac comorbidities instead of HF-specific factors. ConclusionContemporary statistical models applied to nationally representative contemporary real-world data struggle to identify modifiable interventions, suggesting that existing federal programs may penalize without actionable improvements in patient care.