Enhanced Prediction of 30-Day Hospital Readmission in Patients with Congestive Heart Failure Using Physiologic and Laboratory Features in Machine Learning Models: Retrospective Cohort Study (Preprint)

BACKGROUND Models to predict hospital readmission in patients with congestive heart failure (CHF), such as those used by the Centers for Medicare & Medicaid Services (CMS), have limited accuracy. Additional features, available from electronic health records (EHR), have potential to improve accuracy of predictive models. OBJECTIVE The goal of this study was to evaluate the ability of demographic, hospital utilization, medication, laboratory, and physiologic variables to improve prediction of 30-day hospital readmission compared to CMS’s model that utilizes only age, sex, and diagnostic billing codes. METHODS We used the Medical Information Database for Intensive Care (MIMIC-IV) to derive features from patients admitted with a diagnosis of CHF. Machine learning methods were used to develop predictive models and accuracy of models compared using area under the receiver operator characteristic curves (AUROC). RESULTS 6,604 hospital CHF admissions were identified; 30-day hospital readmission occurred for 24.5% of admissions. Models that included features derived from time-series laboratory (AUROC=0.62/95% CI 0.59-0.64) and physiologic (0.63/0.61-0.65) variables performed significantly better than CMS’s base model (0.54/0.51-0.57). These two models included relatively few features related to diagnostic billing codes and a high proportion of laboratory and physiologic features. CONCLUSIONS The addition of features derived from the EHR, particularly those derived from the laboratory and physiologic variable domains, improved predictive accuracy of 30-day hospital readmission over CMS’s base model. Machine learning algorithms that use data readily available from the EHR have potential to improve accuracy of administrative risk adjustment models and reduce bias caused by reliance on diagnostic billing codes.