Improving Clinical Applicability of Heart Failure Readmission Prediction via Automated Feature Engineering

Heart failure (HF) readmission prediction models often rely on manually curated, cross-sectional features and show limited discrimination and calibration. We evaluated whether automated feature engineering via Deep Feature Synthesis (DFS) improves the clinical applicability of HF readmission prediction from lon-gitudinal electronic health record data. Using 355,217 HF hospitalizations from a large U.S. safety-net health system (2010-2025), we compared a clinician-curated baseline feature set to DFS-enhanced features and trained identical models for 30-, 60-, and 90-day read-mission. DFS consistently improved gradient-boosted tree performance, increasing AUROC and AUPRC across all horizons, while logistic regression performance declined. At sensitivity-targeted operating points (80%), DFS improved specificity and positive predictive value for boosted trees, reducing false-positive workload. Calibration also improved for boosted trees at all horizons but not for linear models. These results show that automated feature engineering yields deployment-relevant gains that are strongly model-class dependent. Data and Code AvailabilityThis study uses retrospective electronic health record data from a large urban safety-net healthcare system in the United States. Due to patient privacy, institutional restrictions, and data use agreements, the data are not publicly available. An anonymized version of the code used for data processing, feature engineering, model training, and evaluation will be made available upon acceptance of the paper. Institutional Review Board (IRB)This retrospective study was reviewed and approved by an institutional review board. Full IRB details will be provided in the camera-ready version of the paper if accepted.