Abstract
INTRODUCTION: To develop and internally validate electronic health record (EHR)-based machine-learning models to predict worsening heart failure (WHF) events across care settings and all-cause mortality among adults with mild-to-moderate chronic kidney disease (CKD). METHODS: We studied adults with mild-to-moderate CKD [estimated glomerular filtration rate (eGFR) 30-59 ml/min/1.73 m² or eGFR ≥60 with albuminuria] receiving care in a large health system from 2012 to 2021; outcomes were ascertained through 31 December 2022. Primary outcomes were (i) WHF events-outpatient encounters, emergency department (ED)/observation stays, and hospitalizations-identified using a validated natural language processing algorithm, and (ii) all-cause mortality. Models [extreme gradient boosting (XGBoost)] used an 80:20 train-test split and >500 EHR-derived covariates. Discrimination [area under the curve (AUC)] and calibration (slope) were evaluated overall and across subgroups by age, sex, race and ethnicity, and CKD stage. RESULTS: Among 375 495 adults (mean age 64 ± 16 years; 54% women; 53% non-Hispanic White; mean eGFR 76 ± 26 ml/min/1.73 m²), the WHF model achieved AUC 0.887 (95% CI 0.879-0.893) with calibration slope 0.955; the mortality model achieved AUC 0.875 (95% CI 0.868-0.883) with calibration slope 0.914 in the test set. Performance was consistent across age, sex, and race and ethnicity, with a slight decrement as CKD stage worsened. CONCLUSIONS: Electronic health record-based machine-learning models accurately predicted WHF and mortality in mild-to-moderate CKD with strong calibration across key subgroups. These models are positioned for EHR deployment to support risk-stratified cardiovascular-kidney-metabolic care-prioritizing guideline-directed therapies and care pathways for those at highest risk.