Temporal Validation of a Machine Learning Readmission Model in Heart Failure With Preserved Ejection Fraction and Chronic Kidney Disease

Comprehensive Summary

Deng et al. developed and temporally validated a machine learning model to predict 1-year hospital readmission in patients with heart failure with preserved ejection fraction (HFpEF) and chronic kidney disease (CKD). In this retrospective single-center study, 880 patients were divided into a derivation cohort (n = 750) and an independent temporal validation cohort (n = 130). Feature selection using LASSO and recursive feature elimination identified 10 predictors, and 12 machine learning algorithms were trained using nested cross-validation. The random forest model demonstrated the best performance, achieving an AUC of 0.837 (95% CI: 0.761-0.905) in the validation cohort, outperforming the traditional MAGGIC risk score (AUC: 0.551). SHAP interpretability analysis identified estimated glomerular filtration rate (eGFR) as the strongest predictor, with a notable interaction between high-sensitivity C-reactive protein (hs-CRP) and NT-proBNP. Temporal validation demonstrated consistent model performance despite minor feature drift, with population stability index values staying below 0.25.

Outcomes and Implications

HFpEF patients with CKD represent a high-risk cardiorenal phenotype with substantial readmission risk, and traditional risk scores often perform poorly in this subgroup. This study demonstrates that machine learning models incorporating renal, inflammatory, and hemodynamic markers can improve readmission prediction compared with conventional linear risk scores. The identification of a non-linear interaction between inflammatory markers and hemodynamic stress also provides insight into the pathophysiology of this cardiorenal phenotype. If externally validated, such models could support earlier identification of high-risk patients and guide targeted interventions to reduce hospital readmissions.