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

BACKGROUND

The coexistence of heart failure with preserved ejection fraction and chronic kidney disease is associated with high readmission risks. Existing tools, such as the Meta-Analysis Global Group in Chronic Heart Failure (MAGGIC) score, rely on linear modelling, which limits their ability to capture nonlinear interactions. We developed a machine learning model to predict readmission and assessed its performance against temporal data drift.

METHODS

In this retrospective study, patients were temporally stratified into derivation (n = 750) and independent validation (n = 130) cohorts. Feature selection used the least absolute shrinkage and selection operator and recursive feature elimination with cross-validation. Twelve algorithms were trained using stratified nested cross-validation. Model stability was evaluated using the Population Stability Index (PSI) and rolling metrics.

RESULTS

Ten features were identified, with estimated glomerular filtration rate as the primary predictor. An interaction was observed between high-sensitivity C-reactive protein and N-terminal pro-B-type natriuretic peptide. The random forest model achieved an area under the receiver operating characteristic curve of 0.837 (95% confidence interval, 0.761-0.905) in the validation cohort, compared with 0.551 for the MAGGIC score. Longitudinal analysis indicated temporal stability, with feature PSI scores < 0.25.

CONCLUSIONS

A machine learning model that incorporated renal, inflammatory, and hemodynamic parameters demonstrated greater predictive accuracy for 1-year readmission compared with the MAGGIC score. The model maintained temporal stability, in support of its potential utility for risk stratification.