Machine Learning–Based Model for Predicting Acute Kidney Injury in Patients Hospitalized with Heart Failure: Development and Validation Study

<h2>Abstract</h2><h3>Background</h3> Acute kidney injury (AKI) substantially worsens outcomes in patients hospitalized with heart failure, yet effective early prediction tools remain limited. This study aimed to develop and validate a machine learning–based model for AKI prediction in heart failure patients. <h3>Methods</h3> We retrospectively analyzed 870 patients hospitalized for heart failure between October 2017 and June 2024. Missing values (<30%) were imputed, and feature selection was performed using LASSO and backward stepwise logistic regression. Five machine learning models (XGBoost, LightGBM, Logistic Regression, Support Vector Machine, and Decision Tree) were developed and evaluated using ROC curves, precision–recall curves, calibration plots, and decision curve analysis. <h3>Results</h3> AKI occurred in 271 patients (31.2%). Baseline comparisons showed significant differences in renal function and electrolyte levels between AKI and non-AKI groups (all P<0.05). Ten potential predictors were identified by LASSO, and seven remained significant after logistic regression. Among all models, XGBoost achieved the best discrimination with AUC of 0.927 (95% CI: 0.902–0.951) in the validation set. It showed excellent sensitivity (0.761), specificity (0.967), and positive predictive value (0.911). Calibration and decision curve analyses confirmed strong agreement and net clinical benefit. SHAP analysis indicated chronic kidney disease (OR=2.805, 95% CI: 1.461–5.383) and electrolyte disturbances as key predictors. <h3>Conclusions</h3> The proposed machine learning–based model accurately predicts AKI risk in heart failure patients, outperforming conventional methods. Its interpretability and robust performance suggest promising utility as a clinical decision support tool, warranting external validation.