Development and external validation of a machine learning model based on preoperative nutritional status for predicting acute kidney injury after coronary artery bypass grafting

Background Acute kidney injury (AKI) is a common complication after coronary artery bypass grafting (CABG). Preoperative nutritional status may influence AKI risk, but its predictive value remains unclear. Methods We retrospectively analyzed 811 CABG patients from two centers. Nutritional status was assessed using the Controlling Nutritional Status (CONUT) score, Prognostic Nutritional Index (PNI), and Geriatric Nutritional Risk Index (GNRI). Logistic regression and restricted cubic splines evaluated associations with AKI. The most predictive index, combined with key clinical variables selected via LASSO and Boruta, was used to build six machine-learning models. Model interpretability was assessed using SHAP, and a web-based calculator was deployed. Results All three indices were independently associated with AKI, with PNI performing best (AUC = 0.617). The GBM model showed highest predictive performance with AUCs of 1.000, 0.978, and 0.905 in training, internal, and external validation sets, respectively. SHAP identified PNI, LVEF, and CPB as top contributors. Conclusion Preoperative nutritional status, particularly PNI, is an independent predictor of AKI. An interpretable GBM model incorporating nutritional and clinical variables enables accurate individualized risk assessment.