Atherogenic index of plasma and risk of acute kidney injury in critically ill patients: a multi-cohort study with machine learning and SHAP analysis

Abnormal lipid metabolism poses a risk for acute kidney injury (AKI), a prevalent complication among critically ill patients. Early detection and intervention of AKI are essential; however, reliable lipid-derived predictive biomarkers remain understudied. This study aimed to investigate the relationship between the novel lipid biomarker atherogenic index of plasma (AIP) and AKI in critically ill patients and to construct an AIP-integrated early warning model to identify high-risk patients. Patient data were derived from the MIMIC-IV database (training and internal validation sets) and a local hospital cohort (external validation set). Multivariable logistic regression was used to evaluate the association between AIP and AKI. Restricted cubic spline (RCS) regression was utilized to explore potential nonlinear relationships. 13 machine learning algorithms were applied to develop and validate prediction models. Additionally, the Shapley Additive Explanations (SHAP) method enhance model interpretability. We analyzed 6,062 ICU patients from the MIMIC-IV database and 833 patients from a Chinese hospital cohort. AIP was identified as an independent risk factor for AKI in multivariable logistic regression analyses. RCS regression revealed a nonlinear association between AIP and AKI. The XGBoost + AIP model achieved superior performance with AUCs of 0.8127 (internal) and 0.7228 (external), significantly outperforming the SOFA score (AUC 0.6968). Decision curve analysis (DCA) confirmed its clinical applicability. The SHAP method provides critical validation support for the reliability of the XGBoost model. AIP serves not only as a predictive biomarker but as a metabolic phenotype, potentially enabling AI-guided precision prevention strategies in future digital twin-driven AKI care pathways.