Dynamic machine learning models for the early prediction of acute kidney injury following left ventricular assist device implantation: An explainable AI approach

The risk of developing acute kidney injury (AKI) following left ventricular assist device (LVAD) implantation is high, making its prediction essential. Current tools are limited by relying on static, preoperative data and "black-box" models. Our aim was to create a transparent prediction tool for AKI that integrates dynamic postoperative laboratory results with explainable artificial intelligence (XAI). In a cohort of 138 LVAD patients, AKI was defined using KDIGO criteria within 7 days post-surgery. We evaluated LASSO, Random Forest, and XGBoost models, using the SMOTE method to manage class imbalance within a stratified 10-fold cross-validation. The LASSO model achieved the best performance with a ROC-AUC of 0.755. SHAP analysis revealed that dynamic postoperative laboratory changes were as important as baseline renal function in predicting AKI. Using risk scores calculated 48 h post-surgery, the model successfully predicted whether day 7 AKI would be "persistent" or "transient." Accurately predicting post-LVAD AKI is feasible using an interpretable machine learning model that tracks dynamic data. This approach can serve as a reliable tool for early risk assessment, enabling physicians to identify high-risk patients earlier and implement proactive interventions.