Abstract
BACKGROUND: Early identification of acute kidney injury (AKI) in the intensive care unit (ICU) remains challenging. We aimed to identify key predictors of new-onset AKI within 48 h after ICU admission and renal replacement therapy (RRT) need within 7 days, using explainable artificial intelligence (XAI) with eXtreme Gradient Boosting (XGBoost). We also assessed whether XGBoost improved predictive performance. METHODS: A retrospective cohort study across four ICUs was conducted as part of the SWECRIT biobank project. Blood samples were prospectively obtained at ICU admission and retrospectively analysed. AKI was defined by the Kidney Disease: Improving Global Outcomes (KDIGO) criteria. XAI models were compared with logistic regression models, incorporating emerging biomarkers and routine clinical data at ICU admission. SHapley Additive exPlanations (SHAP) were used to identify key predictors. Discrimination was assessed using the mean area under the receiver operating characteristic curve (AUC). RESULTS: The study included 4732 admissions, with 2603 analysed for new-onset AKI and 4716 for RRT. Top predictors of new-onset AKI were urine output, endostatin, baseline creatinine, lactate, and albumin. Top predictors of RRT were creatinine, urine output, endostatin, neutrophil gelatinase-associated lipocalin (NGAL), and the Simplified Acute Physiology Score (SAPS) 3. Several clinically relevant non-linear relationships were revealed. XGBoost outperformed logistic regression for both new-onset AKI (mean AUC 0.76, 95% CI 0.70-0.81 vs. 0.74, 95% CI 0.68-0.81; p < 0.001) and RRT (mean AUC 0.92, 95% CI 0.89-0.95 vs. 0.90, 95% CI 0.87-0.94; p < 0.001). CONCLUSION: XGBoost identified key predictors of early new-onset AKI and RRT need in the ICU, highlighting both emerging (endostatin, NGAL) and established biomarkers (lactate, albumin), alongside known clinical predictors. It also improved predictive accuracy for both outcomes. Further clinical evaluation of these biomarkers and XAI models is warranted.