An interpretable machine-learning model for early prediction of acute kidney injury in polytrauma patients.

Objective: Acute kidney injury (AKI) is a frequent and clinically important complication after polytrauma, but early risk stratification remains challenging because conventional diagnostic criteria depend on delayed changes in serum creatinine and urine output. This study aimed to develop and interpret an early prediction model for incident AKI in adult ICU patients with polytrauma using routinely available data from the first 6 h after ICU admission. Methods: We conducted a retrospective observational study using MIMIC-IV as the primary development and internal validation dataset. Incident AKI was defined according to KDIGO criteria and assessed between 7 and 72 h after ICU admission. Candidate predictors measured within the first 6 h were processed using training-set imputation, encoding, and standardization. LASSO regression was used for feature selection. Seven machine-learning algorithms were trained using the retained predictors after applying SMOTE combined with random undersampling in the training set only. Model performance was evaluated using discrimination, precision-recall performance, threshold-based metrics, calibration, Brier score, and decision curve analysis. SHAP was used to interpret the selected primary model. A secondary eICU-CRD analysis assessed cross-database reproducibility. Results: The MIMIC-IV cohort included 4,287 adult polytrauma ICU patients, of whom 623 developed AKI. LASSO retained 15 early predictors. In the internal test set, logistic regression achieved the highest AUC and AUPRC among the evaluated models (AUC 0.900, 95% CI 0.878-0.923; AUPRC 0.615, 95% CI 0.533-0.703), with high sensitivity (0.888) and negative predictive value (0.974). In the eICU-CRD cohort, SVM showed the highest external AUC (0.709), while logistic regression demonstrated moderate transportability (AUC 0.678). SHAP identified SOFA score, weight, magnesium, antihypertensive medication exposure, age, respiratory rate, heart rate, platelet count, urine output, and temperature as major contributors. Conclusion: A parsimonious 15-variable logistic regression model using 6-h ICU data showed strong internal performance and clinically interpretable risk signals for early AKI prediction after polytrauma. External validation indicated partial transportability, supporting further prospective validation and local recalibration before clinical implementation.