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 hours after ICU admission.
Approach:
Study Design: Retrospective observational study using MIMIC-IV for model development and internal validation.
Data Processing: Candidate predictors were processed using training-set imputation, encoding, and standardization.
Model Training: LASSO regression was used for feature selection; seven machine-learning algorithms were trained.
Model Evaluation: Model performance was evaluated using various metrics including AUC, AUPRC, and SHAP for interpretation.
Key Findings:
4,287 adult polytrauma ICU patients were included, with 623 developing AKI.
LASSO retained 15 early predictors for AKI.
Logistic regression achieved the highest AUC (0.900) and AUPRC (0.615) in the internal test set.
In the eICU-CRD cohort, SVM showed the highest external AUC (0.709).
Major contributors identified included SOFA score, weight, magnesium, and vital signs.
Interpretation:
A 15-variable logistic regression model demonstrated strong internal performance and interpretable risk signals for early AKI prediction after polytrauma.
Limitations:
Partial transportability of the model to external datasets.
Retrospective design may limit generalizability.
Conclusion:
The study supports further prospective validation and local recalibration before clinical implementation.
