Abstract: FR-PO0006
Prediction of AKI Stage 2+ Within Seven Days of Intubation Using First-Day Intensive Care Unit (ICU) Data
Session Information
- Artificial Intelligence: Clinical Prediction, Decision Support, and Digital Kidney Care
October 23, 2026 | Location: Exhibit Hall A, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Pitogo, Rica Mae Panogalinog, University of California Los Angeles, Los Angeles, California, United States
- Yean, Rithy Techavoan, University of California Los Angeles, Los Angeles, California, United States
- Nagal, Darryl, University of California Los Angeles, Los Angeles, California, United States
Background
Acute kidney injury (AKI) affects 30-68% of mechanically ventilated ICU patients, yet existing models treat ventilation as a covariate rather than the clinical index event. Progression to AKI Stage 2+ substantially increases mortality and renal replacement therapy needs. We developed machine learning models to predict AKI Stage 2+ within 7 days of intubation using the first 24 hours of routine clinical data.
Methods
Using MIMIC-IV (v3.1), we identified 10,085 mechanically ventilated adult ICU stays with no or mild AKI (Stage ≤1) at intubation. Patients with ESRD, prior transplant, elective surgery, or death within 48 hours were excluded. The outcome was progression to AKI Stage 2+ within 168 hours. Missing data were imputed via MICE. Logistic Regression, LASSO, Random Forest, and XGBoost were trained on an 80/20 stratified split with 5-fold cross-validation and Bayesian hyperparameter optimization. Performance was assessed using AUROC, AUPRC, Brier score, and Youden's J-optimized thresholds, with predictions pooled via Rubin's rules.
Results
AKI Stage 2+ occurred in 15.4% of cases. XGBoost with Bayesian tuning achieved the highest discrimination while Logistic Regression and LASSO performed equivalently (DeLong p=0.52). At optimized thresholds, XGBoost yielded 77.8% sensitivity and 68.9% specificity. SHAP analysis identified mean phosphate, total fluid output, mean bicarbonate, and vasopressor exposure as top predictors. No renal-specific features were included.
Conclusion
First-day ICU data predicted AKI Stage 2+ within 7 days of intubation with an AUROC of 0.80 across linear and nonlinear models. Similar discrimination across model families suggests the early predictive signal is predominantly linear, favoring interpretable models for deployment as automated EHR alerts at intubation. External validation is needed.
Acknowledgment
This study utilized the Medical Information Mart for Intensive Care IV (MIMIC-IV v3.1) database, a de-identified electronic health record dataset from Beth Israel Deaconess Medical Center, accessed via PhysioNet under a credentialed data use agreement. The study was deemed exempt from IRB review as MIMIC-IV is fully de-identified per HIPAA Safe Harbor standards. The authors acknowledge the contributions of the MIT Laboratory for Computational Physiology.
Performance Metrics of Different Machine Learning Models Deployed
| Model | AUROC | AUPRC | Brier Score | Optimal Threshold | Youden's J |
| XGBoost (Bayes) | 0.806 | 0.453 | 0.1062 | 0.12 | 0.467 |
| Logistic Regression | 0.803 | 0.448 | 0.1067 | 0.17 | 0.464 |
| LASSO | 0.802 | 0.447 | 0.1067 | 0.20 | 0.469 |
| Random Forest | 0.795 | 0.434 | 0.1082 | 0.19 | 0.454 |
| SVM | 0.787 | 0.423 | 0.1304 | 0.05 | 0.000 |
AUROC and AUPRC Metric Graphs