Causal-Weighted Machine Learning for RRT Prediction in Acute Kidney Injury: Development, External Validation, and Clinical Decision Support
Abstract Background: Acute kidney injury (AKI) affects 20--50% of ICU admissions, and the optimal timing for renal replacement therapy (RRT) initiation remains contested. Methods: We developed and externally validated a leakage-safe, causal-weighted machine learning pipeline predicting 48-hour RRT initiation in patients with KDIGO stage 3 AKI, using harmonized MIMIC-IV and eICU data. A regularized stacking ensemble with inverse-probability-weighted meta-learning combined five base learners (LightGBM, XGBoost, Random Forest, CatBoost, ExtraTrees) via a logistic-regression meta-learner. Results: The model achieved AUC 0.894 internally and 0.807 on an independent, 200+ hospital external cohort, significantly outperforming four of five baseline models (DeLong \((p<0.001)\)) while remaining statistically indistinguishable from a single LightGBM model. Stacking-consistent SHAP identified creatinine and urine output trajectories as dominant risk drivers, 40% of high-risk patients had a feasible short-term recourse below threshold, and fairness assessment showed no clinically meaningful subgroup gaps. Conclusions: A prototype clinical decision support system operationalizes the model for bedside use, motivating prospective evaluation before deployment.
Authors
- Xiaozhu Liao
- LIAO Zhaoyu
- Qiang Du
- Ziyi Chen
- Qiuchen Huang
Institutions
- The University of Melbourne (AU)
- First People’s Hospital of Zunyi (CN)
- RMIT University (AU)
- Tsinghua University (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-13
- DOI
- https://doi.org/10.21203/rs.3.rs-11021990/v1
- Primary Topic
- Acute Kidney Injury Research
- Type
- preprint