Abstract: FR-OR010
Temporal Machine-Learning Model for Predicting Seven-Day RRT-Free Survival in Critically Ill Patients on Continuous Renal Replacement Therapy
Session Information
- Artificial Intelligence and Data Science Transforming Kidney Care: From Algorithms to Action
November 07, 2025 | Location: Room 361A, Convention Center
Abstract Time: 04:30 PM - 04:40 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Yang, Joanna, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Kauffman, Justin, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Pranto, Shehan Irteza, University of Alabama at Birmingham Health System Authority, Birmingham, Alabama, United States
- Takeuchi, Tomonori, University of Alabama at Birmingham Health System Authority, Birmingham, Alabama, United States
- Goldstein, Stuart, Cincinnati Children's Hospital Medical Center, Cincinnati, Ohio, United States
- Chen, Jin, University of Alabama at Birmingham Health System Authority, Birmingham, Alabama, United States
- Lambert, Joshua, University of Cincinnati, Cincinnati, Ohio, United States
- Nadkarni, Girish N., Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Chan, Lili, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Neyra, Javier A., University of Alabama at Birmingham Health System Authority, Birmingham, Alabama, United States
Background
Continuous renal replacement therapy (CRRT) is the preferred RRT modality in hemodynamically unstable patients. Accurate prediction of RRT-free survival using static and temporal features from EHR can improve risk stratification, guide CRRT delivery, and personalize care.
Methods
We utilized data from CRRTnet, a prospective multicenter registry of adult patients on CRRT >24h. TSFresh extracted temporal features over 1–6 day rolling windows. Multiple correspondence analysis (MCA) reduced static categorical feature space, preserving key information and boosting performance. We trained an XGBoost model using 80:20 train-test split to predict 7-day RRT-free survival and used SHAP values from the full dataset to explain the features impact on the model.
Results
We included 1,390 patients (59.6% male, median age 60 [IQR 50-69], 90.2% of registry participants), where 39.6% had 7-day RRT-Free survival. The best model used a 3-day rolling window (AUROC of 0.79 [95% CI 0.77-0.81], AUPRC of 0.84 [95% CI 0.82-0.87]) (Figure 1A). SHAP scores identified key predictors, including lab values (INR, albumin, pH), clinical scores (Charlson Comorbidity Score, neuro GCS), and fluid metrics, many of which are temporal (Figure 1B). MCA dimensions also had high SHAP scores, and the top 5 features per dimension are presented in Figure 1C.
Conclusion
We developed a machine learning model with high AUROC and AUPRC to predict 7-day RRT free survival, incorporating temporal features. Limitations include irregular measurement intervals and imputation (last observations or zero-fill), which may contribute to the consistent performance across rolling windows. This model, once validated and further evaluated, could offer a novel tool for guiding clinical decision-making for critically ill patients on CRRT.
Funding
- NIDDK Support