Abstract: TH-PO019
Machine Learning-Based Risk Prediction Model for ICU Survival After Continuous Renal Replacement Therapy Initiation: A WEROCK Study
Session Information
- AI, Digital Health, Data Science - I
November 02, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Menon, Shina, University of Washington, Seattle, Washington, United States
- Thadani, Sameer, Baylor College of Medicine, Houston, Texas, United States
- Gist, Katja M., Cincinnati Children's Hospital Medical Center, Cincinnati, Ohio, United States
- Soranno, Danielle Elise, Indiana University School of Medicine, Indianapolis, Indiana, United States
- Wu, Danny T.y., Cincinnati Children's Hospital Medical Center, Cincinnati, Ohio, United States
Group or Team Name
- WE-ROCK.
Background
Continuous renal replacement therapy (CRRT) is preferred in critically ill pediatric patients with acute kidney injury (AKI) and/or fluid overload (FO). Prediction of clinical outcomes in patients on CRRT is challenging given heterogeneous clinical practices and study population. Our aim was to predict survival to intensive care unit (ICU) and hospital (HOSP) discharge using machine learning (ML) techniques.
Methods
The Worldwide Exploration of Renal Replacement Outcomes Collaborative in Kidney Disease (WE-ROCK) study is a retrospective international multicenter study (32 centers, 7 nations). Our methods have been published previously. We included 991 patients aged 0-25 years treated with CRRT for AKI and/or FO from 2018-2021. Primary outcomes were survival to ICU and HOSP discharge. Feature selection was done manually through team discussions, resulting in 61 out of 119 demographics variables. The data were split into training(80%) and testing(20%) subsets. The ML process included 5 algorithms (Logistic Regression with L2 regularization, Decision Tree, Random Forest, Gradient Boosting Machine, and Support Vector Machine with linear kernel) and used 5-fold cross validation to train the models. The model performance was determined by the Area Under the Curve of the Receiver Operating Characteristic (AUCROC). The performance of the best trained model on the testing dataset was reported using standard metrics (Table 1).
Results
Random Forest was our highest performing algorithm. Table 1 shows that our models (WE-ROCK) achieved comparable or slightly higher AUROC compared to previous studies in adults.
Conclusion
This is the first ML model to predict survival to ICU and hospital discharge in pediatric patients requiring CRRT. The performance was similar to previously published models in critically ill adults. We will continue refining our models by expanding the input variables, conducting a more sophisticated feature selection, experimenting advanced ML and deep learning algorithms.
Table 1