Kidney Week

Abstract: TH-PO011

Machine-Learning Model for Predicting Early Mortality in Critically Ill Patients Requiring Continuous Kidney Replacement Therapy

Session Information

• Augmented Intelligence for Prediction and Image Analysis
  October 24, 2024 | Location: Exhibit Hall, 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

• Takkavatakarn, Kullaya, Icahn School of Medicine at Mount Sinai, New York, New York, United States

Kullaya Takkavatakarn, MD, PhD

• Lambert, Joshua, University of Cincinnati, Cincinnati, Ohio, United States

Joshua Lambert, PhD

• Kauffman, Justin, Icahn School of Medicine at Mount Sinai, New York, New York, United States

Justin Kauffman, MS

• Takeuchi, Tomonori, The University of Alabama at Birmingham, Birmingham, Alabama, United States

Tomonori Takeuchi, MD

• Cama-Olivares, Augusto, The University of Alabama at Birmingham, Birmingham, Alabama, United States

Augusto Cama-Olivares, MD

• Goldstein, Stuart, Cincinnati Children's Hospital Medical Center, Cincinnati, Ohio, United States

Stuart Goldstein, MD, FASN

• Chen, Jin, The University of Alabama at Birmingham, Birmingham, Alabama, United States

Jin Chen, PhD

• Chan, Lili, Icahn School of Medicine at Mount Sinai, New York, New York, United States

Lili Chan, MD, MS

• Nadkarni, Girish N., Icahn School of Medicine at Mount Sinai, New York, New York, United States

Girish N. Nadkarni, MD, MPH

• Neyra, Javier A., The University of Alabama at Birmingham, Birmingham, Alabama, United States

Javier A. Neyra, MD, MS, FASN

Background

Critically ill patients with acute kidney injury requiring continuous renal replacement therapy (CRRT) have high mortality and require significant resource utilization. Accurately predicting early mortality would facilitate clinical decision-making by identifying high morbid patients who may benefit from personalized care intensity with CRRT. We aimed to develop a model to predict mortality within the first 72 hours after initiating CRRT.

Methods

We obtained data from CRRTnet, a prospective multicenter data registry of adult patients undergoing CRRT for at least 24 hours. The cohort was divided into training and test datasets (80/20). We trained and validated XGBoost model with 10-fold cross-validation to predict mortality within 72 hours of CRRT initiation using patient demographics, laboratory results, vasopressors, fluid balance, and CRRT-related variables. We chose the parameters for the XGBoost model using Bayesian optimization. Bayesian optimization models the anticipated value of an objective function with a gaussian process over the space of all hyper parameters and uses Bayesian updating to choose parameter sets that maximize this function. Area under the receiver operating characteristic (AUROC) and precision-recall curve (AUPRC) were used to evaluate model performances. Shapley additive explanation (SHAP) value was applied to explain the model.

Results

We included 1,446 patients (59% men). Median age was 60 years (IQR 50-69), serum creatinine at CRRT initiation was 5.4 mg/dL (IQR 2.9-6.7), and APACHE II score was 29 (IQR 26-31). Among these patients, 236 (16.3%) died within 72 hours of CRRT initiation. The AUROC was 0.90 (95% CI 0.86 to 0.95), and the AUPRC was 0.64 (0.50 to 0.79) (Fig1A,1B). SHAP plot shows the importance of the top 20 features (Fig1C).

Conclusion

The proposed XGBoost model, the first to incorporate clinical and CRRT-related variables, shows promise in predicting early mortality in critically ill patients undergoing CRRT. Implementing the model could serve as an additional clinical tool for personalizing early CRRT use.

Funding

• Commercial Support – Baxter Healthcare via an Investigator Initiated Research grant award