Predicting Renal Replacement Therapy After Total Arch Surgery Using Machine Learning
Presented During:
Thursday, April 25, 2024: 5:38PM - 7:00PM
Sheraton Times Square
Posted Room Name:
Central Park
Abstract No:
P0258
Submission Type:
Abstract Submission
Authors:
Adam Carroll (1), Nicolas Chanes (1), Michael Kirsch (1), Bo Chang Wu (1), Muhammad Aftab (1), T. Brett Reece (1)
Institutions:
(1) University of Colorado Anschutz, Denver, CO
Submitting Author:
Adam Carroll
-
Contact Me
University of Colorado Anschutz
University of Colorado Anschutz
Co-Author(s):
Nicolas Chanes
-
Contact Me
University of Colorado Anschutz
University of Colorado Anschutz
Michael Kirsch
-
Contact Me
University of Colorado Anschutz
University of Colorado Anschutz
Bo Chang Wu
-
Contact Me
University of Colorado Anschutz
University of Colorado Anschutz
*Muhammad Aftab
-
Contact Me
University of Colorado Anschutz
University of Colorado Anschutz
*T. Brett Reece
-
Contact Me
University of Colorado Anschutz
University of Colorado Anschutz
Presenting Author:
Adam Carroll
-
Contact Me
University of Colorado Anschutz
University of Colorado Anschutz
Abstract:
Objective:
Patients undergoing total arch surgery are at high risk of acute kidney injury, increasing the risk of morbidity and mortality. Identifying patients at risk for severe acute kidney injury may help to improve patient outcomes. We developed a machine-learning model to identify patients at risk for new renal replacement therapy after total arch surgery.
Methods:
From our single institution prospectively maintained database, we identified a total of 235 patients who underwent total arch surgery between June 2009 and October 2022. These patients were randomly divided into training (70%) and testing (30%) sets and various eXtreme gradient boosting (XGBoost) models were constructed to predict the need for postoperative renal replacement therapy (RRT) in the cardiothoracic intensive care unit (CTICU). From the index hospitalization data, we extracted 64 input parameters including demographic information as well as preoperative and intraoperative characteristics. To assess model performance, we utilized multiple evaluation metrics, including accuracy, Brier score, area under the receiver operating characteristic curve (AUC-ROC), and area under the precision-recall curve (AUC-PR, mean average precision). We also employed a SHapley Additive exPlanation (SHAP) violin plot to discern the influence of individual features on the predictions generated by the XGBoost model.
Results:
Postoperative RRT in the CTICU was noted in 25 patients (10.6%) who underwent total arch surgery. The final XGBoost model demonstrated a cross-validation accuracy of 90% and exhibited strong calibration, as indicated by the low Brier score of 0.10. Additionally, the predictor displayed robust performance on the test dataset, achieving an accuracy of 92%. Our top-performing postoperative RRT prediction model attained an AUC-ROC of 0.78 on the training set and an AUC-ROC of 0.88 on the testing set. The SHAP violin plot assisted in elucidating the intricate decision-making process employed by our XGBoost model, offering insights into the top 10 features that exert significant influence on model predictions. Prominent risk factors linked to an elevated risk of postoperative RRT included low preoperative creatinine levels, increased intraoperative blood product transfusion, extended cardiopulmonary bypass durations, and reduced nadir hemoglobin levels.
Conclusions:
Our machine learning model provided insight into preoperative and operative factors associated with increased risk of need for post-operative renal replacement therapy. Lower creatinine, likely a byproduct of decreased muscle mass in a frail population, as well as longer cardiopulmonary bypass with more intraoperative transfusion likely reflecting lower intraoperative DO2 during the procedure were significant risk factors for developing severe kidney injury.
Patients undergoing total arch surgery are at high risk of acute kidney injury, increasing the risk of morbidity and mortality. Identifying patients at risk for severe acute kidney injury may help to improve patient outcomes. We developed a machine-learning model to identify patients at risk for new renal replacement therapy after total arch surgery.
Methods:
From our single institution prospectively maintained database, we identified a total of 235 patients who underwent total arch surgery between June 2009 and October 2022. These patients were randomly divided into training (70%) and testing (30%) sets and various eXtreme gradient boosting (XGBoost) models were constructed to predict the need for postoperative renal replacement therapy (RRT) in the cardiothoracic intensive care unit (CTICU). From the index hospitalization data, we extracted 64 input parameters including demographic information as well as preoperative and intraoperative characteristics. To assess model performance, we utilized multiple evaluation metrics, including accuracy, Brier score, area under the receiver operating characteristic curve (AUC-ROC), and area under the precision-recall curve (AUC-PR, mean average precision). We also employed a SHapley Additive exPlanation (SHAP) violin plot to discern the influence of individual features on the predictions generated by the XGBoost model.
Results:
Postoperative RRT in the CTICU was noted in 25 patients (10.6%) who underwent total arch surgery. The final XGBoost model demonstrated a cross-validation accuracy of 90% and exhibited strong calibration, as indicated by the low Brier score of 0.10. Additionally, the predictor displayed robust performance on the test dataset, achieving an accuracy of 92%. Our top-performing postoperative RRT prediction model attained an AUC-ROC of 0.78 on the training set and an AUC-ROC of 0.88 on the testing set. The SHAP violin plot assisted in elucidating the intricate decision-making process employed by our XGBoost model, offering insights into the top 10 features that exert significant influence on model predictions. Prominent risk factors linked to an elevated risk of postoperative RRT included low preoperative creatinine levels, increased intraoperative blood product transfusion, extended cardiopulmonary bypass durations, and reduced nadir hemoglobin levels.
Conclusions:
Our machine learning model provided insight into preoperative and operative factors associated with increased risk of need for post-operative renal replacement therapy. Lower creatinine, likely a byproduct of decreased muscle mass in a frail population, as well as longer cardiopulmonary bypass with more intraoperative transfusion likely reflecting lower intraoperative DO2 during the procedure were significant risk factors for developing severe kidney injury.
Aortic Symposium:
Aortic Arch
Image or Table
