P141. Harnessing Machine Learning to Forecast Prolonged Intubation in Aortic Surgery Patients
Presented During: POD Abstracts (Available for viewing in the exhibit hall for the duration of the meeting)
University of Colorado, Anschutz Medical Center, Aurora, CO
Denver, CO
United States - Contact Me
Nicolas Chanes is a current PGY-1 surgical resident at the University of Colorado. He obtained his medical degree in 2023 from Louisiana State University School of Medicine in New Orleans. He is an active member of the Aortic Surgery Research Laboratory in the Department of Cardiothoracic Surgery at the University of Colorado. As a data scientist, he has specific interest in utilizing artificial intelligence to enhance outcomes in cardiothoracic surgery.
Thursday, April 25, 2024: 5:38 PM - 7:00 PM
Sheraton Times Square
Room: Central Park
Description
Objective:
Delayed extubation causes significant patient harm, with the potential for increased ventilator associated pneumonia, delirium associated with sedation, and prolonged recovery. Furthermore, delayed extubation places a significant burden on the health care system, with increased in-hospital costs and length of stay. Aortic surgery carries high risk for prolonged ventilation given its comorbid population and the potential for hemodynamic instability. To better predict those at risk for prolonged ventilation, we applied a machine learning model to all aortic surgeries at our institution.
Methods:
All adult patients undergoing aortic surgery from June 2009 to October 2022 (n = 875) were identified from our single institution prospectively maintained database. Patients were randomized 4:1 into training and testing cohorts to develop various eXtreme gradient boosting (XGBoost) models that predicted postoperative prolonged intubation (>24 hours) in the cardiothoracic intensive care unit (CTICU). We identified 64 input parameters from the index hospitalization, including 24 demographic characteristics, 8 preoperative variables, and 32 intraoperative parameters. To achieve the final model, we conducted hyperparameter fine-tuning involving 10-fold cross-validation at each iteration. Model performance was evaluated using multiple measures 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 utilized a SHapley Additive exPlanation (SHAP) violin plot to identify and interpret the impact of individual features on the predictions of the XGBoost model.
Results:
Postoperative prolonged intubation in the CTICU was noted in 81 patients (9.3%) who underwent aortic surgery. The final XGBoost model demonstrated a cross-validation accuracy of 89% and was well-calibrated as evidenced by the low Brier score of 0.09. The predictor also displayed robust performance on the test dataset, achieving an accuracy of 90%. Our best performing postoperative prolonged intubation prediction model achieved an AUC-ROC of 0.82 and an AUC-PR of 0.39. Upon stratification by aortic procedure, the model attained an AUC-ROC of 0.75 for root surgeries, 0.76 for hemiarch cases, and 0.74 for total arch procedures. The SHAP violin plot helped explain the complex decision-making process of our XGBoost model and provided insights into the top 25 key features that significantly influence model prediction. Key factors associated with an increased risk of postoperative prolonged intubation included extended durations of cardiopulmonary bypass and circulatory arrest, increased intraoperative blood product transfusion, advanced age, and a prior history of stroke.
Conclusions:
Our machine learning model accurately predicted those at most risk for prolonged ventilation. More complex surgeries and hemodynamic instability, as reflected by longer intraoperative cardiopulmonary bypass and circulatory arrest times as well as increased intraoperative transfusion, significantly augment the risk of prolonged ventilation. Machine learning empowers clinicians to have better, data-driven discussions with patients, offering personalized insights that allow further optimizations to improve outcomes.
Authors
Nicolas Chanes (1), Adam Carroll (1), Michael Kirsch (1), Bo Chang Wu (1), Muhammad Aftab (1), T. Brett Reece (1)
Institutions
(1) University of Colorado Anschutz, Denver, CO
Delayed extubation causes significant patient harm, with the potential for increased ventilator associated pneumonia, delirium associated with sedation, and prolonged recovery. Furthermore, delayed extubation places a significant burden on the health care system, with increased in-hospital costs and length of stay. Aortic surgery carries high risk for prolonged ventilation given its comorbid population and the potential for hemodynamic instability. To better predict those at risk for prolonged ventilation, we applied a machine learning model to all aortic surgeries at our institution.
Methods:
All adult patients undergoing aortic surgery from June 2009 to October 2022 (n = 875) were identified from our single institution prospectively maintained database. Patients were randomized 4:1 into training and testing cohorts to develop various eXtreme gradient boosting (XGBoost) models that predicted postoperative prolonged intubation (>24 hours) in the cardiothoracic intensive care unit (CTICU). We identified 64 input parameters from the index hospitalization, including 24 demographic characteristics, 8 preoperative variables, and 32 intraoperative parameters. To achieve the final model, we conducted hyperparameter fine-tuning involving 10-fold cross-validation at each iteration. Model performance was evaluated using multiple measures 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 utilized a SHapley Additive exPlanation (SHAP) violin plot to identify and interpret the impact of individual features on the predictions of the XGBoost model.
Results:
Postoperative prolonged intubation in the CTICU was noted in 81 patients (9.3%) who underwent aortic surgery. The final XGBoost model demonstrated a cross-validation accuracy of 89% and was well-calibrated as evidenced by the low Brier score of 0.09. The predictor also displayed robust performance on the test dataset, achieving an accuracy of 90%. Our best performing postoperative prolonged intubation prediction model achieved an AUC-ROC of 0.82 and an AUC-PR of 0.39. Upon stratification by aortic procedure, the model attained an AUC-ROC of 0.75 for root surgeries, 0.76 for hemiarch cases, and 0.74 for total arch procedures. The SHAP violin plot helped explain the complex decision-making process of our XGBoost model and provided insights into the top 25 key features that significantly influence model prediction. Key factors associated with an increased risk of postoperative prolonged intubation included extended durations of cardiopulmonary bypass and circulatory arrest, increased intraoperative blood product transfusion, advanced age, and a prior history of stroke.
Conclusions:
Our machine learning model accurately predicted those at most risk for prolonged ventilation. More complex surgeries and hemodynamic instability, as reflected by longer intraoperative cardiopulmonary bypass and circulatory arrest times as well as increased intraoperative transfusion, significantly augment the risk of prolonged ventilation. Machine learning empowers clinicians to have better, data-driven discussions with patients, offering personalized insights that allow further optimizations to improve outcomes.
Authors
Nicolas Chanes (1), Adam Carroll (1), Michael Kirsch (1), Bo Chang Wu (1), Muhammad Aftab (1), T. Brett Reece (1)
Institutions
(1) University of Colorado Anschutz, Denver, CO
Presentation Duration
PODS will be on display in the exhibit hall for the duration of the meeting during exhibit hall hours. PODS will also be available for viewing on the meeting website. There is no formal presentation associated with your POD, but we encourage you to visit the PODS area during breaks to connect with those viewing.