Predicting Prolonged Intubation After Hemiarch Surgery Using Machine Learning

N. Chanes¹, A. Carroll², M. Aftab³, T. Reece^4
1University of Colorado, Metairie, Louisiana ²University of Colorado School of Medicine, Denver, Colorado ³University of Colorado, Anschutz Medical Center, Aurora, CO, Aurora, Colorado ⁴Univ of Colorado Dept of Surgery Div of Cardiothoracic, Aurora, Colorado

Purpose: Prolonged intubation results in increased risk of ventilator-associated pneumonia, mortality, and prolonged hospital stay(1). Machine learning, which may provide insight into patient outcomes, has had limited application for prolonged intubation after aortic surgery(2). We applied a machine learning algorithm to patients undergoing hemiarch surgery to better elucidate at risk patients.

Methods: All adult patients undergoing hemiarch replacement from January 2010 to August 2022 (n = 579) were identified from our single institution prospectively maintained database. Patients were randomly divided into training (80%) and testing (20%) sets with various eXtreme gradient boosting (XGBoost) models constructed to predict post-operative prolonged intubation (>24 hours) in the cardiothoracic intensive care unit (CTICU). We identified 50 input parameters from the index hospitalization, including 19 demographic characteristics as well 9 pre-operative and 22 intra-operative variables. Hyperparameter fine-tuning with 10-fold cross-validation at each iteration was performed to attain the final model. 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) beeswarm plot to identify and interpret the impact of individual features on the predictions of the XGBoost model.

Results: Development of post-operative prolonged intubation in the CTICU was noted in 38 patients (6.56%) who underwent hemiarch replacement, including urgent/emergent cases and dissections. The final XGBoost model demonstrated a cross-validation accuracy of 94% and was well-calibrated as evidenced by the low Brier score of 0.06. The predictor also demonstrated strong performance on the testing set, achieving an accuracy of 91%. Our best performing CTICU post-operative prolonged intubation prediction model achieved an AUC-ROC of 0.66 and an AUC-PR of 0.25. The SHAP beeswarm 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. Each instance in the dataset was represented by a single dot, with the relative value of each input parameter graded from high to low. Input parameters associated with an increase in post-operative prolonged intubation display positive SHAP values, meanwhile features associated with a decrease in post-operative prolonged intubation display negative SHAP values.

Conclusion: Our final XGBoost machine learning model provided excellent accuracy in identifying individuals at risk of prolonged intubation after hemiarch surgery and highlighted significant predictive features. Better predicting patients at risk for prolonged intubation may alter ventilation strategies and help guide clinical management.

Identify the source of the funding for this research project: Division of Cardiothoracic Surgery, University of Colorado