Prediction model of early postoperative delirium risk based on machine learning algorithm in patients undergoing cardiac surgery

Abstract

Objective To develop an early postoperative delirium (POD) risk prediction model in patients undergoing cardiac surgery based on Extreme Gradient Boosting (XGBoost) and compare its prediction performance with that of a traditional logistic regression (LR) model in order to provide reference for early identification and timely intervention of the condition. Methods A case-control trial was conducted on 684 patients who underwent elective cardiac surgery under general anesthesia due to heart disease in the Second Affiliated Hospital of Army Medical University from March to July 2022. According to the outcome of their 3-day follow-up after operation, the patients were divided into delirium group (n=38) and non-delirium group (n=646). The patients were randomly divided into a training set (479 patients) and a test set (205 patients) at a ratio of 7∶3. LASSO regression analysis was used to screen out important variables related to POD. LR and XGBoost were employed to construct the prediction models. The area under receiver operating characteristic curve (ROC-AUC) of the prediction models and the sensitivity and specificity under the optimal threshold were calculated and the prediction performance of different models was compared. Results The 3-day postoperative delirium rate of the patients undergoing cardiac surgery was 5.56%. Compared with the non-delirium group, the patients in the delirium group were older (P 0.05). Then finally, 5 variables, including age, preoperative peripheral oxygen saturation, preoperative regional cerebral oxygen saturation, preoperative systolic blood pressure and postoperative sleep score, were included for modelling. The AUCs of LR and XGBoost models were 0.732 (95%CI: 0.43~1.000) and 0.659 (95%CI: 0.559~0.759), respectively. LR model had a higher AUC value and better predictive performance, but, its sensitivity was 50%, lower than that of XGBoost model (67%), and its specificity was 100%, higher than the other model (98.5%). Conclusion The predictive performance of the prediction model based on XGBoost, an integrated learning algorithm, is not superior to the traditional LR model for postoperative delirium after cardiac surgery. LR model can well predict the occurrence of delirium after cardiac surgery and provide reference for early intervention and treatment. But XGBoost is more sensitive to the diagnosis of postoperative delirium.

Keywords

• cardiac surgery
• postoperative delirium
• machine learning
• prediction model