Development Of A Prediction Model And Analysis Of Predictive Value Of Postoperative Delirium In Patients With Extensive Burns

Objective:To learn the occurrence of postoperative delirium(POD)in extensive burn patients,analyze the related risk factors,establish an early risk prediction model of POD in extensive burn patients,and analyze the predictive value of the model.Methods:Patients with extensive burns admitted to the burn center of the First Affiliated Hospital of Nanchang University between March 2013 and March 2022 were selected for this study.By collecting relevant data from the enrolled patients,independent risk factors for POD in extensive burn patients were determined by multifactorial logistic regression analysis.The information included in the study subjects was used as the complete dataset,which was randomly divided into a training and validation set(total80% dataset)and a testing set(total 20% dataset)according to 8:2.Predictive models are developed based on nine different machine learning(ML)algorithms including(1)Support Vector Machine(SVM),(2)Logistic Regression(LR),(3)Random Forest(RF),(4)Light Gradient Boosting Machine(Light GBM),(5)Adaptive Boosting(Ada Boost),(6)Gaussian NB(GNB),(7)Complementary NB(CNB),(8)Multi-layers Perceptron(MLP)and(9)e Xtreme Gradient Boosting(XGBoost).Learning training of each model and internal validation are performed in the training and validation sets with a 10-fold cross-validation method,and the best prediction model is selected based on area under the receiver operating curve(AUROC),calibration curve and decision curve analysis(DCA).Based on the Youden index to determine the optimal cut-off value for the best prediction model.The best predictive model performance was further validated with data from the testing set,and SHapley Additive ex Planations(SHAP)was used to explain the best model.Data related to patients with extensive burns admitted to the burn unit of Ganzhou People’s Hospital from September 2017 to September 2022 were collected and used for external validation of the best prediction model to assess its generalization ability.Results:1.In this study,518 patients with extensive burns admitted to the burn center of the First Affiliated Hospital of Nanchang University from March 2013 to March 2022 were collected,and 191 of them had POD,the incidence rate was 36.87%.External validation focused on 118 patients with extensive burns,31 of whom developed POD,the incidence rate of which was 26.27%.2.This study included a total of 48 study variables.The results of their univariate analysis showed statistically significant differences between 30 study variables(P < 0.05).Multifactorial logistic regression analysis of these 30 variables showed the following variables: sex(OR=4.262,95%CI: 2.140～8.485),history of diabetes(OR=4.943,95%CI: 1.349～18.114),physical restraint(OR=8.877,95%CI:5.057～15.584),length of stay in the burn intensive care unit(BICU)(OR=1.183,95%CI: 1.004～1.393),APACHE Ⅱ score(OR=1.111,95%CI: 1.047～1.178),total body surface area(TBSA)(OR=1.042,95%CI: 1.009～1.075)and hemoglobin(OR=0.958,95%CI: 0.941～0.975)were independent risk factors for POD in extensive burn patients.3.The performance of the nine prediction models was compared,and RF performed best in the calibration curve and DCA analysis.In the validation set,the AUROC of RF was 0.840(95%CI:0.762～0.918),the AUROC of XGBoost was 0.810(95% CI: 0.726～0.895),the AUROC of LR was 0.816(95% CI: 0.734～0.899),the AUROC of Light GBM was 0.737(95%CI: 0.636～0.837),Ada Boost’s AUROC was0.767(95%CI: 0.677～0.858),the AUROC of GNB was 0.814(95%CI: 0.731～0.897),the AUROC of CNB was 0.683(95%CI: 0.578～0.788),the AUROC of MLP was0.630(95%CI: 0.522～0.738)and the AUROC of SVM was 0.684(95%CI:0.581～0.788),and the highest AUROC for RF.Therefore,RF was chosen as the best prediction model with an optimal cut-off value of 0.4219.4.The RF was externally validated after visualization as a web calculator.RF performed well in discrimination,calibration and clinical utility in external validation.The accuracy was 77.12%,specificity was 80.46% and sensitivity was 67.74%,with good generalization to clinical use.Conclusion:1.The incidence of POD in patients with extensive burns was 36.87%,suggesting that medical staff should strengthen the screening and identification of patients at high risk of POD in clinical work,and take timely targeted interventions to reduce the incidence of POD and improve the prognosis of patients with extensive burns.2.Multifactorial logistic regression analysis showed that TBSA,length of stay in the BICU,APACHE Ⅱ score,sex,history of diabetes,hemoglobin,and physical restraint were independent risk factors for POD in extensive burn patients.3.Nine POD prediction models for patients with extensive burns were constructed based on ML algorithm,and the prediction efficacy of different prediction models differed.By validating the models,the prediction efficacy of RF was the most,which could effectively identify patients at high risk of POD and provide an effective reference for the development of individualized interventions for this population.4.The predictive model visualized as a web calculator is easy and fast to use.External validation results show that RF has good generalizability and can be extended for use in clinical settings.