A web-based two-stage prediction model for postoperative pulmonary complications using preoperative and intraoperative data.

BACKGROUND

Most existing models for predicting postoperative pulmonary complications (PPCs) rely solely on preoperative variables and lack integration of intraoperative data or clinical tools for application.

METHODS

This study developed and externally validated logistic regression and machine learning (ML) models for predicting PPCs in non-cardiothoracic surgical patients undergoing general anesthesia. A cohort of 997 patients was randomly divided into training and test sets in a 7:3 ratio. The stepwise regression method was adopted to conduct feature screening on the preoperative dataset and the combined preoperative and intraoperative dataset, and a variety of ML algorithm models were developed. Model performance was evaluated using area under the receiver operating characteristic (AUROC), area under the precision-recall curve (AUPRC), brier score, sensitivity, specificity, precision, and F1-score. External validation was performed on an independent cohort of 286 patients from a different institution. The features of the optimized model were interpreted by introducing SHapley Additive exPlanations (SHAP) method.

RESULTS

The AUROC value of the logistic regression model using only preoperative variables in the test set was 0.828. After including intraoperative variables, the CatBoost model demonstrated superior performance, with an AUROC value of 0.927. In the preoperative model, C-reactive protein (CRP), American Society of Anesthesiologists (ASA) physical status classification, age, and smoking status were the most significant predictors. However, operative duration, CRP, age, and intraoperative blood loss were the dominant predictors of postoperative outcomes. External validation confirmed general applicability, with preoperative logistic regression and the Assess Respiratory Risk in Surgical Patients in Catalonia Tool (ARISCAT) yielding AUROC values of 0.827 and 0.796, respectively. The postoperative CatBoost model achieved an AUROC of 0.865, AUPRC of 0.809, sensitivity of 0.787, specificity of 0.921, and the lowest Brier score (0.132).

CONCLUSIONS

The preoperative model enabled early risk stratification, while the postoperative model provided improved accuracy. A publicly accessible web-based calculator was developed to support clinical implementation and facilitate prospective validation.