Related Experiment Videos
A web-based two-stage prediction model for postoperative pulmonary complications using preoperative and
Anran Dai1,2, Hao Liu3, Yuqi Dai4
1Department of Pharmacy, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou 310000, China.
Journal of Anesthesia and Translational Medicine
|June 15, 2026
Summary
This study developed advanced models to predict postoperative pulmonary complications (PPCs) in surgical patients. Machine learning models integrating intraoperative data significantly improved prediction accuracy over preoperative-only models.
Area of Science:
- Anesthesiology and Perioperative Medicine
- Medical Informatics
- Machine Learning in Healthcare
Background:
- Existing models for predicting postoperative pulmonary complications (PPCs) often lack integration of intraoperative data.
- Clinical tools for predicting PPCs require enhancement for broader application.
Purpose of the Study:
- To develop and externally validate logistic regression and machine learning (ML) models for predicting PPCs in non-cardiothoracic surgical patients.
- To compare the predictive performance of models using preoperative variables versus those incorporating intraoperative data.
- To enhance the accuracy and clinical utility of PPC prediction tools.
Main Methods:
- Development and validation of logistic regression and ML models (including CatBoost) on a cohort of 997 patients.
- Feature selection using stepwise regression on preoperative and combined datasets.
- Performance evaluation using AUROC, AUPRC, Brier score, sensitivity, specificity, precision, and F1-score.
- External validation on an independent cohort of 286 patients.
- Interpretation of model features using SHapley Additive exPlanations (SHAP).
Main Results:
- The CatBoost model integrating intraoperative variables achieved a superior AUROC of 0.927 compared to preoperative-only models.
- Key predictors included C-reactive protein (CRP), ASA physical status, age, smoking status (preoperative), and operative duration, CRP, age, intraoperative blood loss (postoperative).
- External validation confirmed generalizability, with the CatBoost model achieving an AUROC of 0.865 and the lowest Brier score (0.132).
Conclusions:
- Preoperative models allow for early risk stratification of patients.
- Postoperative models incorporating intraoperative data offer improved accuracy in predicting PPCs.
- A web-based calculator was developed to facilitate clinical implementation and prospective validation of the models.
