Developing a multimodal prediction model for postoperative pulmonary complications in thoracic surgical patients: a retrospective cohort study

To construct and validate a predictive model for postoperative pulmonary complications (PPCs) in patients undergoing thoracic surgery by integrating general clinical data, intraoperative ventilation parameters and preoperative computed tomography (CT) imaging features, screen out the independent risk factors for PPCs, compare the predictive efficacy of different variable combinations and machine learning algorithms, and provide a scientific tool for clinical precise preoperative risk stratification and individualized intervention. A total of 2000 patients who underwent thoracotomy or video-assisted thoracic surgery in the operating room from January 2017 to December 2019 were retrospectively enrolled, and multi-dimensional data were collected based on the hospital information system (HIS), anesthesia information system and picture archiving and communication system (PACS). Lasso regression was used to screen the core risk factors, and four predictive models were constructed using logistic regression: Clinical Model (CM), Clinical-CT Model (CCTM), Clinical-Ventilation Model (CVM), and Full-Variable Model (FVM). To further optimize predictive performance, the FVM was specifically trained and compared using four machine learning algorithms, namely logistic regression (LR), support vector machine (SVM), Adaptive Boosting (AdaBoost), and eXtreme Gradient Boosting (XGBoost). The model performance was comprehensively evaluated by the area under the receiver operating characteristic curve (AUC), Hosmer–Lemeshow test, and other indicators. The incidence of PPCs in the 2000 patients was 20.9% (418/2000). Lasso regression identified nine core predictive features including ASA physical status classification, body mass index (BMI) and inspired oxygen concentration. Comparison of the performance of the four models showed that FVM had optimal predictive efficacy (AUC=0.816, 95%CI 0.782~0.848). The AUC of the XGBoost algorithm for FVM in the validation set was 0.822 (95%CI 0.782~0.849), which was significantly superior to the models constructed by the other three algorithms, demonstrating strong predictive ability. The multimodal model integrating clinical data, CT imaging characteristics, and intraoperative ventilation parameters showed promising performance for predicting PPCs. Further external validation is required before its clinical implementation.