Objective To establish and compare three prediction models for pulmonary infection after thoracoscopic surgery in lung cancer patients based on multivariate logistic regression, decision tree, and neural network.
Methods Patients with lung cancer who underwent thoracoscopic lung resection in Zhenjiang Third People’s Hospital, Jurong People’s Hospital, and Danyang Third People’s Hospital from October 1, 2022 to March 31, 2025 were selected. Data on patient demographics, physical examination findings, past medical history, pulmonary function indices, surgical conditions, and tumor conditions of the patients were collected. The patients were divided into a pulmonary infection group and a non-pulmonary infection group based on the occurrence of pulmonary infection. The differences between the two groups were compared through univariate analysis. The variables with statistical significance in the univariate analysis were selected as independent variables for multivariate logistic regression, decision tree, and neural network to construct the prediction models, and the sensitivity, specificity, Youden index, and area under the receiver operating characteristic (ROC) curve (AUC) of the three models were compared.
Results A total of 1 262 patients were included in the study, of whom 230 cases had pulmonary infection, with an incidence of 18.22%. Multivariate logistic regression showed that age (≥60 years), gender, smoking history, history of chronic pulmonary diseases, and hospitalization time (≥10 days) were independent risk factors for postoperative pulmonary infection. The decision tree model obtained 4 explanatory variables: age (≥60 years), serum albumin level (<35 g/L), hospitalization time (≥10 days), and smoking history. The neural network model suggested that the top five risk factors were age (≥60 years), serum albumin level (<35 g/L), history of chronic pulmonary diseases, hospitalization time (≥10 days), and smoking history. The accuracy of the three models was 84.9%, 81.5%, and 86.7%, respectively. The sensitivity was 77.2%, 73.2%, and 75.4%; specificity was 80.3%, 76.8%, and 82.7%; Youden index was 0.575, 0.520, and 0.581; and the AUC was 0.831, 0.796, and 0.857, respectively (all P<0.05).
Conclusions Compared with the multivariate logistic regression model and the decision tree model, the neural network model has better predictive performance for pulmonary infection after thoracoscopic surgery in lung cancer patients.