Machine learning prediction of postoperative pulmonary infection in patients who underwent thoracoscopic lung cancer resection: a retrospective case-control study.

Abstract

Background: Accurate identification of patients at high risk of pulmonary infection after thoracoscopic lung cancer resection is important for timely and targeted preventive measures. Methods for determining the risk of pulmonary infection after thoracoscopic lung cancer resection have not been well studied.

Methods: This study was a retrospective case-control research project. The information of 3219 hospitalised patients who underwent thoracoscopic lung cancer resection between January 2019 and December 2023 was obtained from the hospital electronic medical record system. 26 clinical characteristics were obtained from medical and nursing records. The variables were screened using the least absolute contraction and selection operator (LASSO) regression, and the risk prediction models for pulmonary infection after thoracoscopic lung cancer resection was constructed using the following 5 machine learning algorithms: logistic regression model (LR), artificial neural network (ANN), support vector machine (SVM), random forest (RF) and eXtreme gradient boosting (XGB). The model was evaluated using the following metrics: the area under the receiver operator characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and F1 score (F1). Shapley additive explanation (SHAP) was used to interpret the machine learning models.

Results: There were 3219 enrolled patients, 2203 (70%) of whom were assigned to the training cohort and 966 (30%) of whom were assigned to the validation cohort. The AUC range of the five models was 0.883-0.951. The XGB model outperformed the others, with an AUC of 0.951 (95% confidence interval: 0.943-0.964), accuracy of 0.902 (95% confidence interval: 0.886-0.913), sensitivity of 0.927, specificity of 0.864, positive predictive value of 0.898, negative predictive value of 0.824, precision of 0.908, recall of 0.872 and F1 score of 0.815 in the validation group. The model's prediction performance in the 45-65 age group was the best. The AUC of the logistic regression model was 0.948 (95% confidence interval: 0.931-0.957). We transformed the logistic regression model into a nomogram to help clinicians visualise the model and make them more likely to use it to identify the risk of pulmonary infection after thoracoscopic surgery in lung cancer patients.

Conclusions: The establishment of a risk prediction model based on machine learning can help clinical nursing staff identify high-risk patients for pulmonary infection after thoracoscopic lung cancer resection.

Clinical trial number: Not applicable.