From risk factors to individualized prediction: development and internal validation of a machine learning model for early identification of postoperative pneumonia after lung cancer surgery

This is the authors' abstract. We don't add key points for this paper.

Background: Postoperative pneumonia (POP) remains one of the most common infectious complications following lung cancer surgery and is associated with substantial morbidity. Although several risk factors for POP have been identified, clinically applicable approaches for early individualized risk stratification remain limited. This study aimed to develop and internally validate machine learning models for early prediction of POP using routinely available perioperative variables.Methods: In this prospective single-center cohort study, 258 consecutive patients undergoing pulmonary resection for non-small cell lung cancer or pulmonary carcinoid tumors were screened, of whom 200 were included in the final study cohort. POP was defined by newly developed postoperative radiographic abnormalities accompanied by clinical signs of infection requiring antibiotic therapy. Machine learning models, including logistic regression (LR), Naïve Bayes (NB), random forest (RF), and XGBoost (XGB), were developed to predict POP using demographic, clinical, laboratory, surgical, and postoperative day 1 radiographic variables. Model performance was assessed using nested leave-one-out cross-validation and quantified by the area under the receiver operating characteristic curve (AUROC). A nomogram was constructed from the best-performing model to facilitate individualized risk estimation. Different decision thresholds were explored to evaluate clinically relevant trade-offs between sensitivity and specificity.Results: POP occurred in 47 of 200 patients (23.5%), while no POP-related mortality was observed. LR demonstrated the best predictive performance (AUROC 0.701±0.070), followed by NB (0.698±0.103), RF (0.656±0.217), and XGB (0.589±0.179). The most influential predictors included advanced age, perioperative changes in C-reactive protein concentration (ΔCRP), elevated preoperative CRP levels, prolonged operative duration, and abnormal postoperative day 1 chest radiographic findings. At a decision threshold of 25%, the model correctly identified 25 of 47 POP cases (53.2%), whereas increasing the threshold to 32.5% reduced false-positive classifications from 42 to 24 at the expense of missing an additional four POP cases.Conclusions: Machine learning models based on routinely available perioperative variables enabled early individualized prediction of POP following lung cancer surgery. LR provided the best balance between predictive performance, stability, and interpretability, allowing the development of an interpretable nomogram for individualized risk estimation. The model may have potential to support risk-adapted postoperative surveillance and more efficient allocation of healthcare resources; however, external validation is required before clinical implementation.