Development of a machine learning model for preoperative prediction of spread through air spaces in resectable non-small cell lung cancer: A single-center retrospective study.

Spread through air spaces (STAS) is a pathological feature associated with poor prognosis in non-small cell lung cancer (NSCLC). However, its diagnosis currently depends exclusively on postoperative histopathological examination, limiting its utility for preoperative surgical planning. The present study aimed to develop an interpretable machine learning (ML) model using preoperative clinical and semantic CT features to predict STAS in surgically resectable NSCLC. The present study retrospectively analyzed 584 patients with pathologically confirmed NSCLC who underwent surgical resection. A total of five ML algorithms were developed using routinely available preoperative data and evaluated using repeated 5-fold cross-validation to ensure model robustness and mitigate overfitting. The optimal model was selected based on area under receiver operating characteristic curve (AUC). Feature importance was assessed using SHapley Additive exPlanations (SHAP) analysis for interpretability. Among the five models, eXtreme Gradient Boosting (XGBoost) demonstrated the highest predictive performance (mean cross-validated AUC=0.868 on training set; AUC=0.764 on test set). SHAP analysis identified nodule type, lobulation and smoking history as the most influential features associated with STAS. In conclusion, the present study developed a clinically interpretable XGBoost model capable of predicting STAS using readily accessible preoperative features. This model holds promise as a decision-support tool to potentially guide personalized surgical strategies in NSCLC in the future.