Machine learning-based preoperative prediction of spread through air spaces in clinical stage IA non-small cell lung cancer: a single-center study integrating clinical and CT imaging characteristics.

Background: Spread through air spaces (STAS) is a key adverse pathologic feature in clinically early-stage non-small cell lung cancer (NSCLC) and is associated with higher recurrence risk, which can undermine the oncologic safety of sublobar resection. Because STAS is typically confirmed only on postoperative specimens and cannot be reliably determined preoperatively, this study aimed to develop and validate machine learning models integrating preoperative clinical and computed tomography (CT) imaging characteristics to predict STAS in clinical stage IA NSCLC to support surgical procedure selection. Methods: We retrospectively enrolled 445 patients and randomly assigned them to a training cohort (n=312) and a validation cohort (n=133) at a 7:3 ratio. Candidate predictors were selected using minimum redundancy maximum relevance (mRMR) combined with least absolute shrinkage and selection operator (LASSO) regression, and inter-feature correlations were assessed with Spearman rank correlation. We developed and compared seven machine-learning models: Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), K-Nearest Neighbors (KNN), Light Gradient Boosting Machine (LGBM), and naïve Bayes (NB). Model performance was evaluated by the area under the receiver operating characteristic curve (AUC), with AUCs compared using DeLong's test. Confusion-matrix-derived metrics (sensitivity, specificity, accuracy, F1 score, among others) were also reported. Calibration was assessed using calibration curves and the Brier score, and clinical utility was examined via decision curve analysis (DCA). The final model was interpreted with Shapley Additive exPlanations (SHAP) to quantify feature importance. Results: STAS was presented in 26.9% of the training cohort and 30.8% of the validation cohort. mRMR and LASSO yielded highly concordant feature selection results, and nine variables were ultimately retained for model development. Among the seven machine-learning algorithms, XGBoost achieved the best overall performance. The AUC was 0.830 in the training cohort, with a 95% confidence interval (CI) of 0.781-0.878, and 0.780 in the validation cohort, with a 95% CI of 0.702-0.859. Calibration curves demonstrated good agreement between predicted and observed risks, and XGBoost showed the lowest Brier score in the validation cohort (0.171). DCA indicated a favorable net benefit across clinically relevant threshold probabilities. SHAP interpretation highlighted pure-solid nodule, spiculation, nodule volume, pleural retraction, and carcinoembryonic antigen (CEA) as the most influential predictors. Conclusions: The XGBoost model based on preoperative clinical and CT imaging characteristics achieved good discrimination and calibration for predicting STAS in clinical stage IA NSCLC, with potential clinical utility for preoperative risk stratification and surgical decision-making.