Abstract
BACKGROUND: This study aimed to preoperatively predict spread through air spaces (STAS) in stage I lung adenocarcinoma presenting as part-solid and solid nodules by leveraging clinical features and machine learning models, thereby guiding surgical decision-making and enhancing patient counseling. METHODS: A total of 473 patients were retrospectively enrolled, including 353 from our center and 120 from an validation cohort. Predictive features were selected using maximum relevance minimum redundancy (mRMR) and least absolute shrinkage and selection operator (LASSO) algorithms. Seven machine learning models-logistic regression, random forest, support vector machine (SVM), extreme gradient boosting (XGBoost), adaptive boosting (AdaBoost), light gradient boosting machine (LightGBM), and category boosting (CatBoost)-were developed and evaluated using receiver operating characteristic curves, calibration plots, and decision curve analysis (DCA). Feature importance was assessed using Shapley Additive Explanations (SHAP). A web-based nomogram was constructed for clinical application. RESULT: STAS was present in 44.76% of the training set and 50.83% of the validation cohort. Seven predictors were selected to construct the predictive models. The XGBoost model demonstrated superior performance with an AUC of 0.889 (95% CI, 0.852-0.926) in training and 0.856 (95% CI, 0.789-0.928) in validation. The calibration curves in training and validation set exhibited good agreement between the predictions and actual observations. The Decision Curve Analyses (DCA) provide significant clinical utility. SHAP analysis identified the most important predictors for STAS as CEA, vascular convergence, proGRP, age, AFP, smoking history, and CTR. CONCLUSION: The XGBoost model provides robust preoperative prediction of STAS and may assist clinicians in optimizing surgical strategies for patients with stage I lung adenocarcinoma.