Machine learning models based on artificial intelligence-derived CT quantitative parameters and CT semantic features for predicting the invasive risk of subsolid pulmonary nodules: a dual-center study

Purpose To develop and validate machine learning (ML) models integrating artificial intelligence (AI)-derived quantitative CT parameters and CT semantic features for predicting the invasiveness risk of subsolid pulmonary nodules (SSNs). Materials and methods This retrospective dual-center study enrolled 833 patients with surgically resected, pathologically confirmed adenocarcinoma-spectrum SSNs (training, n=586; external validation, n=247). Low risk was defined as atypical adenomatous hyperplasia (AAH), adenocarcinoma in situ (AIS), or minimally invasive adenocarcinoma (MIA); high risk as invasive adenocarcinoma (IAC). The least absolute shrinkage and selection operator (LASSO) selected features from 19 AI parameters, 7 semantic features, and 3 clinical characteristics. Seven ML models were built and validated using ten-fold cross-validation and external validation. SHapley Additive exPlanations (SHAP) provided interpretability. Results LASSO identified nine predictors (margin, spiculation, lobulation, pleural indentation, vascular convergence, long diameter, solid component ratio, skewness, entropy). In internal cross-validation, all seven models showed comparable performance (mean AUC 0.935–0.940). In external validation, random forest (RF) showed the numerically highest AUC (0.959, 95% CI: 0.938–0.980), although pairwise DeLong tests with Bonferroni correction showed no statistically significant differences among models. SHAP analysis identified long diameter, skewness, spiculation, entropy, and solid component ratio as the top five predictors contributing to invasive risk. An online tool based on RF was deployed at https://subsolid-nodule.shinyapps.io/data/ for real-time individual risk calculation. Conclusion Ensemble learning models, particularly RF, integrating AI-derived CT quantitative parameters and CT semantic features accurately predict SSN invasive risk. This non-invasive tool may provide additional information for preoperative risk stratification and multidisciplinary assessment.