Interpretable machine learning for preoperative prediction of ground-glass nodules invasiveness in lung adenocarcinoma: a multicenter study

Accurate, noninvasive prediction of invasiveness in ground-glass nodules (GGNs) is important for surgical planning in lung adenocarcinoma. This multicenter study developed and validated an interpretable machine learning (ML) model based on quantitative and qualitative computed tomography (CT) features. We retrospectively enrolled 860 patients with 1009 GGNs from three centers. Center 1 cases were split into training (n = 609) and internal validation (n = 261) cohorts, while Centers 2 and 3 formed an external validation cohort (n = 139). Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression, and eight ML models were trained with cross-validation. Additional analyses compared the Random Forest (RF) model with consolidation-to-tumor ratio (CTR), assessed pure GGNs (pGGNs), removed visually evident invasive signs, and evaluated overfitting using out-of-bag and bootstrap optimism-corrected estimates. The RF model achieved apparent training, internal validation, and external validation AUCs of 0.955, 0.874, and 0.873, respectively; the RF out-of-bag AUC was 0.835. CTR alone showed lower discrimination than RF in internal and external validation cohorts. Adding CTR to RF did not significantly improve AUC. In pGGN-only analysis, RF achieved AUCs of 0.792 and 0.872 in internal and external validation cohorts, respectively. Removing spiculation and internal vascular signs reduced performance but retained acceptable external discrimination. The RF model showed stable validation performance and interpretable feature contributions for preoperative assessment of GGN invasiveness. Its use should be considered adjunctive, and prospective validation remains necessary before clinical implementation.