Abstract
BACKGROUND: Preoperative detection of occult nodal metastasis (ONM) is essential for treatment planning and prognostic evaluation in lung adenocarcinoma (LUAD). This study aimed to develop and validate radiomics models capable of predicting ONM in patients with stage cT1a-bN0M0 LUAD. METHODS: A total of 1,672 patients from six hospitals were enrolled and stratified into training (n=687), test (n=297) and external validation (n=688) sets. Predictive models including generalized linear model (GLM), support vector machine (SVM), random forest (RF), gradient boosting machine (GBM), and the combined clinical-radiomics (Clinic-Rad) model were constructed. Diagnostic performance was quantified via the area under receiver operating characteristic curve (AUC), with the De Long test used for comparisons. The Mantel test was employed to assess the correlations between radiomics features and pathological/genetic characteristics. RESULTS: In external validation, the Clinic-Rad model integrating clinical predictors and radiomics score demonstrated superior diagnostic efficacy (AUC 0.813±0.019) compared to GLM (0.790±0.021), SVM (0.761±0.023), RF (0.708±0.026), GBM (0.769±0.022) (all P values <0.001). However, no significant intermodel differences were observed in test set, with the Clinic-Rad model achieving an AUC of 0.834±0.023, and GLM, SVM, RF, and GBM yielding AUCs of 0.827±0.024, 0.829±0.025, 0.838±0.023, and 0.826±0.024, respectively (all P values >0.05). The Clinic-Rad model exhibited a pooled sensitivity of 75.8-77.2%, a specificity of 72.0-72.7%, and an accuracy of 72.7-74.4%, with pooled AUC values of 0.802-0.820 and 0.797-0.917 for the solid and subsolid LUAD, respectively. Furthermore, radiomics models outperformed clinical predictors comprising solid-component diameter (AUC: 0.669-0.678), consolidation-to-tumor ratio (CTR) (0.542-0.600), carcinoembryonic antigen (CEA) level (0.571-0.613), and their combination (0.683-0.724) (all P values <0.001). The Mantel test indicated correlations between radiomics signatures and EGFR, ALK, ROS1, and RET expression, as well as histopathological markers of ONM. CONCLUSIONS: Radiomics-based models demonstrate clinical utility in predicting ONM in patients with stage cT1a-bN0M0 LUAD, with the integrated Clinic-Rad model providing superior diagnostic performance.