Objective To develop a clinical-radiomic nomogram model based on CT plain scan, assess its value in predicting spread through air spaces (STAS) in lung adenocarcinoma, and validate its generalizability in an independent external validation set.

Methods A retrospective cohort study was conducted. Patients with lung adenocarcinoma confirmed by postoperative histopathological examination between October 2018 and June 2023 were enrolled from the Eighth Affiliated Hospital of Southern Medical University (Center Ⅰ) and Chencun Hospital, Affiliated to Shunde Hospital of Southern Medical University (Center Ⅱ). The cohort included 201 patients (98 males and 103 females; age (63.9±11.0) years) from Center Ⅰ and 183 patients (84 males and 99 females; age (62.9±10.8) years) from Center Ⅱ. Using a completely randomized grouping method, the 201 patients from Center Ⅰ were divided into a training set (n=140) and an internal validation set (n=61) in a 7∶3 ratio. The 183 patients from Center Ⅱ served as the external validation set. Radiomic features were screened using least absolute shrinkage and selection operator regression. Features retained with non-zero coefficients were used for regression model fitting, combined into a radiomic signature (Radscore), and a radiomic model was developed. Univariate and multivariate logistic regression analyses were performed to identify independent influencing factors for predicting STAS. A logistic regression machine learning algorithm was employed to construct a combined clinical-radiomic model, and a nomogram was plotted. The generalizability of the nomogram model was tested in the external validation set. Receiver operating characteristic curves and area under the curves (AUC) were used to evaluate the diagnostic performance of the nomogram model. Calibration curves were utilized to verify the agreement between the predicted probability and the actual probability. Decision curve analysis was applied to assess the clinical net benefit of the nomogram model.

Results Univariate and multivariate logistic regression analyses indicated that air bronchogram sign (OR=0.428, 95%CI: 0.256–0.716), vacuole sign (OR=0.155, 95%CI: 0.060–0.399), crescent sign (OR=0.216, 95%CI: 0.079–0.595), and consolidation-to-tumor ratio (OR=13.080, 95%CI: 5.403–31.658) were independent influencing factors for predicting STAS in lung adenocarcinoma (all P<0.05). The AUCs of the nomogram model for predicting STAS were 0.923(95%CI: 0.874–0.971) in the training set, 0.922(95%CI: 0.844–1.000) in the internal validation set, and 0.843(95%CI: 0.787–0.900) in the external validation set (all P<0.05), demonstrating high sensitivity of the nomogram model in predicting STAS. The calibration curves showed good agreement between the predicted and actual probabilities. Decision curve analysis revealed that the nomogram model provided a high clinical net benefit within a threshold probability range of 0–0.8.

Conclusions The clinical-radiomic nomogram model based on CT plain scan can accurately assess STAS in lung adenocarcinoma. The model demonstrated satisfactory validation of generalizability in the external validation set.