Joint model based on intratumoral and peritumoral computed tomography radiomics integrated with clinical features for predicting the spread through air spaces in lung adenocarcinoma: a multicenter study.

Background: The presence of spread through air spaces (STAS)-positive tumors is associated with an elevated risk of local and distant recurrence following sublobar resection, a risk that is not observed in patients undergoing lobectomy. Accurate assessment of STAS necessitates postoperative pathological diagnosis. The objective of this study was to develop a joint machine learning model that combines intratumoral and peritumoral computed tomography (CT) radiomics with clinical features for the preoperative prediction of STAS in lung adenocarcinoma, thereby facilitating optimal selection of surgical strategy and treatment plan to mitigate the likelihood of secondary surgery for patients. Methods: This retrospective study collected data from patients with lung adenocarcinoma from three centers. The dataset from one center was randomly divided into a training group and an internal validation group at a ratio of 7:3, while the datasets from the two other centers were used as the external testing cohort. The random forest machine learning algorithm was employed to construct independent radiomics and clinical feature models. Subsequently, logistic regression was applied to develop a mixed model integrating intratumoral, peritumoral, and clinical features. The performance of the model was evaluated via area under the receiver operating characteristic curve (AUC) analysis, and the results were visualized through nomogram representation. Additionally, calibration curves and decision curves were employed to assess both the goodness of fit and clinical utility of the model. Results: The results revealed that vacuole sign (P<0.001), crescent sign (P=0.002), gender (P=0.021), air bronchogram (P=0.001), spiculation (P=0.032), and consolidation tumor ratio (CTR) (P=0.005) were significantly correlated with STAS status. The joint model integrating intratumoral, peritumoral, and clinical features demonstrated exceptional performance in predicting STAS in lung adenocarcinoma. In the training cohort, the model achieved an AUC of 0.985 [95% confidence interval (CI): 0.9697-1.000], with a sensitivity of 93.8% and a specificity of 93.7%. The internal validation cohort exhibited an AUC of 0.988 (95% CI: 0.9703-1.000), sensitivity of 90.9%, and specificity of 95.7%. In the external test cohort, the model maintained robust predictive performance with an AUC of 0.851 (95% CI: 0.7952-0.9073), a sensitivity of 84.6%, and a specificity of 74.3%. The model outperformed individual clinical and radiomics models, as well as other common machine learning classifiers such as extreme gradient boosting (XGBoost), support vector machine (SVM), and multilayer perceptron (MLP) in terms of AUC and overall accuracy. The nomogram visualization of the joint model provides an intuitive tool for clinicians to assess the risk of STAS preoperatively, facilitating the formulation of precise treatment strategies. Conclusions: The joint model developed in this study, integrating intratumoral, peritumoral, and clinical features, serves as a robust tool for predicting the occurrence of STAS in patients with lung adenocarcinoma. This model can aid in the formulation of precise and appropriate surgical strategies.