Objective To explore the value of a combined clinical radiomics model based on machine-learning method in the preoperative prediction of spread through air spaces (STAS) in stage Ⅰ lung cancer.

Methods The clinical, imaging, and pathological data of 165 patients with stage Ⅰ lung cancer confirmed by histopathology were retrospectively analyzed in Huainan Yangguang xinkang Hospital from November 2020 to November 2023. The patients included 91 males and 74 females, aged (61.6±11.2) years, ranging from 28 to 88 years. All patients were divided into STAS-negative and STAS-positive groups according to the results of postoperative histopathology. The target areas of lung cancer in preoperative CT images of 165 patients were mapped, and radiomic features were extracted using ITK-SNAP software. The patients were divided into the training group (115 cases) and the verification group (50 cases) according to a 7∶3 ratio through complete-randomization method. Intra- and inter-group correlation coefficients, univariate analysis, Spearman correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression analysis were used to reduce the dimensionality of the selected features. The feature weight coefficients were calculated, and linear combinations were used to calculate the Radscore. The radiomics model was established according to Logistic method. Independent risk factors for STAS of stage Ⅰ lung cancer were screened by univariate and multivariate Logistic regression analysis. According to Akaechi information criteria (AIC), a clinical model with minimal AIC was obtained by Logistic regression analysis. A combined clinical radiomic model for predicting STAS of stage Ⅰ lung cancer was constructed by combining clinical model and radiomic model, and a nomogram was drawn to visualize the model. A receiver operating characteristic curve was drawn to evaluate the diagnostic efficiency of the model.

Results A total of 165 patients with stage Ⅰ lung cancer were included, specifically 85 in the STAS-negative group and 80 in the STAS-positive group. The area under the curve (AUC) of the radiomics model was 0.895 (95%CI: 0.826–0.950) and 0.814 (95%CI: 0.680–0.929) in the training and verification groups, respectively. The sensitivity was 0.893 and 0.875, and the specificity was 0.797 and 0.731, respectively. Univariate and multivariate Logistic regression analysis showed that CT value and burr sign were independent risk factors for predicting STAS in stage Ⅰ lung cancer. The AUC of the clinical model in the training and verification groups was 0.849 (95%CI: 0.772–0.914) and 0.822 (95%CI: 0.700–0.923), respectively. The sensitivity was 0.786 and 0.583, and the specificity was 0.814 and 0.962, respectively. The AUC of the combined clinical radiomics model was 0.943 (95%CI: 0.894–0.982) and 0.880(95%CI: 0.769–0.971) in the training and verification groups, respectively. The sensitivity was 0.964 and 0.750, and the specificity was 0.847 and 0.962, respectively.

Conclusions The combined clinical radiomics model developed based on machine-learning method can predict STAS of stage Ⅰ lung cancer before surgery. This model can be expected to assist in the selection of accurate diagnosis, treatment, and management strategies for stage Ⅰ lung cancer.