Development of a CT-based integrated model combining clinical features, radiomics, and deep learning to predict STAS in subsolid nodules of stage I lung cancer: a retrospective cohort study.

Background: For stage I lung cancer manifesting as subsolid nodules (SSNs) on computed tomography (CT), precise preoperative estimation of the spread through air spaces (STAS) is crucial for optimizing surgical strategies and enhancing patient prognosis. However, there is currently relatively limited research on the construction of predictive models through multi-feature fusion to enhance prediction performance. This study aimed to develop and internally validate an interpretable multi-modal model for preoperative prediction of STAS in stage I lung cancer with SSNs. Methods: This retrospective, single-center cohort study was conducted with the collection and analysis of relevant data obtained from 1,499 patients with stage I lung cancer showing SSNs on CT imaging and pathologically confirmed by Thoracic Surgery at The First Medical Center of Chinese PLA General Hospital. Radiomics (RAD) and deep learning (DL) features were extracted from the dataset using Python 3.7 and advanced RAD feature extraction techniques, respectively. Furthermore, the cohort was divided into a training cohort (n=1,049) and an independent validation cohort (n=450) at a 7:3 ratio. Four types of predictive models were separated and built by using six machine learning methods. Model performance was assessed via receiver operating characteristic (ROC) curves, decision curve analysis (DCA), and calibration curves. Additionally, the predictions made by the constructed model were interpreted and analyzed by employing the SHapley Additive exPlanations (SHAP). Results: Clinical features such as age, family history of lung cancer, consolidation tumor ratio (CTR), pulmonary nodule density, cyst/cavitation, pleural indentation sign, bronchial changes, neutrophil-to-lymphocyte ratio (NLR), and carbohydrate antigen 125 (CA125) could predict STAS in stage I lung cancer presenting as SSNs. Furthermore, compared to the RAD model [area under the curve (AUC) =0.856, 95% confidence interval (CI): 0.824-0.888 in the validation set], DL model (AUC =0.828, 95% CI: 0.791-0.865 in the validation set), and RAD + DL model (AUC =0.886, 95% CI: 0.850-0.922 in the validation set), clinical-radiomics-deep learning (CRDL) model exhibited superior predictive performance (AUC =0.955, 95% CI: 0.919-0.991 in the validation set). Conclusions: The integrated model combining clinical, RAD, and DL features demonstrates strong predictive performance and improved predictive accuracy for the presence of STAS in stage I lung cancer patients with SSNs, exhibiting potential clinical applicability. Further multi-center prospective validation is required before the model can be applied in clinical practice.