Abstract
BACKGROUND: Early detection of epidermal growth factor receptor (EGFR) is critical for guiding therapeutic decisions in non-small-cell lung cancer (NSCLC). The study aims to develop a predictive model for EGFR mutations with multicohort data. METHODS: The study enrolled 254 NSCLC patients of four cohorts: the Affiliated Hospital of Qingdao University (AHQU, n = 54), the Second Affiliated Hospital of Soochow University (SAHSU, n = 78), TCGA-NSCLC (n = 91), and CPTAC-NSCLC (n = 31). Radiomic features were extracted using the LIFEx software. The least absolute shrinkage and selection operator (LASSO) algorithm was utilized to select predictive features of CT radiomics, clinical data, and RNA sequencing, which were evaluated using receiver operating characteristic (ROC) curves. A nomogram was developed by integrating predictive features. Biological functions were analyzed utilizing RNA sequencing data. RESULTS: Eight radiomic features, four clinical features, and seven genomic features were selected to construct distinct signatures. Through internal 5-fold cross-validation, the first two signatures demonstrated notable discrimination capabilities for distinguishing between mutated and wild-type EGFR, resulting in area under the curve (AUC) values of 0.79 (± 0.08) and 0.74 (± 0.06), respectively. The combination of clinical variables and radiomics signature resulted in an increased AUC of 0.84 (± 0.01). This combined model was named TRANS, representing TTF-1, radiomic signature, AE1/AE3, NapsinA, and stage, which uses radiomics and routine immunohistochemistry markers as inputs. High-risk TRANS was observed to be associated with poor overall survival, and showed relationships with high T cell infiltration and response to PD-1 immunotherapy. CONCLUSIONS: The TRANS model demonstrated favorable ability in predicting EGFR mutation status in NSCLC, providing a valuable approach for optimizing therapeutic strategies in clinical practice.