Abstract
BACKGROUND: An integrated model combining clinical variables, radiomic features, and deep learning was developed to predict EGFR mutation status in patients with lung adenocarcinoma based on pretreatment (18)F-FDG PET/CT imaging. METHODS: In this retrospective study, data from 218 patients-including PET/CT images, EGFR mutation status, and clinical characteristics-were analyzed. Three predictive models were constructed: a clinical model (C), a clinical-radiomics model (CR), and a clinical-radiomics-deep learning model (CRD). RESULTS: The CRD model integrated screened clinical features, as well as ConvNext-based deep learning scores and radiomic scores selected via LASSO regression. It exhibited significantly superior predictive performance to the C model (AUC = 0.599; DeLong test: Z = -3.522, p < 0.001, corrected p = 0.001) and the CR model (AUC = 0.739; DeLong test: Z = -2.197, p = 0.028, corrected p = 0.028), with an AUC of 0.821 for the CRD model. Calibration curves and decision curve analysis confirmed its robustness and potential clinical benefit. A nomogram based on the CRD model was established, enabling individualized risk prediction of EGFR mutation. CONCLUSIONS: This study highlights the potential of integrating clinical, radiomic, and deep learning features as a noninvasive approach for accurately predicting EGFR mutation status in lung adenocarcinoma.