A deep-learning model using chest computed tomography images to predict epidermal growth factor receptor (EGFR) T790M mutation after first-line treatment with EGFR-tyrosine kinase inhibitor in patients with non-small cell lung cancer

To predict the epidermal growth factor receptor (EGFR) T790M status of patients with advanced non-small cell lung cancer (NSCLC) following the first-line first-/second-generation EGFR-tyrosine kinase inhibitor (EGFR-TKI) therapy, the related clinical features and chest computed tomography (CT) images of patients with advanced NSCLC in our hospital were retrospectively collected. All patients who met the criteria were randomly divided into training and validation cohorts. Then, a clinical model with the filtered clinical characteristics and a deep-learning model (DLM) were constructed. The area under the curve (AUC), specificity, sensitivity, accuracy, and decision curve analysis were used to evaluate model performance. In total, 66 patients met the inclusion criteria of the study (training cohort, n = 40; validation cohort, n = 26). EGFR19del and the use of gefitinib were significant (P < 0.05), and then, the clinical model was established using multivariate logistic regression analysis. The AUCs of the clinical model were 0.862 (95% confidence interval [CI], 0.570 – 0.966) and 0.755 (0.566 – 0.943) in the training and validation cohorts, respectively. The AUCs of the DLM from the chest CT image analysis were 0.839 (95% CI, 0.708 – 0.970) and 0.842 (0.680 – 1.000) in the training and validation cohorts, respectively. In the validation cohort, the DLM and clinical model exhibited an accuracy of 0.7308 and 0.5000, specificity of 0.6667 and 0.2000, positive probability values of 0.6429 and 0.4545, and negative probability values of 0.8333 and 0.7500, respectively. The DLM was developed using chest CT images to predict the EGFR T790M status following the first-line first- and second-generation EGFR-TKI treatment of advanced EGFR-positive NSCLC.