Machine Learning-Based CT Radiomics Analysis for Prognostic Prediction in Metastatic Non-Small Cell Lung Cancer Patients With EGFR-T790M Mutation Receiving Third-Generation EGFR-TKI Osimertinib Treatment

Abstract

Background and Purpose As a third-generation EGFR tyrosine kinase inhibitor (TKI), osimertinib is approved for treating advanced non-small cell lung cancer (NSCLC) patients with EGFR-T790M mutation after progression on first- or second-generation EGFR-TKIs such as gefitinib, erlotinib and afatinib. We aim at exploring the feasibility and effectiveness of using radiomic features from chest CT scan to predict the prognosis of metastatic non-small cell lung cancer (NSCLC) patients with EGFR-T790M mutation receiving second-line osimertinib therapy. Methods Contrast-enhanced and unenhanced chest CT images before osimertinib treatment were collected from 201 and 273 metastatic NSCLC patients with EGFR-T790M mutation, respectively. Radiomic features were extracted from the volume of interest. LASSO regression was used to preliminarily evaluate the prognostic values of different radiomic features. We then performed machine learning-based analyses including random forest (RF), support vector machine (SVM), stepwise regression (SR) and LASSO regression with 5-fold cross-validation (CV) to establish the optimal radiomic model for predicting the progression-free survival (PFS) of osimertinib treatment. Finally, a combined clinical-radiomic model was developed and validated using the concordance index (C-index), decision-curve analysis (DCA) and calibration curve analysis. Results Disease progression occurred in 174/273 (63.7%) cases. CT morphological features had no ability in predicting patients’ prognosis in osimertinib treatment. Univariate COX regression followed by LASSO regression analyses identified 23 and 6 radiomic features from the contrast-enhanced and unenhanced CT with prognostic value, respectively. The 23 contrast-enhanced radiomic features were further used to construct radiomic models using different machine learning strategies. Radiomic model built by SR exhibited superior predictive accuracy than RF, SVR or LASSO model (mean C-index of the 5-fold CV: 0.660 vs. 0.560 vs. 0.598 vs. 0.590). Adding the SR radiomic model to the clinical model could remarkably strengthen the C-index of the latter from 0.672 to 0.755. DCA and calibration curve analyses also demonstrated good performance of the combined clinical-radiomic model. Conclusions Radiomic features extracted from the contrast-enhanced chest CT could be used to evaluate metastatic NSCLC patients’ prognosis in osimertinib treatment. Prognostic models combing both radiomic features and clinical factors had a great performance in predicting patients’ outcomes.