Machine learning-based radiomics strategy for prediction of acquired EGFR T790M mutation following treatment with EGFR-TKI in NSCLC

Abstract

Abstract Background: Epidermal growth factor receptor (EGFR)-tyrosine kinase inhibitors (TKIs) are considered the standard first-line therapy for treatment-naive patients with advanced/metastatic non-small cell lung cancer (NSCLC) who have sensitizing EGFR mutations. Currently, there are three generations of EGFR-TKIs available for the treatment of EGFR mutation-positive NSCLC. However, the optimal sequence of administering these drugs to maximize the duration of EGFR signaling inhibition remains uncertain. It is crucial to identify patients at the time of diagnosis who are likely to acquire a Thr790Met (T790M) after treatment with a first- or second-generation EGFR-TKI. Purpose: To develop and validate a machine learning (ML)-based radiomics approach to predict acquired EGFR-T790M mutation following treatment with a first- or second-generation EGFR-TKI in patients with NSCLC harboring EGFR mutations. Methods: A total of 274 advanced NSCLC patients with sensitive EGFR mutation and treatment with first- or second-generation EGFR-TKI were retrospectively collected. Tumor regions of interest were segmented and radiomic features were extracted. The least absolute shrinkage and selection operator (LASSO) algorithm was used for feature selection, 7 ML approaches were used to build radiomics models. The receiver operating characteristic (ROC) curve was established to evaluate the discriminating performance of the radiomic models and combined ones (radiomic features and relevant clinical factors). A nomogram was constructed using the most important clinical factors and the radiomics signatures. Decision curve analysis (DCA) and calibration curve analyses were applied to evaluate clinical utility. Results: In 274 patients, 128 cases harbor EGFR-T790M mutation. One hundred and ninety-two cases were selected into the training cohort and 82 into the independent validation cohort. A total of 1316 radiomics features were extracted and 10 radiomics features most relevant to EGFR-T790M mutation were selected to construct models. In terms of predicting EGFR T790M mutation, the model derived from radiomics features had an area under the receiver operating characteristic (AUC), sensitivity, specificity, and accuracy of 0.80 (95% confidence interval [CI]: 0.79–0.81), 0.85 (0.81–0.89), 0.70 (0.65–0.74), and 0.75 (0.71–0.78), respectively. In addition, the AUC, sensitivity, specificity, and accuracy of the combined model for discriminating EGFR mutation were 0.86 (95% CI: 0.85–0.88), 0.78 (0.72–0.84), 0.76 (0.67–0.85), and 0.77(0.73–0.82), respectively. The DCA and calibration curve analyses confirmed potential clinical usefulness of our nomogram. Conclusions: ML-based radiomics model can identify EGFR-T790M mutation in advanced NSCLC patients with EGFR mutations after treatment with a first- or second-generation EGFR- TKI, which can be conveniently used to discriminate patients with acquired EGFR-T790M mutation at diagnosis from those without. This convenient and noninvasive method may aid in targeted treatment planning for NSCLC patients bearing EGFR mutations.