39P OLIMPIA dataset: Radiomics to predict outcomes in EGFR-mutant non-small cell lung cancer

Non-small cell lung cancer (NSCLC) with a detectable EGFR mutation represents up to 50% of cases depending on the geographic area. There are currently 5 approved tyrosine kinase inhibitors (TKI), including first, second, and third generations. Although osimertinib is currently the standard of care, cost-effectiveness could be improved by identifying patients who will present longer progression-free survival with more accessible treatments. We propose a non-invasive approach to identify risk of progression based on imaging biomarkers (radiomics) from the pre-treatment CT scan. We included 60 histologically proven cases of NSCLC with confirmed EGFR mutations. We evaluated progression at 12-months after starting TKI therapy: 32 patients showed disease progression and 28 did not. We manually segmented lesions in pre-treatment CT scans and extracted radiomic features. We applied machine learning techniques for dimensionality reduction and classification of patient outcomes. We compared the predictive power of this radiomics model to a logistic regression model trained solely on clinical data: gender, age, and smoking status. We used cross validation to calculate diagnostic metrics, reported as mean ± std. The final radiomics model is an ensemble of 12 classifiers trained with 20 features from principal component analysis. For the prediction of disease progression, the radiomics model showed a sensitivity of 0.84 ± 0.12, specificity 0.70 ± 0.41, positive predictive value 0.79 ± 0.23 and negative predictive value 0.67 ± 0.39. Comparing radiomics to the regression with clinical data, they showed respectively an area under the ROC curve of 0.82 ± 0.15 vs. 0.39 ± 0.11, and an area under the precision-recall curve of 0.82 ± 0.18 vs. 0.52 ± 0.13. This suggests that the radiomics model has stronger predictive power than basic clinical data. Up to date, there is no publicly available dataset to target this issue. No previous work has addressed this problem in Latin American populations. Our results are presented as a baseline and we plan to release publicly the current dataset to motivate further studies on this topic. These results suggest that radiomics is a promising approach to predict progression in patients treated with TKI therapy.