Abstract 1174P
Background
Although osimertinib is currently the standard of care, cost-effectiveness could be improved by identifying a subset of patients who will present longer progression-free survival (PFS) with other more accessible treatments. We propose a non-invasive approach based on radiomics and machine learning to identify these patients' risk of progression.
Methods
We included histologically proven cases of NSCLC with confirmed EGFR mutations, from patients who started TKI-therapy between January 2012 and January 2020 and had a 12-months follow-up. The primary endpoint was PFS after 12 months of treatment, obtained from RECIST assessment in CT or from death cause. We performed manual volumetric segmentation of lung lesions in CT scans acquired before EGFR-TKI administration and extracted 70 radiomic features using three filter types, obtaining a total of 182 features per patient. We evaluated eight machine learning algorithms with several hyperparameters configurations using leave-one-out cross-validation (LOOCV).
Results
A total of 44 patients with EGFR variant–positive NSCLC receiving EGFR-TKI therapy were included, of which 19 cases (44%) showed progression at 12 months (age at treatment start: 71 (14) years; 68% females; 16 on erlotinib, 2 on afatinib, 1 on osimertinib) and 25 (66%) presented PFS of more than 12 months (age at treatment start: 74 (9) years; 76% females; 16 on erlotinib, 4 on afatinib, 4 on osimertinib, 1 on gefitinib). Higher LOOCV accuracy was 0.71, obtained using a decision tree that showed a sensitivity of 88% in detecting PFS patients and a negative predictive value of 75% (22 true PFS, 9 true progressions). The selected decision tree used three radiomic features: median intensity in the unfiltered image, minimum intensity in the laplacian-filtered image (sigma=1) and difference entropy in the laplacian-filtered (sigma=3).
Conclusions
These preliminary results are a promising first step in adopting radiomic signatures as risk factors for treatment choice in NSCLC patients. A larger dataset is needed to increase the significance of results, and this will be achieved by following the methodology pipeline developed for this work.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.