115PEvaluation of a predictive radiomics signature for response to immune checkpoint inhibitors (ICIs)

Background: Radiomics (RAD) uses advanced image processing techniques to extract a large set of quantitative texture and geometric features from tumor regions of interest, and subject these to a supervised machine learning protocol to train a classifier, which we exploit to develop a predictive signature of response to ICIs. We previously developed a lesion-based predictive RAD classifier of response for recurrent/metastatic squamous cell carcinoma of the head and neck (RM SCCHN) pts to ICIs based on RAD features extracted from their CT images (Prawira, ESMO 2016). Methods: INSPIRE (NCT02644369) is an investigator-initiated phase II study evaluating biomarkers for pembrolizumab (anti-PD1 monoclonal antibody) in multiple cohorts of pts with advanced solid tumors. The primary endpoint of this project is to validate the previously developed RAD classifier from RM SCCHN pts, with pts from INSPIRE. Texture feature algorithm generation and accuracy determination were as previously described. Cross validation accuracy values were generated for combinations of 3 parameters: fraction, cost, and gamma, yielding a 3 dimensional (3D) accuracy space. Results: Eighty lesions from 23 pts were available for analysis: median age 59, 22% males. Best response: 12 progressive disease, 3 partial response, 8 stable disease (median duration 18 weeks). Primary site: SCCHN/2, triple negative breast cancer/4, high-grade serous ovarian cancer/11, malignant melanoma/2, other advanced cancers/4. Twentyseven lesions were excluded as RECIST 1.1 responses were not yet available. Fiftythree target lesions were contoured. Per lesion RECIST 1.1 radiological outcome: 17 R, 36 NR. Cross validation in the 3D space yielded a set of ROC curves with an accuracy of 71.4% (AUC 0.41, p = 0.7) with 11.2% sensitivity and 99.9% specificity, where specificity corresponds to the proportion of NR tumors classified correctly, and sensitivity to the proportion of R tumors classified correctly. Conclusions: Heterogeneous histologies and low pt numbers may account for the negative result in this study, suggesting that RAD may be histology-specific. Further validation in a large independent cohort of RM SCCHN pts treated with pembrolizumab is planned. Clinical trial identification: INSPIRE (NCT02644369) Legal entity responsible for the study: Princess Margaret Cancer Centre, Drug Development Program Funding: Merck Disclosure: All authors have declared no conflicts of interest.