Abstract PO-035: Development of a robust radiomic biomarker of PDL1 expression and patient survival after first-line immunotherapy for non-small cell lung cancer

Abstract

Abstract Pembrolizumab (PEMBRO), a PD1antagonist, is 1st-line therapy for non-small cell lung cancer (NSCLC) without actionable mutations. While up to 70% of PDL1 enriched NSCLC respond to PEMBRO, those that will fail therapy, with resulting worse overall survival, cannot be reliably predicted. We propose a novel radiomic biomarker which integrates with PDL1 expression data to enhance precision in therapy response prediction. NSCLC patients (n=98) treated with PEMBRO at our institution, either alone or in combination with chemotherapy, with available staging computed tomography (CT) imaging were retrospectively identified. 3D tumor CT volumes were labeled by thoracic radiologists using ITK-SNAP software and radiomic features were extracted from segmented tumor volumes using PyRadiomics (Py) (107 features) and the Cancer Imaging Phenomics Toolkit (CaPTk)(102 features). The impact of imaging acquisition heterogeneity was mitigated using a nested ComBat approach, sequentially parsing parameters as batch covariates to harmonize differences in radiomic features. Harmonization reduced the number of features (KS test p < 0.05; Py: 41.67%, CaPTk: 22.4% feature reduction) significantly different by kernel resolution. Six principal components (PCs) capturing 85% of the variance from the retained robust radiomic features of each software package were extracted and combined into a prognostic model (number of predictors capped at six (deaths = 59) to avoid overfitting). The concordance index for overall survival prediction was computed using a 5-fold cross-validated (CV) Cox proportional hazards model (200 iterations). Results for the ComBat feature PCs model are (C-statistic, 95% CI): CaPTk (0.58, [0.52, 0.62]), Py (0.59, [0.53,0.63]) For patients with available PDL1 expression data (n=67, deaths= 40), we used three radiomic PCs to predict survival for tumors with ≥ 50% PDL1 expression. Results for ComBat feature PCs are (5-fold CV AUC results): CaPTk (0.68), Py (0.71). We also combined the three radiomic PCs with a binary variable indicating the presence of >50% PDL1 expression into a model. The complete dataset was used to evaluate the models’ ability to separate patients above versus below the median prognostic score in a Kaplan Meier curve (for all cases, log-rank test p values < 0.05). The five-fold cross-validated ComBat PCs model had the highest accuracy in the prediction of overall survival among all the tested models. The ComBat PCs model performance is ((5-fold cross-validated C-statistic, 95% CI): Captk (0.60, [0.51, 0.64]), Py model (0.61, [0.52,0.65])). We demonstrated that both toolkits capture similar information from the tumor imaging volume with comparable prediction modeling results. ComBat harmonizes differences in image acquisition parameters and allows the generation of a radiomic model leveraging heterogeneous imaging data typical of the standard of care CT. Combining radiomic feature analysis with tumor PDL1 expression improves NSCLC survival prediction for therapy with PEMBRO. Citation Format: Apurva Singh, Hannah Horng, Joanna Weeks, Michelle Hershman, Leonid Roshkovan, Sharyn Katz, Despina Kontos. Development of a robust radiomic biomarker of PDL1 expression and patient survival after first-line immunotherapy for non-small cell lung cancer [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-035.