329 - Metrics Derived from Architecture of Tumor-Infiltrating Lymphocytes are Associated with Overall Survival in HPV-Positive Oropharyngeal Squamous Cell Carcinoma Patients: Results from NRG/RTOG 0129 and

G. Corredor¹, J. Harris², C. Koyuncu¹, T. Pathak¹, K. Pandav¹, P. Toro³, K. Yang⁴, F. Faraji⁵, P. Castro⁶, V. C. Sandulache⁶, S. A. Koyfman⁷, F. Nguyen⁸, A. S. Garden⁹, J. J. Caudell¹⁰, C. U. Jones¹¹, D. I. Rosenthal⁹, Q. T. Le¹², J. S. Lewis¹³, and A. Madabhushi¹; ¹Emory University, Atlanta, GA, ²American College of Radiology, Philadelphia, PA, ³Cleveland Clinic, Cleveland, OH, ⁴Department of Radiation Oncology, Taussig Cancer Center, Cleveland Clinic, Cleveland, OH, ⁵University of California, San Diego, San Diego, CA, ⁶Baylor College of Medicine, Houston, TX, ⁷Department of Radiation Oncology, Taussig Cancer Institute, Cleveland Clinic, Cleveland, OH, ⁸Centre Hospitalo-Universitaire de Montréal, Montreal, QC, Canada, ⁹The University of Texas MD Anderson Cancer Center, Houston, TX, ¹⁰H. Lee Moffitt Cancer Center and Research Institute, Department of Radiation Oncology, Tampa, FL, ¹¹Sutter Medical Center Sacramento, Roseville, CA, ¹²Stanford University, Stanford, CA, ¹³Mayo Clinic, Phoenix, AZ

Purpose/Objective(s): Subsets of patients (pts) with HPV-positive (HPV+) oropharyngeal squamous cell carcinoma (OPSCC) treated with standard therapies experience poor survival. Thus, there is a need to develop biomarkers to identify higher risk pts. Therefore, we employed image processing and machine learning to quantify the spatial arrangement of tumor-infiltrating lymphocytes (TILs) in tumor hematoxylin and eosin (H&E)-stained slide images for risk stratification in pts with HPV+ OPSCC. The classifier was modeled and tested on retrospective, institutional cohorts and blindly validated on pts from clinical trials RTOG 0129 and 0522. Materials/

Methods: The modeling set (D1) consisted of 491 retrospective pts with OPSCC from 4 independent institutions. The validation set included 466 retrospective pts from 2 new institutions (D2) and 376 pts from the clinical trials RTOG 0129 and 0522 (D3). Computer algorithms were used to identify 2 types of nuclei (TILs & non-TILs) on digitized H&E tumor samples and built clusters for each nucleus type based on proximity. 2,800 metrics related to density, intersection, and neighborhood were computed. The top 12 metrics, determined by the least absolute shrinkage and selection operator, were used to train two Cox regression models, one for predicting risk of overall survival (OS) and another for risk of disease-free survival (DFS). Considering the expected disease specific failure rates, the 15^th percentile risk score on D1 was used as a threshold for stratifying pts on D2 and D3 as either high or low risk. NRG Oncology biostatistics conducted a blinded evaluation of risk scores and labels for cohort D3.
Results: Pts classified as high risk by the OS model in D2 and D3 had significantly worse prognosis compared to their low-risk counterparts with hazard ratios (HRs) of 2.62 (95% confidence interval [CI]: 1.48-4.62, p<0.001, C-index=0.68) and 1.43 (95% CI: 1.02-2.01, p=0.04, C-index=0.55), respectively. The DFS model was prognostic in D2 with HR=2.43 (95% CI: 1.27-4.64, p=0.01, C-index=0.69) but not in D3 with HR=1.13, (95% CI: 0.77-1.66, p=0.53, C-index=0.50).
Conclusion: A computerized image analysis model based on measurements of spatial arrangement of TILs & non-TILs from primary tumor digitized H&E slides, was found to be prognostic for OS in pts with HPV+ OPSCC. This approach has the potential to help in identifying OPSCC pts at a greater risk of poor survival and to allow for treatment strategies that are more specifically tailored to them.