177 - Development and Validation of a Digital Pathology-Based, Multimodal Artificial Intelligence (MMAI) Biomarker in Patients with Head and Neck Cancer
Presenter(s)
S. B. Chinn1, X. Ma2, M. Tierney3, A. Moatadelro2, E. Rosten2, D. I. Rosenthal4, J. J. Caudell5, N. E. Dunlap6, C. U. Jones7, N. M. Woody8, T. J. Galloway9, F. Nguyen10, A. Raben11, G. Shenouda12, D. Blakaj13, S. Firat14, M. Machtay15, Q. T. Le16, M. Schipper17, and S. S. Yom18; 1Rogel Cancer Center, University of Michigan, Ann Arbor, MI, 2Artera, Inc, Los Altos, CA, 3ArteraAI, Los Altos, CA, 4Department of Radiation Oncology, University of Texas MD Anderson Cancer Center, Houston, TX, 5H. Lee Moffitt Cancer Center and Research Institute, Department of Radiation Oncology, Tampa, FL, 6The James Graham Brown Cancer Center at University of Louisville, Louisville, KY, 7Sutter Medical Center Sacramento, Roseville, CA, 8Department of Radiation Oncology, Taussig Cancer Institute, Cleveland Clinic, Cleveland, OH, 9Department of Radiation Oncology, Fox Chase Cancer Center, Philadelphia, PA, 10Department of Radiation Oncology, Centre Hospitalier de l'Université de Montréal (CHUM), Montréal, QC, Canada, 11Radiation Oncologists, PA; Christiana Care, Newark, DE, 12McGill University Health Center, Montreal, QC, Canada, 13Ohio State University, Columbus, OH, 14Medical College of Wisconsin, Milwaukee, WI, 15Penn State Cancer Institute and College of Medicine, Hershey, PA, 16Stanford University, Stanford, CA, 17University of Michigan, Ann Arbor, MI, 18University of California San Francisco, San Francisco, CA
Purpose/Objective(s): Prognostication of clinical outcomes of head and neck cancers (HNC) is challenging and complicated by the heterogeneity of these cancers. We hypothesized that a digital pathology (DP)-based MMAI approach to biomarker development could provide a quantitative prognostic classifier.
Materials/Methods: We used clinical data and pre-treatment DP images of tumor specimens from 6 NRG Oncology phase III trials (RTOG 0129, 0522, 1016, 9501, 0234, HN002), dividing them into a training set (N=1836) and two validation (val) cohorts: radiation (RT)-based (N=655) and surgery-based (N=167). An MMAI model was developed to predict progression-free survival (PFS) for RT-treated patients, disease-free survival (DFS) for surgically treated patients and overall survival (OS) for both groups. The model generated an image-derived score from DP images which was combined with 8 clinical input variables (age, sex, performance status, HPV/p16 status, smoking status, T-, N-classification, primary tumor site) to generate an “MMAI score.” The locked model was tested in the val cohorts. Missing HPV/p16 data in the val cohorts were imputed with missingness indicator approach (further confirmed by multiple imputation). Model performance was assessed using Cox proportional hazards models and Harrell’s c-indices. Hazard ratios [HR (95% CI)] are per 1 SD.
Results: In both val cohorts, image and MMAI scores were significantly associated with PFS/DFS and OS and remained significantly prognostic after adjusting for AJCC stage, HPV/p16, or primary tumor site (Table 1). The image score remained significantly prognostic for PFS/DFS after adjusting for all 8 clinical input variables [RT:aHR=1.16 (1.03-1.30), p=0.01; Surgery:aHR=1.24 (1.01-1.53), p=0.01]. C-indices for MMAI scores were numerically higher than those for AJCC stage, HPV/p16, or primary tumor site. Comparing 10th v 90th percentile MMAI scores, estimated 3-yr rates of progression/recurrence were 9% vs 34% for p16+ RT-treated patients and 70% vs 90% for p16- surgically-treated patients.
Conclusion: This is the first DP-based MMAI biomarker validated in patients with locoregionally advanced HNC, independent of stage and HPV status. This work suggests that DP images contain important prognostic information beyond those captured by known clinical features.
Table 1: Association between biomarker scores and clinical endpoints in univariable and multivariable analysis| PFS/DFS | OS | ||||||||
| Cohort | Model Score | UVA | p-value | MVA* | p-value | UVA | p-value | MVA* | p-value |
| Definitive RT | MMAI | 2.02(1.78-2.30) | <0.001 | 1.86(1.57-2.21) | <0.001 | 2.42(2.09-2.82) | <0.001 | 2.34(1.91-2.87) | <0.001 |
| Image | 1.31(1.18-1.44) | <0.001 | 1.16(1.04-1.29) | 0.007 | 1.37(1.23-1.52) | <0.001 | 1.19(1.06-1.34) | 0.003 | |
| Definitive Surgery | MMAI | 1.53(1.27-1.85) | <0.001 | 1.54(1.16-2.04) | 0.003 | 1.55(1.26-1.90) | <0.001 | 1.63(1.19-2.23) | 0.002 |
| Image | 1.36(1.14-1.62) | <0.001 | 1.24(1.01-1.54) | 0.043 | 1.35(1.11-1.63) | <0.002 | 1.25(0.99-1.56) | 0.057 | |
| * Adjusted for AJCC stage, HPV/p16, and primary tumor site. | |||||||||