Presenter(s)
A. Sabbagh1, J. Y. Qian1, A. Singh2, Y. Wu3, F. Yang4, J. Suggitt5, T. Treechairusame6, E. S. Polanco3, Z. Zhang7, D. Mah8, K. Sine8, A. Shim9, H. Lin9, J. J. Kang10, S. M. McBride1, N. Riaz1, D. Y. Gelblum1, A. Shamseddine1, Y. Yu1, J. Huryn2, S. Yom2, J. Cracchiolo2, I. Ganly2, M. Cohen2, E. Sherman11, A. Ho3, R. J. Wong2, E. C. Dee1, C. Estilo2, and N. Y. Lee1; 1Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, 2Department of Surgery, Memorial Sloan Kettering Cancer Center, New York, NY, 3Memorial Sloan Kettering Cancer Center, New York, NY, 4Department of Radiation Oncology, Mayo Clinic, Phoenix, AZ, 5NYU Grossman School of Medicine, New York, NY, 6Division of Radiation Oncology, Faculty of Medicine Siriraj Hospital, Mahidol University, Bangkok, Thailand, 7Department of Epidemiology and Biostatistics, Memorial Sloan Kettering Cancer Center, New York, NY, 8ProCure Proton Therapy Center, Somerset, NJ, 9New York Proton Center, New York, NY, 10Memorial Sloan Kettering Cancer Center, New Haven, CT, 11Department of Medicine, Memorial Sloan Kettering Cancer Center, New York, NY
Purpose/Objective(s):
Osteoradionecrosis (ORN) is a serious complication of radiation therapy in head and neck cancer. We have previously showcased that proton therapy, concurrent chemotherapy, and smoking are independent risk factors of ORN. This study investigates the utility of machine learning methods in identifying other variables associated with the development of ORN, including relationships such as interaction effects that may not be readily captured using traditional analysis frameworks.Materials/Methods:
Data was collected from consecutive patients with oropharyngeal squamous cell carcinoma at a single high-volume academic institution, treated between January 2013 and December 2023. Baseline characteristics were obtained from electronic medical records. The outcome of interest (ORN) was cross-referenced with the institution’s Dental Service database of patients with ORN to verify that all cases have been included. An extreme gradient boosting (XGBoost) model was fit on the entire dataset using baseline clinical characteristics (age at diagnosis, sex, smoking status, HPV status, T and N stage, use of concurrent and/or induction chemotherapy, use of proton therapy, RT dose, and postoperative status) with outcome defined as the development of ORN within 3 years of completing treatment. Shapley Additive Explanations (SHAP) plots were used to explore associations and interaction effects between the variables and outcome. Associations of interest were tested on univariate analysis (Cox proportional hazards model) with the outcome of interest defined as time-dependent development of ORN. If an association was found to exist, the variables were subsequently included in a multivariate model.Results: The analysis included data from 1564 patients. A SHAP summary plot was generated based on the XGBoost model described above. Age at diagnosis was the most predictive feature based on the SHAP plot. In the univariate model, the association between age and ORN did not meet statistical significance (HR 1, p = 0.901). SHAP dependence plots were then generated and showed an association between age > 70 and postoperative status, with the presence of both leading the XGBoost model to predict a higher risk of ORN within 3 years. This interaction term was tested on univariate analysis, and found to be statistically significantly associated with ORN (HR 5.20, p = 0.03). It was then included in a multivariate model along with known risk factors (smoking status, concurrent chemotherapy, and proton use) as well as age > 70 and postoperative status. The association between the interaction term and ORN remained present even when adjusting for the remaining variables (HR 7.2, p = 0.011).
Conclusion: An interaction exists between age and postoperative status on the risk of ORN development. This association was detected using machine learning. Additional studies and larger cohorts are needed to further characterize this relationship and guide the standard of care in older patients with head and neck cancers.