Mathematical predication models to optimize post-treatment surveillance in HPV-associated oropharyngeal cancer.

Abstract

6027 Background: In this study we develop post-treatment imaging surveillance schedules for locally advanced oropharyngeal carcinoma (OPC) specific to the unique recurrence patterns of tumor stage and HPV status, using mathematical models. Current post-treatment imaging surveillance recommendations for OPC are not evidence based. The exception is the use of a positron emission tomography (PET) scan at 3 months post-treatment, after which practice across institutions diverge. An optimized and personalized surveillance schedule for OPC patients can minimize costs and diagnostic delays. Methods: A Markov multi-state model defining local and distant recurrences was trained using 2159 patients from the National Cancer Database. Patients from 2010-2015 treated at an academic or major cancer center with curative radiotherapy were included. Tumors must have been stage III to IVB (AJCC 7th edition) with known p16/HPV status. Model performance was then successfully externally validated using the 2016 International Collaboration on Oropharyngeal cancer Network for Staging (ICON-S) study. Optimized radiographic surveillance schedules were created using this model, assuming a PET at month 3 and including 0 to 6 additional computed tomography (CT) scans of the neck and chest. Optimization was done for minimization of latency, defined as time between disease recurrence and radiographic discovery. Results: Model-selected schedules varied significantly from commonly utilized-surveillance schedules (such as imaging every 3 months within the first year from treatment) and showed lower mean diagnostic latency for every stage and HPV status (shown in Table). In the lowest risk cohort (Stage III HPV+), the optimized schedule had a sensitivity of 65% and latency of 3.1 months. In the highest risk group (Stage IVB HPV-), the optimized schedule had a sensitivity of 76% and latency of 1.9 months. Conclusions: Mathematical model optimization for HPV status and stage is feasible and produces non-intuitive results. These results could be used to inform surveillance if payors reimburse for fewer total scans. Across all cohorts, each added CT scan increases surveillance sensitivity and decreases latency. Incorporation of physical exam and direct visualization results into the model are still needed. Future steps include cost effectiveness research and prospective clinical trials.[Table: see text]