295PDevelopment and validation of M1 substages for previously untreated metastatic nasopharyngeal carcinoma

Abstract Background We aim at subdividing M1 stage to better predict survival of metastatic nasopharyngeal carcinoma (NPC) patients whose outcomes could vary greatly. Methods Patients with previously untreated metastatic NPC (training cohort) were recruited prospectively from 2007 to 2018 and were re-staged based on 8th edition of American Joint Committee on Cancer system. All patients had baseline plasma EBV DNA at diagnosis of metastasis. Characteristics of metastases (site, number and size of metastatic lesions) were confirmed by MRI and PET-CT. We used recursive partitioning analysis (RPA) incorporating baseline plasma EBV DNA and/or metastatic characteristics with internal validations to subdivide M1 stage. The two models were externally validated using an independent data set of 67 NPC patients who were non-metastatic at diagnosis but later developed distant metastases after radical treatment (validation cohort). Performance of survival prediction between the two models was compared with paired t-test under 1000 bootstrapping samples. Results The training cohort of 69 patients had a median follow-up of 40.8 months and 3-year overall survival (OS) of 36%. Model 1 incorporating pre-treatment plasma EBV DNA subdivided M1 stage into two groups: M1a (EBV DNA ≤2500 copies/ml; OS 74%) and M1b (EBV DNA >2500 copies/ml; OS 17%) (P 2500 copies/ml) (HR 4.7 (95% CI 1.9-11.5); P=.001) and metastatic site (coexisting liver and bone metastasis) (HR 2.2 (1.0-4.7); P=.046) were prognostic of OS. Model 1 demonstrated better model fit in predicting OS (Model 1: mean AIC 246.9 (95% CI 187.8-303.6) vs Model 2: mean AIC 257.7 (200.6-313.2); P Conclusions A novel RPA-based M1 stage set incorporating baseline plasma EBV DNA had a significantly better survival prediction, providing important values on prognosis and treatment decision making. Legal entity responsible for the study The authors. Funding Has not received any funding. Disclosure All authors have declared no conflicts of interest.