ESTRO 2024 - Abstract Book

S1280

Clinical - Head & neck

ESTRO 2024

Currently used tumor staging systems for nasopharyngeal carcinoma (NPC) are not optimal to accurately predict outcomes. Patients with similar tumour stages that receive similar treatment regimens can have different outcomes. Factors like plasma Epstein Barr Virus (EBV) DNA, tumor volume, and smoking index have been identified as significant prognostic factors. The aim of this study was to develop prediction models using a comprehensive set of routinely used clinical parameters for overall survival (OS), progression-free survival (PFS), locoregional recurrence (LRR), and distant metastasis (DM) in a large cohort of locoregionally advanced NPC patients after (chemo)IMRT and to compare model performance with the stage-only model.

Material/Methods:

The study population was composed of 1455 patients with complete endpoints at 5-year follow-up. The following variables were included as candidate variables: sex, age, BMI, WHO performance status, T-stage, N-stage, AJCC stage, number of days between first symptoms to treatment (interval), pack years, plasma EBV DNA level, and GTV volume. Variables were dichotomized based on the optimal cut-off value assessed by the Kaplan-Meier curve analysis. The dataset was split into a training (70%) and a test set (30%) stratified on the number of events and the start date of treatment. Multivariate imputation on missing data was applied 10 times to both sets separately. Among the patients, pack year was missing for 4% of cases, 53% for EBV DNA, and 30% for tumor volume. Pre selection was performed as follows: features with a p-value<0.1 association with the endpoint in univariable analysis were selected. For highly correlated features (i.e., Pearson correlation coefficient > 0.8), only the feature with the highest association with the endpoint was considered as a variable. A stepwise forward selection Cox regression model was applied using the selected candidate variables based on the Bayesian Information Criterion (BIC) standard. For the final feature selection, variables selected more than half the time were included in the final model. The pooled c-indexes with the 95% confidence intervals were calculated. In the final multivariable Cox regression model, age, T-stage, N-stage, and plasma EBV DNA were selected for OS. Age, T-stage, N-stage, and pack years were selected with PFS. Model performance in terms of c-index of the OS and PFS model performed significantly better than the models with AJCC-stage only: In the test set the c-index of OS increased from 0.604 to 0.703, and for PFS, it improved from 0.640 to 0.684. When stratification of patients into low and high-risk groups based on the median value of model predictions, a significant difference in the Kaplan-Meier curves was observed for OS and PFS. The 5-year OS values for the low and high-risk groups were: 76.1% vs. 60.7% in the training set and 70.5% vs. 59.1% in the test set (p<0.001) and the 5-year PFS values were: 71.4% vs. 57.9% in the training set and 74.1% vs 59.5% in the test set. For the LRR model, only T-stage was a significant predictor. The LRR model exhibited a lower c-index (0.534) compared to the AJCC-stage model (c-index of 0.648). The stratification of patients into high and low-risk for LRR showed a statistically significant difference in the training set with a 5-year LRR-free survival rate of 94.7% for the low-risk group and 90.9% for the high-risk group (P=0.021). However, this difference was not statistically significant in the test set (P=0.4). For the DM model, the feature selection included pack years, GTV-volume, and plasma EBV DNA. The c-index of the DM model decreased from 0.638 to 0.606 when compared to the AJCC-stage model. However, significant stratifications of patients into low and high-risk were achieved in both the training set and the test set, with the 5-year DM-free survival rates of 69.9% compared to 62.2% in the training set (P<0.001), and 65.9% compared to 60.2% in the test set (P=0.036). Results:

Table 1. Multivariable analysis models for overall survival (OS), progression-free survival (PFS), locoregional recurrence (LRR), and distant metastasis (DM).