Session Item

Current dose de- or escalation trials for head and neck cancer (HNC) patients generally use tumor staging (including HPV status) as patient eligibility criteria. Model-based patient selection of low, intermediate and high-risk patients could improve the effectiveness of personalized dose strategies. The aim is to develop a robust international overall survival risk-stratification model based more than 4500 HNC patients.

The inclusion criteria were HNC patients with squamous cell carcinomas treated with (chemo)radiation, without prior HN treatment. Data from 3 different institutes was split into 4 cohorts: a training (n=2241), independent test (n=786), and 2 external validation cohorts: Cohort 1 (n=1087) and Cohort 2 (n=497). Data imputation was only used in the training cohort; all other data was complete.

We constructed: 1) a non-spatial clinical variable model to establish the outcome risk estimation based on the large-scale data, followed by 2) optional radiomics (spatial) component to further prediction improvement.

Training of the Cox regression models was performed with bootstrapped forward selection. The clinical models were validated in the independent test and 2 validation cohorts. Subsequently, patients were stratified into high, intermediate and low risk overall survival probability based on the predicted 2-year mortality risk; with a priori thresholds of >25% for high risk and <5% for low risk.

The additional radiomics component was developed in imaging sub cohorts (Table). The selected radiomics predictors were added to the linear predictor from the final clinical model.

In the multivariable analyses Performance score, AJCC^8th stage, pack years, and Age were selected for the prediction of overall survival (NB: AJCC^8th stage is based on T and N stage, tumor site and HPV status). Model performance was stable over different cohorts with c-indices ranging from 0.72-0.76 (Table). The prediction model was highly discriminative for stratifying high, intermediate and low risk patients (Figure); the cumulative 5-year overall survival ranged from 92-98% for the low risk group and from 17-46% for the high risk group. The clinical prediction model outperformed clinical standard-of-care AJCC^8th stage prognosis (Table).

In smaller imaging cohorts, the addition of selected radiomics features to the clinical model’s linear predictor further improved the performance in training and validation cohort 1(Table).

This international multi-institutional dataset allowed for the development and validation of a robust overall survival risk-stratification model. The clinical model showed exceptional distinction capacity to select low and high-risk patients for potential dose de-escalation and escalation strategies. Additionally, our right-censoring-aware prediction model approach provides a framework for robust clinical prediction (i.e. without refitting), and has simultaneous flexibility to add image features for further improved risk estimation.