ESTRO 2024 - Abstract Book

S1269

Clinical - Head & neck

ESTRO 2024

1 Centre of Cancer and Organ Diseases, Copenhagen University Hospital - Rigshospitalet, Department of Oncology, Copenhagen, Denmark. 2 Sahlgrenska University Hospital, Department of Oncology, Gothenburg, Sweden. 3 Institute of Clinical Sciences, Sahlgrenska Academy, University of Gothenburg, Department of Medical Radiation Sciences, Gothenburg, Sweden. 4 Sahlgrenska University Hospital, Therapeutic Radiation Physics, Department of Medical Physics and Biomedical Engineering, Gothenburg, Sweden. 5 Centre of Head and Orthopaedics, Copenhagen University Hospital - Rigshospitalet, Department of Otorhinolaryngology, Head and Neck Surgery and Audiology, Copenhagen, Denmark, Denmark. 6 University of Copenhagen, Faculty of Health and Medical Sciences, Copenhagen, Denmark

Purpose/Objective:

A thorough validation process is necessary for risk models to be clinically useful. For the model to add value, the predictive power of the risk model should also exceed that of easily accessible clinical variables. Prediction of loco regional failure (LRF) is important when selecting patients for trials. Here, a previously published multi-endpoint model[1] was validated in an independent data set with data from two institutions. As second aim, we compare the multi-endpoint model´s performance for LRF prediction to that of UICC staging.

Material/Methods:

The independent validation cohort consisted of 741 head and neck cancer patients treated with primary radiotherapy between 2013-2019 (533 from Institution 1 and 208 from Institution 2, different country). The multi endpoint model was developed in a cohort of 560 patients (from Copenhagen Denmark 2009-2012) and was published in 2017[1]. In short, the model simultaneously predicted LRF, distant metastasis (DM) and death with no evidence of disease (death NED) in a competing risk setting with input variables: Tumor subsite (with p16+ oropharynx as separate site), Tobacco, T- and N-stage, Gross Tumor Volume (GTV), Age, Performance Status and “Candidate for cisplatin”. The model was applied to the validation set patients and individual 3-year risks of LRF, DM and Death NED were calculated. Expected and observed risks were compared by calibration plots and Brier score[2,3]. The discriminatory ability was evaluated with risk group stratified cumulative incidence plots and a time-dependent (3-yr) concordance index (AUC)[3]. UICC stage was determined for all patients, both using version 8 and version 7 (where possible). The observed 3-year LRF was determined for high- and low risk patients following different strategies for risk stratification (risk model estimates vs UICC stage).

Results:

The calibration plots can be seen in Figure 1. The multi-endpoint model generally overestimated the risk of LRF, while the DM and Death NED predictions were more accurately calibrated. The Brier scores showed that individual risk estimates for LRF did not significantly differ from assigning observed group mean to all patients. As for discriminatory ability, the 3-year AUC for LRF was well above 0.5 (0.65 (0.61-0.70)). The cumulative incidence plots showed that the model was able to stratify patients based on risk estimates (Figure 2). UICC staging by version 8 performed equally well, while UICC version 7 had inferior discriminatory ability. Table 1 shows the observed 3-year LRF for different strategies of risk stratification.