ESTRO 2020 Abstract book

S326 ESTRO 2020

low risk patients. There was a good calibration for DM, but with a small range of assigned risks. The model under- estimated the risk of Death NED (Figure 1).

Conclusion These preliminary results show that a model based on radiomics significantly predicts OS in TNM-7 Stage-III and - IV HNSCC patients in an external validation dataset while it is not able to significantly stratify the patients in the prospective validation dataset. The lower patient count and larger variance in the prospective dataset may explain this, as predictive performance as measured by the C- Index is much higher. The relatively low predictive performance in external validation may be explained by differences in patient characteristics and treatment, and ComBat based on machine and/or reconstruction method instead of centers may also improve signature generalizability. Funding European Union Horizon 2020 research/innovation program (689715). Dutch Cancer Society (Design - KWF-A6C7072). Dutch Cancer Society (KWF Kankerbestrijding), (12085/2018-2) PD-0544 Validation of a multi-endpoint risk model for clinical outcome in head-and-neck cancer K. Håkansson 1 , J. Kjems 1 , J.H. Rasmussen 2 , L. Specht 1 , J. Friborg 1 , I.R. Vogelius 1 1 Rigshospitalet Copenhagen University Hospital, Department of Oncology- Section of Radiotherapy, Copenhagen, Denmark ; 2 Rigshospitalet Copenhagen University Hospital, Department of Otorhinolaryngology- Head and Neck Surgery and Audiology, Copenhagen, Denmark Purpose or Objective It is essential that published risk models are validated to be clinically relevant, and if they are to be used for trial design/choice of treatment, other endpoints than survival need to be modeled. Here, a previously published multi- endpoint prognostic model for head-and-neck cancer squamous cell carcinoma (HNSCC) is validated in an independent set of patients. Material and Methods A previously published multi-endpoint model 1 was evaluated in an independent validation dataset of 307 HNSCC patients treated at the same institution but after the original (training) dataset. The model included the endpoints of loco-regional failure (LRF), distant metastasis (DM) and death with no evidence of disease (NED) using competing risk analysis, based on the following prediction variables: Smoking status, T stage, N stage, tumor subsite (p16+ OPSCC vs p16- OPSCC, larynx, oral cavity and hypopharynx), gross tumor volume, age (for Death NED only) and WHO performance status (for Death NED only). The predictive performance at three years after start of radiotherapy was evaluated using a time-dependent C- index (AUC) for discriminative performance, and a plot of observed vs predicted risks for calibration performance. For the calibration plot, the validation patients were divided into quartiles of predicted risk. The AUC:s in the validation set were compared to the values found for the training set in the previous publication. Results The multi-endpoint model showed a discrimination performance for DM and Death NED in the validation set in the same order as previously published for the training set (Table 1). The AUC for LRF was lower in the validation set than in the training set (63% vs 73%). The lower ends of the confidence intervals all exceed the 'coin-toss' level at 50% (Table 1). The calibration plot for LRF showed a deviation for patients in the 3 rd risk quartile, but fitted well for the