Improving survival prediction of patients treated with external beam radiotherapy for dysphagia in esophageal cancer using prediction models.

358 Background: The recently published POLDER trial investigated the effects of external beam radiotherapy (EBRT) on dysphagia caused by incurable esophageal cancer. As the effects of EBRT were presumed not to be immediate, an estimated life expectancy of minimally three months was required for inclusion. However, nearly a third of the included patients died within three months. The aim of this study was to investigate if the use of prediction models could have improved the physician’s estimation of the patient’s survival, and thus the eligibility for EBRT treatment. Methods: Data from the POLDER trial (N = 110) were linked to the Netherlands Cancer Registry to retrieve additional patient, tumor and treatment characteristics. Two published prediction models (the SOURCE model and Steyerberg model) were used to predict overall survival for all patients included in the POLDER trial. Predicted survival probabilities were dichotomized (predicted deceased/alive at three months) and the positive predictive value, negative predictive value, sensitivity, specificity and the area under the curve (AUC) were used to evaluate the predictive performance. DeLong’s test was used to test the difference between the AUCs of the SOURCE and Steyerberg models for statistical significance. Results: In the POLDER trial, 35 patients were unjustly presumed to survive three months. Predicting survival at three months, the SOURCE and Steyerberg model had an AUC of 0.76 and 0.60 respectively. The difference between the AUCs of the models was significant (p =.017). Under optimal survival cut-off scores, SOURCE would have incorrectly predicted 16 patients to survive three months. For the Steyerberg model this was 22 patients. Furthermore, using SOURCE under these cut-off scores, seven patients were incorrectly predicted to not survive three months compared to 18 patients using the Steyerberg model. Conclusions: Results showed that the SOURCE and Steyerberg models could have improved survival predictions compared to clinical judgement alone. The SOURCE model was found to be a more useful decision aid than the Steyerberg model as it was more accurate. Accepting that a small proportion of patients are incorrectly predicted not to survive three months and are not considered for EBRT treatment, we recommend using the SOURCE model for patients that are considered for palliative treatment of dysphagia caused by esophageal cancer.