Improving the prognostication of patients receiving palliative radiotherapy

Background: We aim to develop a prognostic model for patients undergoing palliative radiotherapy (RT), and to compare its performance with alternative prognostic models.Methods: A total of 288 patients were prospectively recruited and randomly split into a 60% training cohort and a 40% testing cohort for model development and validation. Prognosticators were identified via Cox regression analyses. Additional parametric survival models with the identified prognosticators included as covariates were fitted. Model fit and discrimination ability of various fitted models were compared to identify the best model. Patients were grouped according to the number of prognostic factors of the NRF model and TEACHH model for comparison.Results: In the training cohort, neutrophil-lymphocyte ratio (NLR) ≥6, inpatient status at start of RT, prior chemotherapy within 3 months of starting RT, and Palliative Performance Scale score ≤50% were independently and adversely associated with OS. Amongst the various models considered, the log-logistic accelerated failure time (AFT) model had the best performance [log-logistic AFT: explained variation R2 (R-square) =0.314, Royston and Sauerbrei’s D-statistics (D-stats) =1.956, Harrell’s concordance index (c-index) =0.686; Cox: R-square =0.242, D-stats =1.158, c-index =0.688; Weibull AFT: R-square =0.334, D-stats =1.450, c-index =0.688; flexible Royston-Parmar PO(3): R-square =0.293, D-stats =1.863, c-index =0.685]. This model also achieved good calibration and discrimination (D-stats =1.812, c-index =0.683) in the testing cohort, and performed better than the NRF model (D-stats =0.510, c-index =0.567) and the TEACHH model (D-stats =0.656, c-index =0.580).Conclusions: We developed a prognostic model for Asian patients undergoing palliative RT which had better performance than the NRF and TEACHH models. Further work to externally validate this model are required.