Marieke van Son

151 PREDICTION MODELS FOR BF AFTER FOCAL SALVAGE HDR-BT and PSADT are strong predictors for BF after focal salvage treatment for radiorecur- rent PCa. While we did not investigate the predictive value of PSA nadir alone, we did incorporate it in our model by using PSA reduction. We argue that this might be a better predictor than PSA nadir, given its dependence on pre-salvage PSA. Furthermore, PSA nadir is also influenced by other factors, such as prostate volume[10]. We did not assess pre-salvage Gleason score as a potential predictor, as biopsies were not performed from the end of 2017 onwards (leading to 44.7% missing values). Also, while some have identified variables from the primary tumor and/or treatment as predictors, we did not investigate any primary tumor characteristics because of our limited sample size and missing data in these characteristics. Furthermore, the predictive value of these variables in focal salvage studies seems limited[15]. With an extended sample size and follow-up, we could potentially investigate the added value of some of these predictors. There are several strengths to our study. Missing data for candidate pre-salvage predictors was very low (0.7%) due to prospective data collection. The inclusion of patients treated off-protocol also makes the study sample more representative and increases external validity. Furthermore, candidate predictors for multivariable analysis were selected based on literature and clinical knowledge rather than by performing univariable analysis, thereby minimizing the occurrence of type-I errors[25]. The online dynamic nomograms we created are helpful tools to quickly assess and visualize in- dividual predicted bDFS. The study has some limitations. First, external validation of this model is necessary. Several other focal salvage strategies have been described, all with minor differences with respect to eligibility of patients. Therefore, such cohorts offer an opportunity for external validation. Especially since both models use predictors that are known to be related to PCa progression and none of them are treatment specific. External validation of our models could lead to adjustment of these models and thereby improve predic- tive accuracy and be applicable to other focal salvage modalities. Despite taking into account the sample size, some overfitting is indicated by the suboptimal shrinkage factors of 0.85 and 0.81, indicating 15% and 19% optimism, respectively. Furthermore, limiting the number of candidate variables might have led to missing important pre- dictors, such as DSFI [15]. Consequently, the C-statistic of 0.73 of the first model might be improved by including other potential predictors when sample size has increased. Third, length of follow-up was relatively short with a median of 25.1 months, thus the models perform optimal within a timeframe of approximately two years. Fourth, tumor volume was based on the delineated GTV. Although GTV delineation was based on mpMRI and PSMA PET-CT, which improves the estimation of tumor volume compared to mpMRI alone[31], interobserver variability due to the lack of delineation guidelines will be present and influences the accuracy and predictive value of this variable. 8