ESTRO 2023 - Abstract Book

S451

Sunday 14 May 2023

ESTRO 2023

Netherlands Comprehensive Cancer Organization, Utrecht, The Netherlands; 6 Department of Urology, Radboud university Medical Center, Nijmegen, The Netherlands; 7 Zuyd University of Applied Sciences, ,, Heerlen, The Netherlands; 8 Fontys University of Applied Sciences, ,, Eindhoven, The Netherlands; 9 Institute for Health Sciences, Radboud university medical centre, Nijmegen, The Netherlands Purpose or Objective For men with localized prostate cancer, choosing the optimal treatment can be challenging since each option has different side effects, such as erectile dysfunction (ED), which negatively impacts their quality of life. Our previous findings demonstrate that logistic regression models are able to identify patients at high risk of ED. However, other algorithms such as Bayesian networks may be even more successful, as they can intricately represent the causal relations between available variables to probabilistically determine if a patient will have ED after treatment with optimal performance. Therefore, this study aims to develop and internally validate a clinically plausible Bayesian network structure to predict one-year ED in prostate cancer patients by combining expert knowledge and evidence from data. Materials and Methods A subset of the ProZIB dataset, which was collected by the Netherlands Comprehensive Cancer Organization (Integraal Kankercentrum Nederland; IKNL) from 69 Dutch hospitals, was used for this study. The subset contained Patient-Reported Outcome Measures (PROMs) and clinical information of 964 localized prostate cancer cases. All patients and variables with missing information were excluded and ED (Yes or No) at one year was considered the endpoint of interest. The initial causal relationship (arrows or arcs) between the variables in the Bayesian network structure was specified by expert prostate cancer treating physicians. Evidence of a relationship with the outcome was based on variable importance from a bootstrapped (B=400) naïve Bayes model. A cut-off threshold of 75 was chosen to either retain or remove relationships that are above or below the threshold respectively to avoid possible model overfit. The performance of the resulting Bayesian network structure was evaluated based on a 5-fold cross-validation area under the curve (AUC) and calibration plots. Results 964 patients with localized prostate cancer were identified from 69 Dutch hospitals. 233 patients with missing information were excluded from this analysis, resulting in a dataset consisting of 731 patients. The variables ‘BMI’ and “volume of prostate via MRI” were excluded from the study due to their high percentage (51% and 76% respectively) of missing information. The median age of patients in this study was 69 (45 – 86) years. Based on the variable importance results, only therapy was retained from the clinical variable to have an influence on the outcome while feeling depressed and Sexual functioning were dropped from the PROMs variables. The resulting structure had 13 covariates and 13 arcs with ED being a child to six variables (Figure 1). The mean AUC of the structure was 0.960 (0.948 - 0.972) with good calibration (Figure 2).

Conclusion