S1228
Clinical - Urology
ESTRO 2026
University of Texas at Austin, Austin, USA. 5 Oden Institute for Computational Engineering and Sciences, The University of Texas at Austin, Austin, TX, USA Purpose/Objective: To develop a Bayesian mechanistic framework for patient-specific prediction of prostate-specific antigen (PSA) dynamics after external beam radiotherapy (EBRT) [1, 2], aiming to improve early detection and risk assessment of biochemical relapse in prostate cancer patients. Material/Methods: Single-center longitudinal PSA measurements from 166 patients undergoing EBRT were analyzed using a mechanistic model capturing PSA dynamics [1]. Patient-specific model parameters were estimated through Bayesian calibration [3], enabling uncertainty quantification and probabilistic calculation of key model-based biomarkers of biochemical relapse, namely proliferation rate, ratio of tumor proliferation and radiation-induced cell death rates, time to progression, PSA nadir (i.e., lowest post-treatment PSA), and time to PSA nadir since EBRT termination. We quantitatively characterize risk by estimating the α - superquantiles of the posterior distributions of these biomarkers with α =0.95 [3], thereby capturing the 5% upper or lower tail behavior of their distributions. Hence, superquantile-based biomarker predictive performance was assessed through ROC curve analyses. Additionally, our predictive biomarkers were evaluated at multiple post-treatment time horizons (1 to 10 years post-EBRT) to analyze how the predictive ability evolves over time since EBRT completion. To assess clinical benefit, we defined a Days Gained to Biochemical Relapse Diagnosis (DGBRD) metric, representing the time gained by model-based relapse detection compared with standard of care PSA-based criteria (e.g., nadir+2 ng/mL). Results: Our Bayesian mechanistic framework enabled reconstruction and prediction of patient-specific PSA trajectories across the cohort with median (IQR) of the α -superquantile of the RMSE of 0.48 (0.31, 0.65) ng/mL and 0.69 (0.31, 1.20) ng/mL, respectively. Superquantile-based ROC curve analysis of the model- based biomarkers showed that time to progression and proliferation rate exhibited strong discriminative performance between relapsing and non-relapsing patients (AUC > 0.95), while the ratio between tumor proliferation and radiation-induced cell death rates, PSA nadir, and time to PSA nadir also showed high accuracy (AUC > 0.85). For time to progression and proliferation rate, the triaging performance improved over time achieving AUC > 0.70 4 years post-EBRT. The DGBRD analysis further demonstrated that these two biomarkers anticipate biochemical relapse detection relative to clinical PSA criteria by median (IQR) 283 (90,
significant difference was observed between PBT and IMRT in matched-pair analysis: 42% (39/92) versus 35% (32/92), p=0.29. Analysis of all patients confirmed this finding: 45% (75/168) versus 36% (34/95), p=0.16. Unexpectedly, PBT showed significantly higher GI toxicity in both matched (26% vs. 9%, p=0.002) and unmatched analysis (30% vs. 10%, p<0.001). GU toxicity rates were comparable between groups in matched analysis (25% vs. 30%, p=0.41) and unmatched analysis (26% vs. 31%, p=0.45). Conclusion: In this prospective matched-pair comparison, prostate-only proton therapy did not demonstrate the hypothesized reduction in combined moderate-to- severe GI and GU toxicity compared to photon IMRT. Contrary to expectations, PBT was associated with significantly higher GI toxicity rates. These findings challenge the assumption of superior clinical outcomes with protons for prostate cancer compared to state-of-the art image guide IMRT with rectal ballon application and warrant further investigation into technical factors, dose distribution uncertainties, and patient selection. Cost-effectiveness of PBT for prostate cancer requires critical re-evaluation.
Keywords: Prostate cancer, Proton therapy, adverse events
Digital Poster 2381
Personalized risk assessment of post-radiotherapy biochemical relapse in prostate cancer using a Bayesian mechanistic framework Miguel Anxo Vicente Pardal 1 , Chiara L. Deantoni 2 , Nadia Di Muzio 2,3 , Anirban Chaudhuri 4 , Ernesto A.B.F. Lima 5 , Guillermo Lorenzo 1,5 1 Group of Numerical Methods in Engineering, Department of Mathematics, University of A Coruna, A Coruna, Spain. 2 Department of Radiation Oncology, IRCCS San Raffaele Scientific Institute, Milan, Italy. 3 Department of Medicine, “Vita-Salute”, San Raffaele University, Milan, Italy. 4 Oden Institute for Computational Engineering and Sciences, The
Made with FlippingBook - Share PDF online