S1303
Clinical - Urology
ESTRO 2026
Feature importance consistently identified PSA before sRT, positive surgical margins, SVI, and pT stage as top predictors. The Youden-optimal PSA cutoff was 0.60 ng/mL (AUC = 0.80; 95% CI 0.73–0.86). Patients with PSA ≥ 0.6 ng/mL had shorter median BFF ( ≈ 30 months) than PSA < 0.6 ng/mL (not reached), log-rank p < 0.001.
Digital Poster Highlight 4665 Machine Learning–Driven Prediction of Biochemical Failure After Salvage Radiotherapy for Prostate Cancer Hrvoje Br č i ć 1 , Dino Beli ć 2 , Ante Matani ć 2 , Jure Murgi ć 1,3 , Ana Fröbe 1,4 1 Clinic for oncology and nuclear medicine, Sestre milosrdnice UHC, Zagreb, Croatia. 2 Clinic for radiotherapy and oncology, UHC Osijek, Osijek, Croatia. 3 School of medicine, Croatian Catholic University of Zagreb, Zagreb, Croatia. 4 School of dental medicine, University of Zagreb, Zagreb, Croatia Purpose/Objective: After radical prostatectomy (RP), salvage radiotherapy (sRT) remains the only potentially curative treatment for biochemical recurrence (BCR). However, outcomes following sRT vary widely, and precise identification of patients at high risk of biochemical failure (BF) is lacking. The aim of this study was to identify predictors of BF after sRT and to compare the predictive performance of conventional regression and machine learning (ML)–based models in a single-institution cohort. Material/Methods: Data from 378 patients treated at the University Hospital Centre Sestre milosrdnice, Zagreb, Croatia, who underwent sRT for BF after RP were retrospectively analysed. Clinical, pathological, and treatment variables—including age, PSA before sRT, pathological T and N stage, extracapsular extension (ECE), seminal vesicle invasion (SVI), surgical margin status, and androgen deprivation therapy (ADT)—were extracted from the institutional registry. Analyses were performed in Python (v3.11) using pandas, scikit-learn, lifelines, matplotlib, and SciPy. Missing data were handled by a surrogate-split–like imputation (numeric = − 999; categorical = “MISSING”). Predictive models included L1-regularized logistic regression (Lasso), Classification and Regression Trees (CART), and Random Forests, validated by repeated 5 × 3-fold stratified cross-validation (seed = 42). Model performance was evaluated by AUC, accuracy, sensitivity, specificity, and F1 score. Permutation testing assessed AUC significance, and Hanley–McNeil analysis estimated power. PSA before sRT was analyzed by ROC to define the Youden-optimal cutoff. Kaplan–Meier curves compared BFF survival by this PSA threshold (log-rank test). Results: Biochemical failure occurred in one-third of patients (N=109). Mean cross-validated performance: Random Forest AUC = 0.81 ± 0.04, L1-Logistic Regression AUC = 0.78 ± 0.05, CART AUC = 0.74 ± 0.06. All permutation tests were significant (p < 0.01). Power analysis confirmed >95% power for Random Forest (AUC > 0.8).
Conclusion: ML-based models, particularly Random Forests, outperformed traditional regression in predicting biochemical failure after sRT. Pre-sRT PSA level was the single strongest predictor, with a clinically meaningful cut-off of 0.6 ng/mL identifying high-risk patients. Integration of ML algorithms with standard clinical variables may refine patient selection and timing of sRT, improving personalization of post- prostatectomy management. Validation in external, multi-institutional cohorts is needed. References: 1. Valdes G, Chang AJ, Interian Y, Owen K, Jensen ST, Ungar LH, Cunha A, Solberg TD, Hsu IC. Salvage HDR Brachytherapy: Multiple Hypothesis Testing Versus Machine Learning Analysis. Int J Radiat Oncol Biol Phys. 2018 Jul 1;101(3):694-703. 2. Tang S, Zhang H, Liang J, Tang S, Li L, Li Y, Xu Y, Wang D, Zhou Y. Prostate cancer treatment recommendation study based on machine learning and SHAP interpreter. Cancer Sci. 2024 Nov;115(11):3755-3766. Keywords: BF, SRT, prostate_cancer
Made with FlippingBook - Share PDF online