ESTRO 2025 - Abstract Book

S2055

Clinical - Urology

ESTRO 2025

The training cohort consisted of patients where ADT was not deemed to influence PSA (no ADT, ADT ceased at least 12-months before follow-up). Validation was performed on both the full test cohort (variable ADT status during follow-up), and a subset of patients who continued ADT during follow-up (n=132). A multi-task, population-based, Gaussian process (GP) model (MagmaClustR) was trained to define clusters representing proto-typical PSA trajectories (training data). The posterior predictive distribution of each cluster (training), and the posterior mixture probabilities of each patient (validation), were calculated to assign patients to a cluster. Kaplan-Meier analysis was performed, dichotomizing by PSA cluster. The end point of the study was BCR (nadir+2ng/ml). Results: Cohorts were characteristically different for all baseline characteristics (age, T-stage, Gleason grade, baseline-PSA, p<0.001, Chi-square, Kruskal-Wallis). 60%-70% of patients in the training and full validation cohort were high-risk. 96% in the ADT-influenced validation cohort were high-risk. 73/244 patients in the training cohort did not receive ADT; 148/244 had short- term (≤2 years), and 22/244 long -term ADT (≥2 years). All patients in the validation cohort were prescribed ADT (short -term: n=237/460, long-term: n=223/460), and all patients in the ADT-influenced validation cohort had long-term ADT (2-3 years). Three distinct PSA trajectories were identified (figure 1). Kaplan-Meier analysis showed BCR risk was significantly higher for those with unstable PSA (p<0.0001, figure 2A). Results were consistent in both validation cohorts (p<0.0001, figure 2B-C). There was no correlation between PSA cluster and clinical variable for any cohort. There was no correlation between ADT-influence during follow-up and cluster in the full validation cohort (p=0.29, Chi-square).