Individual patient data (IPD) analysis of early PSA nadir in ARASENS, LATITUDE, and TITAN: Training and validation of a novel model.

192 Background: Early PSA nadir after systemic therapy in mHSPC is a predictor of overall survival. There are no prediction models that have been validated in randomized clinical trials (RCTs) to help identify patients who experience early PSA response. Using IPD from three phase III randomized trials, ARASENS, LATITUDE, and TITAN, we trained and validated a model to predict early PSA nadir in mHSPC patients. Methods: Eligible trials that randomized mHSPC patients to receive an androgen receptor pathway inhibitor (ARPI) were identified through Medline and clinicaltrials.gov. Three trials were identified that had available IPD through online data-sharing portals, and a pre-specified analysis plan was approved. Early PSA nadir was defined as ≤0.2 ng/mL by 6 months of random allocation. Patients who received androgen deprivation (ADT) (+/- docetaxel [doce]) with ARPI, were split randomly 60:40 into a training and a testing cohort. Patients who received ADT monotherapy (+/- doce) as standard of care (SOC) were used as a validation cohort. A random forest classifier model was constructed in the training cohort with 10-fold cross-validation. Included variables were age, performance status, body mass index (BMI), Gleason score, metastatic stage at diagnosis, visceral metastasis, PSA and hemoglobin at baseline, use of docetaxel, and receipt of prior local therapy (RP and/or RT). After internal validation in the testing cohort, the locked model was applied to the validation cohort, and performance was assessed with area under curve (AUC) and Brier score. Results: Data was available for 3434 patients. Overall, 1718 patients received SOC plus ARPI and 1716 patients received SOC alone. The training cohort consisted of 1030 patients while the testing cohort and validation cohort consisted of 688 and 1716 patients, respectively. The top 5 variables in order of importance were baseline PSA, hemoglobin, BMI, age, and ECOG performance status. The AUC for the testing and validation cohort was 0.75 and 0.77, respectively. When stratified by tertile of predicted probability, the proportion of PSA nadir was 32%, 52%, and 83% in the testing cohort and 7%, 15%, and 41% in the validation cohort, respectively. The Brier scores for the testing and validation cohort were 0.20 and 0.26, respectively. The modest calibration (i.e., higher Brier score) in the validation (SOC alone) cohort could be attributed to slight overprediction of early PSA nadir probability by a model trained in SOC plus ARPI group. Conclusions: To our knowledge, this is the first trained and validated model to predict early PSA nadir by 6 months of treatment initiation in mHSPC using data from multiple phase III RCTs. This model could provide clinical utility to guide treatment and monitoring strategies, as well as conducting clinical trials to enrich patient accrual for those clinically impacted by early PSA nadir.