Classification tree phenotypes for early treatment-limiting toxicity after androgen receptor pathway inhibitor initiation in metastatic hormone-sensitive prostate cancer from the IRONMAN registry.

5095 Background: Androgen receptor pathway inhibitors (ARPIs) improve survival in metastatic hormone-sensitive prostate cancer (mHSPC), but early treatment-limiting toxicity remains common. Using the IRONMAN (International Registry for Men with Advanced Prostate Cancer; NCT03151629) registry, we developed an interpretable classification tree that yields simple rule-based 1-year risk strata for a composite adverse outcome (AO) after ARPI initiation. Methods: Men with mHSPC initiating first-line ARPI with at least 12 months follow-up or AO within 365 days were included (n=583). AO was the first serious adverse event (SAE) or ARPI discontinuation for toxicity or other non-progression reasons; progression-attributed events were not counted. Candidate peri-ARPI initiation predictors (demographics, labs, treatments/medications, patient-reported outcomes) were screened for missingness, imbalance, and collinearity, reduced by stepwise logistic regression, and entered a recursive-partitioning classification tree with restricted depth and node size. Discrimination was summarized by area under the receiver operating characteristic curve (AUC) and repeated 3-fold cross-validation. Terminal nodes were collapsed into low, intermediate, and high-risk strata by predicted 1-year AO. Results: Overall, 122/583 men (20.9%) had AO within 1 year (85 SAEs; 37 discontinuations [28 toxicity, 9 other non-progression]). AO-free survival did not differ by ARPI agent (log-rank p=0.40); ARPIs were pooled. Peri-ARPI docetaxel exposure (triplet therapy) was uncommon (2.4%). The final tree retained nine predictors, including European Organization for Research and Treatment of Cancer global health rating (1-7), hemoglobin, worst pain (0-10), lactate dehydrogenase (LDH), age, docetaxel, and opioid agonist and systemic steroid use. In-sample AUC was 0.76 (95% CI 0.71-0.81); cross-validated AUC was 0.60 (95% CI 0.55-0.65). Collapsed strata showed distinct 1-year AO risk (Table 1). High-risk phenotypes reflected poor health/symptoms with anemia and/or elevated LDH, and a treatment-intensity phenotype with triplet therapy plus opioids and steroids. Conclusions: A small classification tree identified three clinically transparent phenotypes of early AO after ARPI initiation in mHSPC and can be applied as simple decision rules to support monitoring and shared decision-making. External validation is needed. Tree risk strata and 1-year AO risk. Risk stratum n (%) 1-year AO %, (95% CI) Low 422 (72) 10.7 (7.7–13.6) Intermediate 80 (14) 26.3 (16.6–35.9) High 81 (14) 69.1 (59.1–79.2)