Abstract 4400: Machine learning (ML) on real-world data (RWD) of front-line (1L) metastatic castration resistance prostate cancer (mCRPC) patients for dynamic prediction of time to tx discontinuation (TTD)

Background: Abiraterone (abi) and enzalutamide (enza) are two novel androgen therapies (NAT) for 1L treatment (tx) for mCRPC. As there is no head-to-head randomized controlled clinical trial (RCT) between Abi and Enza, predicting 1L comparative effectiveness of these two drugs for mCRPC, especially in patient groups under-represented in RCTs is critical. ML was performed on electronic health records (EHR) to evaluate risk factors of TTD in 1L mCRPC patients and identify predictive markers of patient subgroups with differential outcomes from enza vs. abi. Methods: Patients with curated mCRPC diagnosis and 1L tx between Jan 2012-May 2022 were identified in a ConcertAI® oncology EHR database. For dynamic prediction of TTD, survival XGBoost models were built with four different timepoints as index: 1L start, and 30, 60, and 90 days from 1L start. A 60:20:20 split was used for train, validation, and test sets. TTD was defined as either tx end or death. Risk factors and interaction terms were identified using SHapley Additive exPlanations (SHAP) and marginal hazard ratios. Results: The sample size ranged between 1580-1636 patients (Table 1). Holdout testset cumulative dynamic AUROC for the models ranged between 0.65-0.75. PSA rise from 1L start, elevated PSA, liver enzymes, and platelets to lymphocytes ratio, low hemoglobin and albumin, prior NAT tx, intake of pain medications, lower BMI, and high comorbidity were risk factors for TTD across all models. Tx with enza or abi was predictive of lower risk of TTD. While enza was better than abi in the overall population, SHAP interaction terms revealed that enza had worse TTD than abi in patients with higher ALP and NATs did not have any TTD benefit over chemo for patients with lower HGB. Conclusions: ML on RWD identified prognostic markers for 1L mCRPC tx discontinuation and predictive markers of patient subpopulations with different outcomes from enza vs. abi. Table 1. Summary of TTD Models: Cohort characteristics, model performance, risk factors: Mean HR (95% CI) Index of models 1L Tx-start 1L Tx start + 30 days 1L Tx start + 60 days 1L Tx start + 90 days N patients (event prevalence) 1636 (89 %) 1611 (89.8%) 1585 (91.3) 1570 (92.2) Model performance: hold-out CD-AUC (mean ± SD) 0.65±0.04 0.70±0.03 0.75±0.04 0.75±0.05 Risk factors Current Tx: enzalutamide (vs no) 0.81 (0.61, 0.92) 0.82 (0.76,0.86) 0.94 (0.91,0.97) 0.97 (0.96,0.99) Current Tx: abiraterone (vs no) 0.84 (0.68, 0.94) 0.91 (0.88,0.94) 0.96 (0.94, 0.98)