Prognostic and Predictive Performance of Routine Clinicopathologic Variables in 10,535 Men Enrolled on Randomized Phase III Trials in Localized Prostate Cancer.

Abstract

PURPOSE/OBJECTIVE(S)\nStandard clinicopathologic variables including T-category, Gleason grade, and prostate-specific antigen (PSA) guide the management of localized prostate cancer. The prognostic and predictive performance of these variables has been assessed primarily in retrospective cohorts without standardized quality assurance. Given the substantial heterogeneity in oncologic outcomes for men with localized prostate cancer, we hypothesized that standard clinicopathologic variables have inherent limitations. We therefore sought to assess their performance in a meta-analysis of 10 phase III randomized trials.\n\n\nMATERIALS/METHODS\nIndividual patient data was obtained from 10 radiotherapy-related RCTs (NRG/RTOG 9202, 9408, 9413, 9910, 0126; EORTC 22863, 22961, 22991; DART 01/05 GICOR, and Ottawa 0101). The cohort was split 2:1 into training and validation cohorts. In the training set, multivariable cox regression and random forest models were developed. Concordance indices (c-indices) and bootstrapped 95% confidence intervals were calculated for models based on age, performance status, T-category, Gleason score, and PSA. Interactions with androgen deprivation therapy (ADT) treatment effects were also tested. Endpoints included distant metastasis (DM), metastasis-free survival (MFS), and overall survival (OS).\n\n\nRESULTS\nA total of 10,535 patients were included. The training cohort (n\u202f=\u202f7382) was used to develop an optimized clinicopathologic model. Using the locked-model in the validation cohort (n\u202f=\u202f3153), the 10-year c-index was 0.70 (95% CI: 0.66-0.74), 0.61 (0.59-0.63), and 0.61 (0.59 - 0.63), for 10-year DM, MFS, and OS, respectively. Similar results were obtained using random forest models. In trials testing the addition of short-term (ST) ADT, or those testing the addition of long-term (LT) ADT, there was no evidence of treatment effect modification by any variable for any endpoint (all P-interaction > 0.05). There was, however, significant heterogeneity in the absolute benefit derived from ADT intensification; the empirical 10-year absolute DM risk reduction with LT ADT over ST ADT ranged from 4.1% to 18.0% depending on modelled baseline risk.\n\n\nCONCLUSION\nIn this large cohort with long-term follow-up and trial-level quality assurance, models based on standard clinicopathologic variables had limitations in accurately estimating prognosis. The proportional benefit of ADT use was consistent, but the absolute magnitude of benefit varied greatly by predicted baseline risk of DM. Incorporation of genomic prognostic information into risk stratification to better estimate absolute ADT benefit are needed and ongoing (NCT04513717).