The Use and Validation of Oncology Endpoint Proxies in Real-World Studies: Targeted Literature Review Mapping 64 Retrospective and Validation Real-World Studies

OBJECTIVES: Oncology clinical endpoints are not always recorded in real-world data (RWD). Proxy endpoints are increasingly being used in real-world studies (RWS). This targeted literature review (TLR) mapped studies using clinical endpoints proxies, proxying methods used, and any acknowledgment and mitigation of potential bias associated with proxy use.

METHODS: EMBASE^® searches (inception-May 2024) and desktop searches identified oncology RWS using clinical endpoint proxies. One reviewer conducted screening and data extraction with subject-matter expert quality-check.

RESULTS: We included 64 RWS (14 [22%] methodological RWS assessing proxy performance and 50 [78%] retrospective natural history or effectiveness RWS using proxies). Most retrospective studies used breast cancer data (n=10), from United States (n=32) registries (n=21), and proxied the following endpoints: progression-free survival (PFS n=23, recurrence n=12, overall mortality/survival n=7, adverse events n=5, progression n=4, cause-specific survival n=3, [time-to] discontinuation n=3 and treatment response n=1). Bias was not discussed in 39 (78%) studies. Limited reporting of methods could be attributed to space constraints (n=26 were conference abstracts). To examine proxy performance, 11 studies compared their findings to the literature including trials (n=7) or to gold-standard chart abstractor in the same RWD (n=3), compared different proxies’ performance in the same RWD (n=1) or performed sensitivity analyses with different proxy definitions (n=2). The 14 methodological RWS evaluated proxy performance (mostly PFS, n=8 and death/survival, n=4) by comparing the proxy to other outcome ascertainment method (e.g. manual chart abstraction, n=5), comparing RW to trial results (n=3), correlating real-world outcomes (e.g. time-to-discontinuation to survival, n=6), comparing different proxy definitions (n=2) and comparing proxies in different RW datasets (n=2).

CONCLUSIONS: Proxy use in RWS focuses on PFS and is often inadequately reported, limiting replicability. Twenty-five studies provided some information regarding proxy performance. Consensus is needed regarding well-performing proxies, along with consistency in assessing bias in RWS using proxied endpoints, e.g., quantitative bias analysis.