(C-174) Validating Early-stage Cancer Clinical Outcomes for Melanoma and Breast Cancer with Real World Data

Background: Reliable real-world (RW) clinical outcome measures are needed to support health technology assessment, provider and regulatory decision-making. While long-term endpoints such as overall survival have been well validated, there is a need to develop a stepwise method to evaluate the concordance between trial and real-world estimates.

Objectives: To describe stepwise evaluation of the replicability of RW endpoints compared to estimates from Phase 3 clinical trials in two early-stage cancer types.

Methods: This retrospective observational study used electronic health record data from patients treated in US community oncology settings. RW proxies for KEYNOTE trial (KN) eligibility criteria were applied 1/ with elements available in curated clinical data and 2/ with supplemental data from custom chart abstraction (CA). Multiple cohorts were created to assess the impact of missing eligibility data on endpoint concordance. RW constructs for clinical trial endpoint definition were established. Kaplan-Meier curves were used to estimate RW endpoints. Cox proportional hazard ratios and generalized linear model risk ratios compared RW estimates of recurrence-free (RFS), distant metastasis-free (DMFS), event-free (EFS) survival, and pathological complete response (pCR) with corresponding KN control arm estimates from published results. Matching Adjusted Indirect Comparison (MAIC) statistically adjusted differences in available demographic and clinical characteristics. The timing of all underlying data sources referencing each endpoint event in patients’ charts was evaluated.

Results: Key eligibility criteria from clinically curated data were sufficient to achieve concordance between trial control arm estimates and RW endpoints. Missing eligibility data resulted in non-concordance only for DMFS and RFS in the melanoma population. ECOG was the most undocumented variable. Overall, positive confounding for RFS, DMFS, and pCR was adjusted through the addition of CA and MAIC. Selection bias was suggested for RFS and EFS after MAIC, possibly due to post-index collider bias. At least 77% of events were indicated by two sources within 30 days of each other.

Conclusions: RW endpoints RFS and DMFS for melanoma and EFS and pCR for breast cancer were found to be concordant with clinical trial estimates when key study design aspects were aligned, and fit-for-purpose data are leveraged. The magnitude of response to design and analytic approaches is distinct for each endpoint. In almost all cases, multiple data sources substantiated events in RW data. These findings increase confidence in the reliability of early-stage RW outcomes for understanding treatment effectiveness.