6C-02 - Application of a Principled Framework to the Robust Emulation of the CheckMate649 Trial in First Line Gastric Cancer Using a Nationwide Electronic Health Record (EHR)-Derived Database

Background: The target trial emulation framework gained popularity in real world evidence (RWE). However, experience with emulating complex oncology trials is limited and principled approaches are needed to increase confidence in oncological RWE.

Objectives: To emulate the CheckMate649 (CM649) trial combining the target trial emulation framework and principled approaches to missing data.

Methods: CM649 assessed first line (1L) nivolumab + chemotherapy (FOLFOX/CAPOX) vs. chemotherapy alone in untreated, non-HER2+ patients with advanced/metastatic gastric, gastro-oesophageal junction, or oesophageal adenocarcinoma. We used the nationwide Flatiron Health EHR-derived deidentified database and first assessed the feasibility to emulate all key CM649 trial characteristics. To emulate 1L treatment, we identified patients who initiated all components of the respective treatment arms within 14 days of advanced/metastatic diagnosis (advDx) between 04/2021-12/2024. For patients with incomplete resection, advDx was assigned to the primary surgery date prior to treatment initiation. To avoid immortal person-time, we defined follow-up from the last date of the exposure ascertainment window (index date) to the date of all-cause mortality/censoring (overall survival). Key eligibility criteria and covariates were measured within 90 days before/on the index date and structurally assessed for potential missingness mechanisms. The main analysis involved multiple imputation of covariates which were used to compute propensity score overlap weights (OW). Cohort balance was assessed across the range of imputed datasets. We estimated marginal hazard ratios (HR) for each imputed and weighted dataset via proportional hazards models and used Rubin’s rule to pool estimates and compute confidence intervals (CI). We additionally emulated subgroup analyses by PD-L1 cut-offs. Sensitivity analyses employed alternative exposure definitions and tipping points for covariates deemed to be missing not at random. Agreement between emulated and CM649 results was assessed using three metrics: statistical significance (same directionality), estimate agreement (HR within CM649 CI) and standardized difference.

Results: We identified 624 eligible patients and OW led to optimal balance between treatment arms. The emulated effect (HR 0.74 [95% CI 0.59-0.92]) closely aligned with CM649 trial results (0.78 [0.71-0.87]) and agreed across all three metrics. Some differences were observed in subgroups leading to a total agreement of 71% across all analyses and metrics. Sensitivity analyses did not meaningfully alter results.

Conclusions: This principled target trial framework led to high agreement with results of CM649 and may be used to calibrate feasibility of RWE in oncology.