P32: Oncology trial emulation using real-world electronic health record data: Results of the Coalition to Advance Real-World Evidence through Randomized Controlled Trial Emulation (CARE) initiative

2025
Levy N, Sheridan P, Campbell U, Lenis D, O'Doherty I, Estrin A, Gautam N, Zhen T, Belli A, Carrigan G, Chan A, Chen J, Chia V, Dhopeshwarkar N, Eckert JC, Fernandes L, Greshock J, Hendricks-Sturrup R, Huang J, Jiao X, Khosla S, Lunacsek O, McRoy L, Natanzon Y, Obviosa O, Pace ND, Pinheiro S, Rees M, Rider J, Rimawi M, Robinson T, Rodriguez-Watson C, Sangli C, Sarsour K, Schneeweiss S, Shapiro MA, Stewart M, Taylor A, Wang CK, Wang S, Zhang Y, Madsen AM
Read more

Methods

Following feasibility assessments of candidate RCTs in available U.S. data sources, we emulated the KEYNOTE-189 (metastatic NSCLC) trial of first-line pembrolizumab+chemotherapy vs. chemotherapy in two electronic health record datasets (DS1 and DS2) and the PALOMA-2 (advanced breast cancer) trial of first-line palbociclib+letrozole vs. letrozole in DS1. Trial entry criteria were applied, as feasible. Treatment status was based on first-line regimens (using data partner-defined line of therapy algorithms) initiated during a fixed ascertainment period. Inverse probability of treatment weighting was used to control baseline confounding. Cox proportional hazards models were used to estimate the primary outcome(s). RWD-based estimates were assessed for qualitative agreement (same direction/magnitude) with RCT results.

Results

The KEYNOTE-189 emulation real-world progression-free survival (rwPFS) hazard ratio (HR) in DS2 was of similar magnitude to the RCT finding, whereas the DS1 result did not demonstrate qualitative agreement [RCT: HR=0.52 (0.43, 0.64); DS2: HR=0.64 (0.47, 0.84); DS1: HR=0.81 (0.65, 1.00)]. KEYNOTE-189 emulation real-world overall survival estimates differed from the RCT results [RCT: 0.49 (0.38, 0.64), DS2: 0.89 (0.63, 1.29), DS1: 1.18 (0.95, 1.44)]. The PALOMA-2 emulation rwPFS HR also differed from RCT findings [RCT: HR=0.58 (0.46, 0.72); DS1: HR=0.84 (0.61, 1.23)].

Conclusions

Our results highlight that RWD oncology emulation conclusions depend on dataset features (e.g., care setting, therapy uptake, data completeness), treatment modality, and real-world clinical care. Future work should emphasize fit-for-purpose RWD selection and consideration of real-world care patterns to generate robust, interpretable real-world evidence.