An emulation of the KEYNOTE-189 trial using electronic health records

As a pilot exercise to assess feasibility of using routine clinical practice data to duplicate randomized control trial (RCT) design, we attempted to emulate the KEYNOTE-189 RCT using real-world data (RWD), and report findings from comparing the results.

We used a US electronic health record (EHR) database linked with a tumor registry to retrospectively construct a study cohort. Consistent with the RCT, patients with metastatic non-squamous non-small cell lung cancer were included; patients with prior first-line treatment for metastatic disease, primary non-lung malignancies, and EGFR/ALK mutations were excluded. Mortality outcomes in initiators of pembrolizumab and chemotherapy vs. chemotherapy alone were compared in intent-to-treat analyses. The mortality hazard ratio and 12-month survival probabilities were estimated using Cox regression and the Kaplan-Meier estimator, respectively. Inverse probability of treatment weighting was used to control for potential baseline confounders.

There were 589 pembrolizumab initiators and 1,265 chemotherapy-only initiators. The mortality hazard ratio was 0.95 (95% CI: 0.78, 1.16) in the RWD study versus 0.49 (95% CI: 0.38, 0.64) in the RCT. The 12-month survival probabilities were 0.60 (95% CI: 0.54, 0.65) vs. 0.58 (95% CI: 0.55, 0.62) in the pembrolizumab and chemotherapy groups, respectively, compared with 0.69 (95% CI: 0.64, 0.74) and 0.49 (95% CI: 0.42, 0.56) in the RCT. In the RWD study, substantial treatment crossover was observed, and the results were robust to sensitivity analyses. A post-hoc subgroup analysis of de novo metastatic patients, identified using the linked tumor registry only, was aligned with the RCT.

Results of this EHR-based emulation were incongruous with those of the benchmark RCT, but consistent with other investigators’ emulation attempts. Treatment crossover and accuracy of captured diagnoses may explain these findings and should be considered in population selection and other features of future RWD study designs for oncology treatment questions.