P338: Comparing Static and Dynamic Bayesian Borrowing Approaches to Parametric Survival Models Informed by Historical Control Data
Poster Presenter
Jacqueline Vanderpuye-Orgle
Vice President
Parexel United States
Objectives
To compare survival extrapolations from static (SB) and dynamic (DB) Bayesian borrowing models informed by historical control data, applied to reconstructed interim data for pembrolizumab+ chemotherapy+/-bevacizumab (PEMBRO+CHEMO+/-BEV, vs CHEMO+/-BEV) in advanced cervical cancer from KEYNOTE-826.
Method
Survival data from KEYNOTE-826 (22 months median follow-up [fu]) were supplemented with data from the GOG-240 study (50 months maximum fu). Extrapolations from SB, DB, and vague prior models were compared and their accuracy judged against final data from KEYNOTE-826 (39 months median fu).
Results
In the SB model for the CHEMO+/-BEV arm, the preferred value for the fixed discounting parameter, chosen based on the Watanabe-Akaike Information Criterion (WAIC), was 0.95. Hence, the SB model borrowed very strongly from the historical control data that had been reweighted to reflect BEV use in KEYNOTE-826. The SB model yielded extrapolations for the control arm that had very low uncertainty (5-year survival: 8.6% [95% credible interval: 6.0-11.4%] SB vs 6.3% [1.3-12.7%] vague) and were accurate (2.5-year survival: 27.4% [24.3-30.6%] SB vs 31.6% observed). However, the SB model performed poorly for the PEMBRO+CHEMO+/-BEV arm. In the DB model, where the extent of borrowing from the historical study data is determined by multiple distributed commensurability parameters, a half-normal distribution was selected as the preferred hierarchical prior out of five candidates, based on the WAIC. The DB model demonstrated a less strong tendency to leverage historical data for the CHEMO+/-BEV arm compared to the SB model, and thus the predictions had higher uncertainty (5-year survival: 8.3% [3.6-13.3%] DB). For the PEMBRO+CHEMO+/-BEV arm, the limited borrowing led to projections that were less uncertain and more realistic than the uninformed model (5-year survival: 17.4% [10.0-24.6%] DB vs 14.3% [3.3-25.7%] vague), while remaining conservative (2.5-year survival: 41.3% [34.9-47.2%] DB vs 47.4% observed). Compared to the a priori borrowing propensity, the DB model learned to borrow more strongly for the shape and scale parameters of the generalized gamma distribution that was used to represent survival outcomes, while avoiding strong borrowing for the location parameter. Across the five candidate DB models employing alternative commensurate priors, survival estimates were consistent (e.g., range of 5-year survival estimates for PEMBRO+CHEMO+/-BEV: 16.7-18.0%) and all models had reasonable goodness-of-fit (range of WAIC <5).
Conclusion
Borrowing from historical control data by either static or dynamic approaches greatly improved confidence in forecasting the positive long-term treatment effect of PEMBRO+CHEMO+/-BEV that was evident in final data from KEYNOTE-826. The simpler SB model was sufficient to accurately extrapolate survival in the CHEMO+/-BEV arm with markedly reduced uncertainty and was in fact superior to DB for application to the control arm, since it capitalized on the high degree of commensurability between KEYNOTE-826 and the external data. The more sophisticated DB approach enabled limited and data-adaptive borrowing of selected features from the historical CHEMO+/-BEV data to inform projections for the experimental arm, and thus better captured the sustained benefit of PEMBRO+CHEMO+/-BEV compared to the uninformed model. Nonetheless, the estimates remained conservative, suggesting that the treatment waning assumption implicit in the DB model was overly pessimistic. Sensitivity analyses confirmed that survival estimates from DB were robust to the choice of commensurate prior. This study supports the use of Bayesian borrowing approaches leveraging historical control data from prior randomized trials to minimize uncertainty in health technology assessments and improve the reliability of longer-term treatment effect estimates, for ongoing studies in advanced cancers where overall survival data are immature.
