OBJECTIVES:

Heterogeneous response to an intervention is a common feature of survival patterns in randomized clinical trials (RCTs) of immuno-oncology therapies. This effect is usually inadequately represented by standard parametric models. Instead, it is desirable to employ parametric mixture models (PMMs), which represent the cohort as a combination of two latent subpopulations with distinct survival curves. However, classical PMMs are often too complex to parameterize reliably given the limited observations in RCT data.

METHODS:

We propose a Bayesian PMM (B-PMM) framework wherein a proportion of patients in the experimental arm are assumed to follow a survival pattern similar to that of the control arm, via specification of appropriate prior distributions. Vaguer priors are defined for the second subpopulation, representing patients who exhibit an improved response owing to the experimental treatment effect, and for the mixture fraction. We apply B-PMMs to digitized overall survival data for Pembrolizumab (versus Docetaxel) in advanced non-small cell lung cancer from the 5-year database lock of the phase 3 KEYNOTE-010 study.

RESULTS:

The B-PMM informed by control arm observations leads to clinically plausible survival predictions for responder and non-responder subpopulations of the experimental arm, with reasonable uncertainty. The model estimates that approximately 25% of patients belong to the responder subpopulation. When PMMs are fitted without leveraging additional information, a good visual fit to the observed data is obtained, but no meaningful inferences can be made on the subpopulations owing to high uncertainty arising from lack of information.

CONCLUSIONS:

The proposed B-PMM approach overcomes the problem of limited observations in RCT data to reliably infer the nature of heterogeneous response from flexible survival models. The key hypothesis that a proportion of patients do not respond to the experimental treatment leads to a model that is clinically interpretable and introduces the mixture (“responder”) fraction of the B-PMMs as a relevant predictive quantity.