After reading several posts on Bayesian methods I am curious if this approach may be used to design a good radiation dose de-escalation trial.
Background: Patients of cervical cancers receive the same dose to the entire pelvis. This stems from observations done several decades ago which established at below certain dose thresholds there is an increased risk of disease recurrence.
However, over my 10 years of experience in treating these cancers, I have rarely observed failures in the pelvis in areas where a lymph node was not present at the beginning. This form of failure is a very rare event and this brings us to the research question - can we reduce the dose to these areas safely (in order to reduce toxicity).
We estimate that less than 1 - 2 percent of patients will have a failure in this region, so a modest reduction in dose is possibly worth exploring. We certainly do not want this risk to go beyond 5 percent under any circumstances. (Time period of 3 - 5 years).
A simple binomial sample size (1 sided type I error 5% and power 80%) with a null proportion of 1% and an alternative proportion of 5% results in a sample size of 448 patients (224 in each arm).
Is there a way we can use Bayesian methods to either :
- Design a study with a smaller sample size?
- Design a study wherein we get the information whether reduced dose levels is safe earlier where toxicity is the primary endpoint for example.
If we are to do this we can reasonably think that at least 5 - 7 years would be required for accrual only. Note that for toxicity (acute) the sample size requirements are much more modest.
Note that similar dose de-escalation trials have been conducted in Head Neck Cancers but interestingly not as non-inferiority studies but as studies where toxicity reduction was the primary endpoint. That is of course rational as the primary aim of dose reduction is a reduction in toxicity. However, I am unsure as to how reliable we can consider their statement related to “similarity” of recurrence rates to be.
Case in point this study https://doi.org/10.1016/j.radonc.2016.08.009 (it seems to be open access).