What can nonstandard trial designs accomplish?
They can set up your clinical trial to answer multiple questions, explore multiple therapy options, assess multiple sub-populations, improve speed and sample size reductions, and provide various administrative efficiencies [such as Phase II-III trials, shared control groups, or improvements to recruiting and logistics]. As a result, costs are decreased, time is saved, and patients and therapies are better matched.
What have we achieved to date?
We have been working on mathematical techniques to make new clinical trial designs more legible. In Michael's PhD thesis (100-page PDF here - chapter 5 is the relevant part), he shows that for a large class of designs, if you can run large-scale simulations of a trial, you can turn those simulations into a bona fide proof about its Type I Error function [which is one of the most vexing issues for FDA reviewers].
There are a few other requirements, but for more details, please refer to Michael's talk delivered to the FDA on this topic in September 2021.
What about the FDA?
In Phases I and II, where the FDA does not regulate design, the use of new innovative trial designs has been exploding. The FDA’s leadership is encouraging progress in Phase III [notably Janet Woodcock has long promoted innovative trial designs. See her paper here].
These trials are genuinely complicated and very difficult to assess. Understandably, the FDA needs to be conservative in its approach.
The FDA CID program is doing several fascinating trials! We expect they are going to establish some really useful “case law” in the process.
What are we focusing on right now?
We are producing a basic proof-of-concept implementation, which will be published with a paper and an analysis of a Phase II-III design
What features are needed?
Fast Bayesian simulations with Laplace approximations.
Adaptive gridding.
Various small modifications to the core engine.
Modifying methods to evaluate bias of treatment effect estimators.
What about cloud computing?
Our paper will likely require it, since the scale of calculations required for Laplace approximations is likely to be huge.
After we submit the paper, the next goal is to host an online prototype.
It will help us identify future stumbling blocks and would be beneficial for users and future funding proposals.
What are our challenges?
Simulation problems are “too hard” at the necessary scale and can’t be done in realistic cases.
Software is too clunky and no one ever wants to use it.
How are we going to fund our work?
Grant applications from both inside and outside of academia.
Engaging with researchers who are already working in this space.
Philanthropy.