Our comment to the FDA on AI-assisted statistical review

A focused job for AI, with the reviewer in charge

Adaptive dose-escalation designs like BOIN, mTPI, EWOC, and BLRM are now routine in oncology and first-in-human submissions. Each one carries study-specific choices: boundary probabilities, priors, simulation assumptions, and decision rules. Those choices are hard to verify quickly, and when they are set wrong they can expose participants to unnecessary toxicity or produce results that answer nothing.

So our comment proposes using AI for one task it handles well: reviewing the statistics in early-phase protocols before a trial starts. The system produces a structured, prioritized work product, and a human reviewer confirms, modifies, or dismisses each item on it.