Protecting the Label: Abstention, Prespecification, and Alpha Allocation for AI Derived Subgroups in Confirmatory Trials

Abstract Phase 3 programs fail when a treatment that benefits some patients is tested in a population where many do not respond, and the overall effect is diluted below significance. A subgroup found after unblinding cannot rescue such a program, because no type I error was allocated to it. A subgroup can support a label only if it was defined before unblinding, by a rule an independent statistician can apply, and tested under a multiplicity strategy fixed in advance. Most classifiers produced by modern machine learning cannot be written into a protocol in that form, because membership depends on a fitted probability, distance, or similarity. We argue that abstention resolves this. A system that returns No Call where the evidence does not support an assignment can express membership as a compact logical rule over baseline values: deterministic, frozen before unblinding, and verifiable without access to the model. Model Derived Subgroups (MDS) of one to four variables, distilled from higher dimensional models, have this form. Carried into an all comers trial under weighted graphical multiple testing, an MDS gives the sponsor two paths to a label. If the overall population succeeds, the subgroup is tested at full alpha at no cost to the primary claim. If the overall population fails, a prespecified split allocation keeps the subgroup hypothesis testable, and a positive result can support a label in the population the rule defines. We set out the allocation choices, their costs, and what the rule must satisfy to define a label. (249 words)