CNS trials reach for Bayesian neurophysiology while blinding quietly leaks

Three items dropped this week that, taken together, describe a CNS pharmacotherapy beat reaching for methodological sophistication on the front end while still leaving the foundations — blinding integrity, endpoint definition, multiplicity — to be sorted out later.

The headline protocol is NCT07227948, a 75-patient randomised, double-blind, placebo-controlled trial of once-weekly semaglutide (0.25–1.0 mg, dose-escalating) as an adjunct to weekly CBT in treatment-seeking adults with cocaine use disorder. There are no FDA-approved pharmacotherapies for CUD, so any credible signal here matters. The design is ambitious in the way small addiction trials increasingly are: four co-primary endpoints spanning late positive potential (LPP) cue reactivity on EEG, behavioural-economics cocaine demand, the Cocaine Craving Questionnaire, and use confirmed by urine drug screens, all analysed in a Bayesian intention-to-treat framework.

That is a lot of methodological surface area for 75 patients. The protocol abstract does not specify the prior distributions, the posterior probability thresholds that would define a “positive” result, interim rules, or any multiplicity strategy across the four primaries. None of those omissions is fatal at protocol-publication stage — they may well live in the SAP — but they are exactly the details that determine whether this trial produces an interpretable signal or a Rorschach test. Bayesian small-N designs in reward neuroscience are precisely the setting where prior elicitation and decision calibration do the heavy lifting; the active methodological frontier on robust mixture priors and Bayesian equipoise calibration speaks directly to both gaps, and neither is mentioned.

Blinding is the unspoken problem

Semaglutide has a recognisable GI side-effect profile. In a treatment-seeking population on weekly subcutaneous injections, functional unblinding is a real risk, and three of the four primary endpoints (LPP, demand, craving) are either subjective or expectancy-sensitive. The protocol does not describe an active placebo or a blinding-assessment endpoint, which lands it squarely in the territory mapped this week by a Journal of Clinical Epidemiology scoping review on active-placebo design. The review’s two findings are unflattering and overdue: pharmacological RCTs that use active placebos rarely document how the control was developed, and the term “active placebo” itself is used inconsistently enough that meta-analyses assuming control-arm comparability are doing more imputation than they admit.

This is the same journal that, in April, retracted the 2025 paper claiming treatment effects in pharmacological RCTs are “mainly due to placebo.” The scoping review is a more defensible, descriptive contribution from an adjacent research stream — but it lands in a literature that is actively re-litigating what placebo controls even mean. For CNS trials with subjective primaries and side-effect-heavy actives, that debate is not academic.

EMA opens the smoking-cessation guideline

The third item is the EMA’s concept paper, published 29 May, beginning the formal revision of its smoking-cessation development guideline. Concept papers are non-binding scoping documents, and the substantive methodology will only surface in a draft 12–24 months out. But the drivers named are the ones biometrics leads should expect: alignment with ICH E9(R1) estimands (the existing guideline predates it), scope extension to non-combustible nicotine products (vapes, pouches, heated tobacco), and — though the concept paper does not commit — almost certainly a reconsideration of continuous abstinence versus point prevalence as primary, with biochemical verification and intercurrent-event handling (relapse, product switching, behavioural-support initiation) defined explicitly rather than left to sponsor discretion.

The through-line across all three items is that CNS pharmacotherapy trials are being asked to carry more methodological weight — estimands, Bayesian inference, neurophysiological biomarkers, multi-product scope — on top of design fundamentals that the field has not actually closed out. A multi-modal Bayesian primary endpoint stack is a defensible bet in a treatment-gap indication. It is also a structure where unblinding, an unspecified prior, and unaddressed multiplicity can quietly conspire to produce a posterior that looks decisive and is not.

Protocol read: Treat any small-N Bayesian CNS protocol that does not pre-specify priors, posterior decision thresholds, and a blinding-integrity check as not yet reviewable, however sophisticated the endpoint stack looks on the cover page.

What to do now:

• For any small-N Bayesian CNS protocol in-house, require the SAP to pre-specify priors, posterior decision thresholds, and a multiplicity strategy across co-primaries before database lock of the pilot cohort.
• Add a blinding-integrity assessment (post-treatment guess plus side-effect attribution) to GLP-1 and other side-effect-heavy CNS protocols, and pre-specify it as a sensitivity covariate.
• Flag the EMA smoking-cessation consultation for the regulatory calendar and begin an estimands gap assessment on any active nicotine-dependence programme now, not at draft-guideline publication.