AI-Assisted Aggregate Report Drafting

Generative AI drafting has become a normal part of PBRER preparation for the same reason it became normal in narrative writing (Module 5) and, to some extent, causality suggestion (Module 6): given well-structured input data, a model can assemble a fluent first-pass draft of predictable, pattern-following content far faster than a medical writer starting from a blank page. For a PBRER specifically, this shows up most usefully in the data summary and exposure sections — content that follows a fairly predictable structure once the underlying tabulated data is available.

The reliability drops sharply, though, exactly where the report matters most: the benefit-risk analysis conclusion, and the signal and risk evaluation section it depends on. A model summarising findings across potentially hundreds of individual cases can produce prose that reads as fluent and confident regardless of whether it has correctly weighted the actual balance of evidence — and a subtly mischaracterised signal status, understated in a generated draft, is a genuinely dangerous kind of error precisely because it doesn’t look wrong on the page. It reads as competent, well-organised prose; the flaw is in what it’s actually claiming, not how it’s phrased.

This is the PBRER-scale version of exactly the risk Module 5 flagged for individual narratives and Module 6 flagged for causality suggestions: a generated draft that sounds right is not the same as a generated draft that is right, and the gap between those two is hardest to spot precisely in the sections requiring the most integrative judgment — which, for a PBRER, is unambiguously the benefit-risk analysis and the signal evaluation feeding into it.

CIOMS Working Group XIV’s human-in-the-loop principle, which this course has applied consistently across MedDRA coding, narrative drafting, and causality assessment, applies here without modification, arguably at higher stakes than any single-case application: a qualified human medical writer and reviewer verifies every generated draft against the actual underlying evidence before a PBRER is finalised, with particular scrutiny on the benefit-risk conclusion — because unlike a single miscoded case, an inaccurate aggregate report conclusion can misinform a regulatory judgment affecting the entire patient population exposed to the product.