A Niacinamide Rebuttal Letter, and the TriNetX Methodology Fight It's Part Of

A formal reply posted online September 7, 2026 in the Journal of the American Academy of Dermatology answers a comment from Li and colleagues on an October 2025 TriNetX cohort study by Yousef Salem and coauthors, which had reported lower rates of STEMI, peripheral vascular disease, and cardiac arrest among dermatology patients on oral niacinamide. The response, credited to Salem, Mahnoor Liaqat, Nadeen Gonna, Dina Zamil, Ferris Abu-Ghosh, Anwar Merie, Dimitri Cassimatis, and Ida Orengo, sits behind a ScienceDirect paywall, so the specific rebuttals it makes to Li et al. cannot be confirmed from open sources. What can be established is the evidentiary category both papers belong to, and that category is where the real story lives.

The original Salem study drew on TriNetX, a federated electronic health record network that lets researchers assemble matched patient cohorts through a query interface rather than a chart review. Patients with at least two non-melanoma skin cancers or bullous pemphigoid who started niacinamide at 500 mg or more were propensity-matched 1:1 against niacinamide-naive patients, yielding 3,231 patients per arm and relative risk reductions such as 0.547 for STEMI. That is retrospective, code-based association evidence: exposure and outcomes are pulled from ICD billing codes, not verified against dosing records or chart-confirmed diagnoses, and the design cannot rule out confounding by indication. Bullous pemphigoid patients started on niacinamide are frequently co-managed with tetracyclines as part of a long-standing combination regimen, so any cardiovascular signal attributed to niacinamide could partly reflect the comorbidities or comedications that led a physician to choose that regimen in the first place.

The Salem paper was explicit that it was answering a specific alarm: a 2024 Nature Medicine study by Ferrell and colleagues, built on prospective metabolomics cohorts totaling more than 4,000 patients, found that terminal niacin metabolites called 2PY and 4PY tracked with 1.6 to 2 times higher three-year rates of major adverse cardiac events, and showed in mouse vascular endothelium that 4PY induces VCAM-1 and leukocyte adhesion, a plausible mechanism for vessel inflammation. That paper is mechanistic and prospective; the TriNetX cohort studies that followed, including Salem et al., are retrospective and associational. A reassurance finding from the weaker evidence type does not resolve a mechanistic concern raised by the stronger one, it only adds a second, differently limited data point.

The Li comment and this response also land inside a much larger reckoning over how TriNetX studies get built and reported. An investigation published by Science in June 2026 tracked TriNetX-titled publications rising from 33 five years ago to roughly 2,700 in 2025, with most first authors being medical trainees using the platform's push-button cohort builder, and, since January 2026, a conversational AI query interface that critics say lowers the bar for producing superficially rigorous but bias-prone studies further still. A separate methodological paper in the European Journal of Epidemiology, published February 21, 2026 by Liu and colleagues, examined published TriNetX studies describing an index-event design the platform cannot actually configure, found 13 such papers, and then asked seven generative AI tools for index-event advice, six of which recommended at least one impossible strategy, pointing at AI-assisted methods writing as a likely source of the errors. That paper has already generated its own reply-and-response cycle, distinct from the Salem-Li exchange but in the same journal ecosystem and about the same underlying problem.

A related February 2026 letter argued that TriNetX-based authors should publish the platform's auto-generated cohort-definition summary alongside their papers, since ambiguity in something as basic as whether an exclusion criterion applies before or across the entire patient record can change results. None of this determines whether Li et al.'s specific critique of the niacinamide paper is correct, since the full text of both the comment and the response require institutional access neither disclosed publicly. But the pattern across this exchange and its neighbors is a shift already underway in the dermatology literature: retrospective TriNetX findings are being published fast, then re-litigated in comment-and-reply pairs almost as fast, over the same recurring flaws, immortal time bias, code-based exposure ascertainment, and index-event definitions the platform was never built to support in the form some papers claim.