48,840 supplement adverse event reports, screened four ways — and what a signal isn't

Dietary supplements are consumed by over half of US adults, and post-market safety surveillance of them remains limited [s1]. An analysis published in Food and Chemical Toxicology on 20 June works with the main record that does exist — the FDA's Center for Food Safety and Applied Nutrition Adverse Event Reporting System, or CAERS — and applies pharmacovigilance methods normally used on prescription drugs [s1].

Before any of the findings: the paper is unusually careful about what its own output means, and that caution is the most important thing to carry out of it.

What the analysis did

The study applied four disproportionality methods — the proportional reporting ratio, reporting odds ratio, Gamma Poisson Shrinker and Bayesian confidence propagation neural network — along with CUSUM temporal detection and demographic stratification, to 48,840 dietary supplement adverse event reports drawn from CAERS covering 2004 to 2025 [s1].

Disproportionality analysis does not measure risk. It asks whether a particular product-event combination appears in a reporting database more often than the database's overall pattern would predict. A signal means "this pairing stands out in the reports", not "this product caused this outcome".

The paper reinforces that. The analysis was conducted at the product-name level rather than the ingredient level, and the paper states directly that the signals it reports are product-name-level signals to be interpreted as flags for further investigation rather than confirmed ingredient-specific risks [s1].

The findings

Of 4,779 product-adverse event pairs, 3,017 — 63.1% — were detected by three or more of the four methods [s1]. After consolidating variant names for the most heavily fragmented products, the paper estimates this resolves to somewhere between 1,800 and 2,200 distinct product-level signals [s1].

Kratom dominated the highest-severity tier, with the Kratom-death pairing ranked highest overall (proportional reporting ratio 19.7, N = 178) [s1].

Hepatotoxicity signals clustered in herbal and botanical products and in weight-loss products [s1]. Among products that could be classified into a category, herbal and botanical supplements carried the highest serious outcome rate at 78.0%, with an adjusted odds ratio of 2.08 (95% CI 1.53-2.84) [s1].

The temporal method flagged 451 emerging signals, including what the paper describes as preliminary hepatic enzyme elevations for AG1 and Nutrafol over 2023-2025 [s1]. The word "preliminary" is the author's, and the product-name-level caveat above applies with full force to these two: an emerging report cluster is the beginning of an investigation, not its conclusion.

On statistical robustness, 97.3% of consensus signals survived false discovery rate correction, and the method recovered 81% of established international supplement safety signals [s1] — a sensitivity check indicating the approach finds known problems, which is the relevant test of whether its novel findings are worth pursuing.

The limitation the paper names about itself

The category-level results come with a caveat the author states plainly: 54% of reports fell into an "Other" category, which limits how far category-specific estimates generalise to all supplements within each nominal category [s1]. In other words, the 78.0% serious-outcome rate for herbal and botanical products describes the classifiable subset, and more than half the data was not classifiable.

The limitations of the database itself

CAERS is a voluntary, spontaneous reporting system, and every conclusion drawn from one inherits the same structural problems. Reporting is incomplete and uneven. Products that attract media attention attract reports. A product used by millions will generate more reports than one used by thousands regardless of its safety profile, because the database records numerators without denominators — there is no record of how many people took a product without incident.

This is why disproportionality analysis exists as a screening tool rather than a measurement tool, and why the paper's own conclusion is framed as regulatory prioritisation: it argues the findings support prioritising herbal, botanical and weight-loss supplement categories for scrutiny, and identifies emerging signals warranting continued monitoring [s1]. That is a statement about where regulators should look next, not about what any individual product does to any individual person.

The paper is also the work of a single author, and its outputs are screening signals that have not been independently confirmed against clinical case review.

What a reader can take from it

Not a verdict on any product. What the analysis supports is narrower and still useful: that within the reports Americans have voluntarily filed over two decades, serious outcomes concentrate in the herbal, botanical and weight-loss end of the market rather than being spread evenly across it [s1], and that liver-related events are the recurring theme in those clusters [s1].

What to watch

Whether the FDA acts on the categories the analysis prioritises. Whether any of the 451 emerging signals is followed by a case-series or clinical investigation that can distinguish a reporting artefact from a real hazard. And whether the product-name fragmentation problem the paper had to work around — the same product appearing under many names — is addressed in how CAERS records data, since it currently makes the database harder to analyse than it needs to be.

Sources

1. [s1] Disproportionality analysis of dietary supplement adverse events in the FDA CAERS database, 2004-2025. Food and Chemical Toxicology, published online 20 June 2026. https://doi.org/10.1016/j.fct.2026.116225