Original research · Open data

The Source-Disclosure Gap: Starch Excipients in 1,360 EU Centrally-Authorised Medicines

An excipient census of European Medicines Agency product information leaflets. 218 of 1,360 products contain a starch or dextrin excipient; exactly half name no botanical source. No product declares wheat starch.

An excipient census of European Medicines Agency product information leaflets, and what it means for patients avoiding gluten

Abstract

Patients with celiac disease are routinely advised to check whether their medications contain gluten. Whether that check is possible from a product label has never been measured at scale for the European market.

We parsed excipient lists from the product information leaflets of 1,360 European Medicines Agency centrally-authorised medicines and enumerated every starch and dextrin term. No product in the corpus declares wheat starch as an excipient. However, 218 products (16.0%) contain a starch or dextrin excipient, and exactly half of those — 109 of 218 — do not name a botanical source anywhere in the leaflet. Sodium starch glycolate, present in 87 products under 17 distinct spellings, never names a source in any of them.

We further find that excipient nomenclature in EU leaflets is substantially unstandardised: 2,975 distinct excipient strings collapse to 1,842 after conservative normalisation, a fragmentation rate of 38.1%, with 327 substances appearing under multiple spellings.

The EU's mandatory wheat-starch declaration requirement appears to be working, in that no product declares wheat starch. But the finding a patient actually needs — this starch is not wheat — is available on the label for only half of starch-containing products, and the absence of a controlled vocabulary makes automated verification unreliable. We characterise this as a source-disclosure gap: a regime that successfully prevents undisclosed wheat starch, yet leaves the patient unable to confirm that from the leaflet alone.

1. Why this has not been measured

The question "does my medication contain gluten?" has a large clinical literature on the answer and almost none on the answerability.

Existing work is US-centric and proceeds by contacting manufacturers. The most recent and most rigorous example is Matson et al. (2025), which reviewed 689 NDCs across the top 100 pediatric medications dispensed in the United States, cross-referencing Lexicomp and contacting manufacturers directly. Their result is a useful benchmark: 50.2% of NDCs were confirmed gluten-free by the manufacturer, 22.6% confirmed gluten-free but with no guarantee against cross-contamination, 25.7% could not be confirmed at all, and 1.5% were confirmed to contain gluten.

That study measures what manufacturers will tell you when asked. It does not measure what the label tells you unprompted, and it does not cover Europe, where the regulatory regime is materially different: EMA guidance (CHMP/704219/2013) requires that when wheat starch is used as an excipient in a centrally-authorised product, the leaflet must declare it, and the product may be described as gluten-free only if residual gluten is below 20 ppm. US labelling of excipient sources is, by contrast, voluntary — the FDA's 2017 draft guidance on gluten in drug products remains draft.

The natural inference is that the EU regime is the safer one for celiac patients. That inference has never been tested against the actual corpus of EU leaflets. This paper tests it.

2. Method

2.1 Corpus

The EMA publishes product information leaflets for centrally-authorised medicines. We indexed 1,851 products and successfully parsed excipient sections from 1,360 (73.5%). Of the remainder, 447 failed parsing (leaflet structure variation, PDF-only publication, or section-heading variants the parser did not recognise) and 44 returned no document.

This is a coverage limitation, not a sample. The 1,360 parsed products are not a random subset — products with irregular leaflet structure are systematically more likely to have failed. Figures below are proportions within the parsed corpus and should not be extrapolated to the full EMA register without that caveat.

Extraction date: 30 April 2026. Centrally-authorised products only; the large population of nationally-authorised generics (BfArM, MHRA, AIFA, ANSM and others) is outside this corpus entirely and is discussed in §5.

2.2 Excipient normalisation

Raw excipient strings were lowercased; parenthetical content, hydration-state qualifiers (anhydrous, monohydrate, dihydrate), E-numbers and pH-adjustment annotations were stripped; remaining tokens were sorted to make comparison order-insensitive. This collapses iron oxide yellow and yellow iron oxide to a single key.

Normalisation is deliberately conservative. It does not resolve chemically equivalent but lexically unrelated names (hypromellose / hydroxypropyl methylcellulose). The fragmentation figure reported here is therefore a lower bound.

2.3 Starch identification

A product was counted as starch-containing if any excipient string matched starch, amylum, dextrin, or dextrate. A product was counted as source-disclosing if any of its starch strings named a botanical origin (maize, corn, potato, wheat, rice, tapioca, pea).

Note that maltodextrin is captured by this pattern. Maltodextrin is gluten-free even when wheat-derived, because processing removes the protein fraction; it is included in the census for completeness but is not a risk term. isomalt was explicitly excluded from gluten-term matching, as an unfiltered malt pattern otherwise dominates results with a chemically unrelated sugar alcohol.

All figures are reproducible by running analysis.py against ema_data.json.

3. Results

3.1 The corpus at a glance

3.2 What EU medicines are actually made of

To our knowledge this is the first published frequency table of excipients across the EU centrally-authorised register.

Lactose monohydrate at 15.0% is worth flagging for this patient population specifically: secondary lactose intolerance is common in untreated and newly-diagnosed celiac disease, so a substantial minority of EU medicines carry a second excipient concern for the same patients.

3.3 Nomenclature is not standardised

38.1% of distinct excipient strings are surface variants of another string in the same corpus. 327 substances appear under more than one spelling.

The worst cases are severe. Microcrystalline cellulose appears under 25 distinct strings; sodium hydroxide under 36; water for injections under 50. Much of this is driven by leaflets embedding cross-references directly into the excipient list — strings such as sodium starch glycolate (see section 2 'Itovebi contains lactose, sodium') are parsed as the excipient name because that is, structurally, where they appear.

This is not a cosmetic problem. It means that any patient, pharmacist, or software system searching an EU leaflet for a specific excipient by name will produce false negatives at a meaningful rate, and that the rate is unmeasurable without exactly this kind of normalisation pass.

3.4 The starch census

218 of 1,360 products (16.0%) contain a starch or dextrin excipient, expressed across 84 distinct strings.

Where a source is named, it is almost always maize: 107 products name maize, 2 name corn, 1 names pea. No product in the corpus names wheat, and no product names potato or rice.

3.5 Sodium starch glycolate is the blind spot

The single largest contributor to the undisclosed half is sodium starch glycolate, a common tablet disintegrant.

Sodium starch glycolate is most commonly manufactured from potato starch, and in practice is very unlikely to be wheat-derived. But that is pharmaceutical background knowledge, not information on the label. For 87 EU products — 6.4% of the parsed corpus — a patient reading the leaflet has no in-document basis for determining what the starch is.

3.6 Voluntary gluten-free annotation is vanishingly rare

Exactly two products in 1,360 voluntarily annotate a starch excipient as gluten-free:

• Truvada — Pregelatinised starch (gluten free)
• Combivir — sodium starch glycollate (gluten free)

Both are Gilead/ViiV antiretrovirals. 0.9% of starch-containing products take this step. The regulatory floor is, in practice, also the ceiling: manufacturers declare wheat starch when required and almost never volunteer reassurance when not.

4. Discussion

Correction, August 2026. A subsequent within-molecule study (The Vocabulary Beats the Mandate) found that for 86 medicines authorised in both markets, US labels name the starch's botanical source for 88.4% of molecules against 38.4% for EU leaflets, with 43 of 43 discordant pairs favouring the US. The comparison in §4.1 below is between different drug populations and therefore confounds regime with product mix. The characterisation of US label silence as uninformative is too harsh: US silence is rare. Read §4.1 with that correction.

4.1 Two regimes, two failure modes

Reading this alongside Matson et al. (2025) produces a sharper comparison than the usual "EU good, US bad" framing.

The two figures measure different things and should not be read as "the EU is worse." The US 25.7% is what remains unresolved after the patient has done the maximum available work. The EU 50.0% is what the label alone fails to resolve, before any manufacturer contact.

The substantive point is that the EU mandate transfers the burden of proof but does not eliminate the gap. Under a mandatory-declaration regime, the absence of a wheat declaration is genuinely informative in a way it is not in the US — it is evidence, not silence. But it is indirect evidence, and the patient is asked to trust a compliance process they cannot audit from the document in their hand.

4.2 The inference chain, stated explicitly

A patient in the EU reasoning from a leaflet that lists sodium starch glycolate with no source must chain the following:

1. Wheat starch must be declared under CHMP/704219/2013.
2. This product is centrally authorised and therefore in scope.
3. The leaflet does not declare wheat starch.
4. The manufacturer has complied.
5. Therefore this starch is not wheat.

Steps 1 through 3 are checkable. Step 4 is not. And our nomenclature finding weakens step 3 more than it first appears: with 38.1% string fragmentation, "the leaflet does not declare wheat starch" is a claim about a text search, and text searches over unstandardised vocabulary fail silently.

4.3 What would close the gap

A single, low-cost change would resolve most of this: require the botanical source of any starch-derived excipient, not only wheat. The 109 products already naming maize demonstrate that the field exists, is populated when manufacturers choose to, and costs nothing. Extending it from a wheat-triggered obligation to a universal one converts an inference chain into a fact on the page.

Secondarily, a controlled vocabulary for excipient names in leaflets — or even just a requirement that the excipient list contain excipient names and not embedded cross-references to other leaflet sections — would make automated verification tractable for the first time.

5. Limitations

We state these plainly, because a reader deciding whether to cite this should be able to see the edges.

• Coverage, not sample. 1,360 of 1,851 products parsed. Parse failures are non-random and correlate with irregular leaflet structure.
• Centrally-authorised only. Most mainstream European generics are nationally authorised and absent from this corpus entirely. Generic products are precisely where excipient variability is highest, so the true source-disclosure gap across all EU medicines is plausibly worse than 50%.
• Label-derived, not laboratory-derived. This study measures what leaflets disclose. It does not measure gluten content. No claim is made about any product's actual gluten status, and nothing here should be read as clearing or condemning a specific medicine.
• Cross-contamination is out of scope. Leaflets say nothing about shared manufacturing lines. Matson et al. found 22.6% of US NDCs were gluten-free-but-not-guaranteed on exactly this basis.
• Point-in-time. Extracted 30 April 2026. Formulations and leaflets change without notice.
• Fragmentation is a lower bound, per §2.2.
• Not clinically reviewed. The author is an independent developer, not a pharmacist or physician. This is a data-enumeration study, not clinical guidance. Patients should confirm any medication question with their pharmacist or prescriber.

6. Reproducibility

Everything in this paper regenerates from two files:

python3 analysis.py > results.json # every figure cited above

ema-starch-census-2026.csv contains one row per starch-containing product (n=218) with the verbatim excipient strings, source-disclosure flag, and normalised botanical source. Released CC0 — no attribution required, though it is appreciated and makes corrections easier to route.

Corrections: https://glutenscreen.org/about.html#corrections

References

Matson KL, et al. Gluten-Free Options for the Top 100 Pediatric Medications. Journal of Pediatric Pharmacology and Therapeutics 2025;30(3):367–371. doi:10.5863/JPPT-23-00084. PMCID: PMC12172667.

European Medicines Agency. Wheat starch containing gluten as an excipient in medicinal products for human use. CHMP/704219/2013.

US Food and Drug Administration. Gluten in Drug Products and Associated Labeling Recommendations. Draft guidance, December 2017.

King AR. Gluten Content of the Top 200 Medications. Hospital Pharmacy 2013;48(9):736–743.

Catassi C, et al. A prospective, double-blind, placebo-controlled trial to establish a safe gluten threshold for patients with coeliac disease. American Journal of Clinical Nutrition 2007;85(1):160–166.

Lizano-Díez I, et al. Gluten-derived excipients in analgesics, NSAIDs and antiasthmatics. 2024. PMC11958770.

Plogsted S. Gluten-free drug list. glutenfreedrugs.com.

Gluten Screen is a free, no-signup medication screening tool. No advertising, affiliate links, or manufacturer funding. Editorial policy: https://glutenscreen.org/about.html#editorial-policy