Data Quality in Regulatory Labeling: A Source of Recalls Nobody Tracks

How Often Labeling Drives a Recall: What the Data Says

Start with the numbers, and with their limits, because this subject attracts invented percentages. Three sources hold up to scrutiny.

The first is a peer-reviewed analysis of FDA drug recall data from 2012 through 2023, published in the Journal of Pharmaceutical and Biomedical Analysis in 2024. The authors found that on average 330 drug recalls were initiated each year, that a recall ran an average of 1.3 years from initiation to termination, and that each recall involved an average of 400,000 product units. The most frequent cause was impurities and contaminants at 37 percent, followed by control issues at 28 percent and labeling and packaging at 19 percent. Class I recalls, the category for problems that can cause serious harm or death, accounted for 14 percent of events.¹

The second is an earlier study in the Journal of the American Pharmacists Association covering January 2017 through September 2019. Of 195 drug recalls in that window, 29, or 14.9 percent, were attributed to labeling issues rather than product quality.² The two studies used different windows and different category definitions (one groups labeling with packaging, the other separates it), which is why the percentages differ. Both put labeling in the top three causes.

The third source is the FDA enforcement report data itself, which is public through the openFDA drug enforcement endpoint. That dataset draws on the FDA Recall Enterprise System, runs from 2004 to the present, and is updated weekly. openFDA is explicit that its results are unvalidated and that FDA does not update a recall’s status after classification, so treat the figures below as a description of the published record, not a validated statistic.³ Each enforcement report entry carries the recalling firm, the classification, the product description, the code information, and a “reason for recall” field that is descriptive text rather than a coded value.⁴

Because the reason is free text, counting labeling recalls means searching the text. Sakara Digital ran that search against the dataset as updated on 19 August 2026. Of 17,899 drug enforcement entries in the full dataset, 1,248 carried the word “labeling” in the reason-for-recall field. Of those, 101 were Class I, 804 were Class II, and 343 were Class III. Narrower phrases show where the harm concentrates: 84 entries cite a label error on declared strength (22 of them Class I), 164 cite a mix-up (55 Class I), and 16 cite a barcode.³

Two cautions on the openFDA figures. Enforcement entries are one row per product code per recall event, so a single event that covers several NDCs appears several times; the Class I potassium chloride overwrap event of 2025 appears as multiple rows for exactly that reason. And a text search for “labeling” undercounts, because some entries describe a wrong label without using the word. The point is not the precise share. The point is that the labeling category produces Class I events every year and that the record is detailed enough to learn from.

Source: Sakara Digital text search of the openFDA drug enforcement dataset by report date, data as updated 19 August 2026. Entries are per product code, not per event.³

The EU record says the same thing in a different vocabulary

The European picture comes from two documents twenty years apart. In January 2007 the EMEA (as EMA was then named) published its first analysis of quality defects reported on centrally authorized products. Of 65 defects reported in 2005, 20 resulted in recalls. When the defects were classified by category, “Product Information Literature” was the most common at 14 defects, or 23.1 percent of the total, ahead of deviations from the marketing authorization and ancillary material problems. The examples given were wrong shelf life, wrong batch number, and wrong blue box text. The paper notes that these defects normally lead to Class 2 recalls, that a wrong product name, active substance, or strength would normally be Class 1, and that the whole upstream manufacturing effort is wasted when a batch that is technically sound has to be recalled because of the printed information attached to it.⁵

The EMA’s annual report for 2024 shows the category has not gone away. The Agency received 395 suspected quality defect notifications in 2024, the highest in recent years. Of these, 221 were confirmed and led to batch recalls of nine centrally authorized products, none of them Class 1. The listed main reasons include product label issues, described as missing or incorrect batch numbers, and product packaging issues, described as mix-ups and damaged containers, alongside laboratory control and contamination causes.⁶ Marketing authorization holders are obliged to report any defect that could result in a recall, and where a defect presents a serious risk, national competent authorities inform each other through the rapid alert system.⁷

Five Alerts From 2025 and the Pattern Behind Them

Statistics describe the size of the problem. Individual alerts describe its mechanism. The MHRA publishes every medicines recall and defect notification as a numbered alert with the batch, the error, and the parts of the pack that were correct, which makes them unusually useful for root-cause thinking. Five from 2025 illustrate the range.

The wrong strength on one face of the carton

On 17 April 2025 the MHRA issued Class 2 recall EL(25)A/17 for one batch of lercanidipine hydrochloride 20 mg tablets. The packs were labeled as 10 mg on some sides of the carton. The correct strength, 20 mg, was printed on the top of the carton and on the blister strips.⁸ Two months later, on 12 June 2025, Class 2 recall EL(25)A/27 covered two parallel-imported batches of Inhixa 12,000 IU (120 mg)/0.8 mL enoxaparin syringes repackaged for the UK market. One side of the carton read 12,000 IU (20 mg)/0.8 mL. The other three sides, the patient leaflet, and the syringe labels were correct. Around 14 defective packs of ten syringes each had reached pharmacies.⁹

Both alerts share a feature that should trouble any data owner: the same fact, the strength, appeared correctly on most surfaces of the same pack and incorrectly on one. That is not a printing fault. A printing fault affects the whole run. It is a content fault, where one panel of the artwork carried a value that had diverged from the rest. In the enoxaparin case the difference between 20 mg and 120 mg is a single missing character, which is exactly the kind of error a human proofreader passes over when the surrounding text looks familiar, and exactly the kind of error a field-level comparison against master data catches in seconds.

The wrong product on the blister

On 4 March 2025, Class 2 recall EL(25)A/10 withdrew a batch of a retailer’s own-brand paracetamol 500 mg tablets because the foil blister inside the carton stated “Aspirin 300 mg Dispersible Tablets.” The manufacturer retrieved and inspected the batch and confirmed the tablets themselves were paracetamol.¹⁰ Here the carton was right and the primary pack was wrong. A blister foil is a printed packaging component with its own item code, its own artwork, and its own supplier. When a manufacturer runs several products on similar foil, the foil for one product can be issued against the packaging order for another, or the artwork for one foil can be approved from the wrong source text. Either way, the carton and the blister are two different data objects that were never reconciled against each other before release.

The barcode that says something else

The MHRA uses a fourth class, the Class 4 defect notification, for a “caution in use” where the product is not recalled. Two in 2025 concern barcodes. On 3 July 2025, notification EL(25)A/33 reported that the EAN barcode on cartons of one batch of simvastatin 10 mg tablets, when scanned, identified the product as paracetamol 250 mg/5 mL oral solution. The name, strength, and form printed on the carton were correct. Pharmacies were told not to use the batch in robotic or automated dispensing systems.¹¹ On 21 August 2025, notification EL(25)A/41 reported that the EAN/GTIN barcode on fexofenadine 120 mg tablets scanned as naratriptan 2.5 mg tablets. The affected list ran to more than twenty batches, the earliest first distributed in September 2023, which means the wrong barcode had been in the market for close to two years before it was caught.¹²

These two are the purest data errors in the set. Nothing a human reads on the carton was wrong. The machine-readable content was wrong, and it was wrong in a way that only a scanner would notice. In November 2025 England’s patient safety commissioner publicly warned that barcode and labeling errors on medicine packs can put the wrong medicine, dose, or formulation in front of a patient, after pharmacists reported recurring barcode data problems and missing 2D codes.¹³ Automated dispensing, ward scanning, and closed-loop medication administration all depend on the barcode being the truth. A pack whose barcode claims to be a different product defeats every one of those safeguards.

The same mechanism in the FDA record

The FDA enforcement data for 2025 reads the same way. A Class I entry from January 2025 records a label error on declared strength where some cartons were incorrectly labeled and the blister strips inside reflected the correct strength. A Class I entry from March 2025 records potassium chloride injection overwraps labeled 10 mEq that may have contained 20 mEq containers; the same event recurs in November. A Class I entry from July 2025 records cartons of cefazolin for injection found to contain vials labeled as penicillin G potassium. A Class II entry from August 2025 records a batch released with the expiry date printed as February 2027 instead of January 2027. A Class II entry from December 2025 records incorrect RFID tag labels applied by a repackaging firm.³ Older barcode entries include a 2023 Class III recall in which one of every forty unit-dose blisters carried a barcode that scanned a 125 mcg dose as 200 mcg, and a 2022 Class III recall in which a sodium chloride 23.4 percent product scanned as rocuronium bromide.³

Design guidance from the NHS Specialist Pharmacy Service, updated in December 2025, makes the clinical side of this explicit: the medicine name, strength, form, and route must be given prominence, and packs that look alike pose a confusion risk that design alone cannot remove.¹⁴ Good design is necessary. It is not a control on whether the prominent value is the right value.

The Data Behind a Label

To manage labeling as data, you first have to see it as data. A finished pack is the meeting point of at least six data flows, and most organizations manage them in six different places.

Regulatory content: SPL in the US, product information in the EU

In the United States, the content of labeling has been an electronic submission since the electronic labeling rule took effect on 8 June 2004. The requirements live in 21 CFR 314.50(l) for NDAs, 314.94(d) for ANDAs, 601.14(b) for BLAs, and 314.81(b) for annual reports. FDA adopted the HL7 Structured Product Labeling standard, an XML format built on the Clinical Document Architecture, and its April 2005 guidance on content of labeling explains why: SPL allows comparison of text section by section and comparison of specific drug information data elements, and it can be used to exchange the information needed for drug listing, which removes redundant data collection.¹⁵ Today SPL is the mechanism for establishment registration and drug listing, for labeling content, for REMS documents, and for lot distribution reports, and FDA’s online label repository holds more than 140,000 SPL records.¹⁶

SPL matters for this discussion because it is structured. The strength of an active ingredient is not a string in a paragraph; it is a numerator, a denominator, and units, and FDA’s validation rejects a file where those are missing or zero. The product data elements section carries the proprietary name, the established name, the dosage form, the route, the NDC product and package codes, and the package description. FDA’s own training material on SPL technical errors spells out the rules that bite on data quality: the product name fields must not contain the strength or dosage form; if a product appears in more than one place in the document, its generic name, dosage form, UNII, and strength must be identical everywhere; and if an NDC package code has been submitted before, the package description must be identical to the earlier submission.¹⁷ In other words, FDA already runs a consistency check on the SPL. What it cannot do is check the SPL against your artwork or your ERP.

Regulators do check the listing against the physical label when they look. A December 2025 FDA warning letter cited a firm for at least 23 drug listings in which the proprietary name, the non-proprietary name, or the product labeling did not match the active ingredient given in the SPL, a violation of 21 CFR 207.49(a)(4), alongside a practice of labeling active ingredients with a generic placeholder name instead of the drug name.¹⁸ And FDA has been inactivating drug listings that are not certified as current, not updated in over a year, or tied to an unregistered establishment, because stale listing data degrades the NDC Directory that pharmacies, payers, and other systems rely on.¹⁹

In the EU the equivalent content is the product information: the summary of product characteristics, the labeling, and the package leaflet, drafted against the QRD template. The structure is looser than SPL. It is a document with conventions, not a data model with validation rules, which is one reason the EMA’s electronic product information program discussed later matters.

Artwork: version control and approval

Artwork is where regulatory content becomes a printed component. Each carton, label, leaflet, and blister foil is a separate artwork file with its own version history, dieline, and approval record. In a mature operation the artwork brief is generated from approved text and the artwork proof is compared back to that text. In a typical operation the brief is assembled by a coordinator from the latest approved Word document, an email from regulatory affairs, and a spreadsheet of country variants, and the comparison is a visual proofread.

Translation management for multi-market packs

For centrally authorized products in the EU, the applicant must deliver translations of the full adopted English product information into all EU languages plus Icelandic and Norwegian within five days of the CHMP opinion, at Day 215. Member states review the translations by Day 229, and the applicant returns final versions incorporating those comments by Day 235. The EMA advises starting translation well before the opinion, around Day 180, and states that poor translations or poor implementation of member state comments can delay transmission to the European Commission. For Type II variations the same cycle runs on Day +5, Day +19, and Day +25.²⁰ That is a short window in which more than twenty parallel versions of a safety-critical text are created, reviewed, and frozen.

Multilingual packs add another layer. The QRD compilation of decisions on stylistic matters explains that abbreviations or short terms may be accepted for multilingual packs where the applicant has agreed them with the Agency, and that where the other languages differ from the English this must be reflected in the labeling annex.²¹ Research on leaflet translation has long noted that translation happens after the opinion and under time pressure, that in some markets the majority of leaflets are translated by pharmacists rather than trained translators, and that user testing of a leaflet is mandatory for only one language version, usually the English one.²² The translated versions are, in effect, released on trust.

Barcodes and serialization

Two barcode regimes coexist on a US pack. Under 21 CFR 201.25 every human prescription drug, biological product, and OTC drug covered by the rule must carry a linear barcode containing at least the NDC.²³ Under the Drug Supply Chain Security Act, each package must also carry a product identifier: the NDC combined with a unique alphanumeric serial number of up to 20 characters, plus the lot number and expiration date, in both human-readable form and a 2D data matrix barcode. Homogeneous cases may use a linear or 2D code. FDA’s June 2021 question-and-answer guidance sets out how the two requirements relate and how the human-readable portion should appear.²⁴

The GTIN encoded in that data matrix is governed by the GS1 Healthcare GTIN Allocation Rules. Those rules state that any change to the regulatory filing of a product, for example a change in formulation, usage, or concentration, leads to a new GTIN, that a declared change in net content such as the number of tablets in a pack requires a new GTIN, and that brand owners who hold the product specifications must properly allocate and maintain their GTINs so that trading partners can distinguish products.²⁵ A GTIN is therefore a master data attribute with a change rule of its own, and the simvastatin and fexofenadine notifications above are what it looks like when that attribute is wrong.

The master data underneath all of it

Every flow above draws on the same small set of facts, and the table below shows how many places each fact is held.

Four to six copies of each fact, each in a system with its own change process. No regulation requires those copies to be reconciled with one another. 21 CFR 211.125 requires that labeling issued for a batch be examined for identity and conformity to the master or batch production record and that issued, used, and returned quantities be reconciled, and 211.130 requires that packaging and labeling materials be examined for suitability and correctness before packaging begins.²⁶ EU GMP Chapter 5 requires that cut labels be stored in closed containers to avoid mix-ups, that printing of code numbers and expiry dates be checked and recorded, that electronic code readers and label counters be verified, that on-line control include checking whether overprinting is correct, and that any significant discrepancy in the reconciliation of printed materials be investigated before release.²⁷ All of that checks the physical component against the batch record. None of it asks whether the batch record, the artwork, and the SPL were built from the same data.

Where the Chain Breaks: Five Failure Points

Across the recall record and the audits Sakara Digital has seen in pharma and biotech packaging operations, the divergences cluster at five points.

Notice what is absent from that list: careless operators. The cGMP controls on line clearance, cut labels, and reconciliation are designed for operator error and they work. The failure points above are all upstream of the packaging line, in the offices where data is created and copied, and the packaging line controls cannot see them because the component issued to the line matches the batch record. The error was already in both.

Controls That Hold: One Source of Truth and a Three-Way Reconciliation

The control model has three parts. Each one is a data management practice, not a proofreading practice.

One source of truth for label content

Pick one system to hold the approved value of every label data element and make every other system a consumer of it. For most pharma and biotech companies the candidate is the regulatory information management system or the SPL authoring environment, because that is where the approved value is legally defined. The ERP item master then holds a copy with a documented feed, not an independently maintained value. The artwork system pulls the brief from the same source. The serialization system pulls the GTIN, NDC, and lot and expiry format from it. Translation memories are keyed to the product and presentation, not to the sentence, so a segment approved for one strength cannot be proposed for another.

The hard part is not the technology. It is the decision about ownership. Someone has to own the strength as a data element, with authority to reject a copy that disagrees, and that person is rarely appointed today because the strength “belongs” to regulatory, to supply chain, to artwork, and to quality all at once. The data governance work that assigns stewards to label data elements is the same work described for master data generally in Sakara Digital’s earlier articles, applied to the narrowest and highest-risk dataset in the company.

A three-way reconciliation before release

The single control that would have caught every alert reviewed above is a documented reconciliation of three artifacts before a packaging component is released for use: the submitted SPL or approved product information, the approved artwork, and the ERP master data for the material and batch. It is a data comparison, not a proofread, and it runs field by field.

Country-pack variants as data, not as files

Replace the variant spreadsheet with a variant table in the system of record: one row per market presentation, with the values that differ by market held as attributes and the values that must not differ held once. When a shared value changes, every variant inherits it and the artwork requests are generated from the table, which turns “which cartons did we miss” from a memory exercise into a query.

What a Labeling Data Quality Audit Looks Like

The reconciliation above is a transactional control that runs per change. A labeling data quality audit is the periodic check that the control exists and works, and that the copies have not drifted between changes. It is a data audit, so it is built from tests with pass and fail criteria rather than from interviews. Sakara Digital runs it as a sampled comparison across systems with a fixed set of tests.

Three practical notes. First, sample by risk, not at random: prioritize presentations with multiple strengths under one brand, multilingual packs, products handled by contract packagers or repackagers, and anything that changed in the last twelve months. Second, run the barcode decode test with a real scanner against a real proof or a retained pack, because barcode images embedded in artwork files can be replaced at the printer. Third, report the findings in the vocabulary of data, with the system, the field, and the divergence, so that the corrective action goes to the data owner rather than to the packaging supervisor.

The audit also produces a metric that most quality dashboards lack: the count of label data elements with more than one authoritative value. On the first run that number is rarely zero. On the third run it should be.

Labeling Change Control and the Tools: What They Fix and What They Relocate

Labeling changes arrive from many directions: safety labeling changes, new indications, manufacturing site changes, shelf-life updates, pack redesigns, market additions, and regulatory format changes. Each one enters a change control process that was built for manufacturing changes and treats the label as a document to be updated, not as a set of data elements to be propagated. The result is the disconnect in failure point one: the regulatory change closes when the submission is accepted, the artwork change closes when the proof is approved, and nothing checks that the two closed on the same content.

Change control that follows the data

The adjustment is small in concept and large in practice. A labeling change request should list the data elements it changes, not just the documents. For each element, the request should enumerate every system and every component that holds a copy, and the change cannot close until each copy shows the new value and the reconciliation report is attached. Regulatory affairs remains the owner of the approved value. Quality owns the closure criteria. The artwork team, the ERP data team, and the serialization team each own a task within the change rather than a separate change of their own. Organizations with a backlog of open labeling changes usually find that the backlog is a symptom of exactly this fragmentation, and the operating plan for clearing it in Sakara Digital’s earlier article applies directly.

Artwork management systems

Artwork management systems are the most widely deployed tool in this space, and they solve real problems. They give every component a controlled version history, route approvals with electronic signatures, enforce that the printer receives only the released version, and, in the better products, compare a new proof to the previous version pixel by pixel or text by text and highlight every difference. That last feature closes the failure mode where an unintended change slips into a revision.

What they do not do, unless the implementation is designed for it, is check the artwork against anything outside the artwork system. The comparison is proof against previous proof, or proof against the brief. If the brief was assembled from a superseded label text, the artwork system will faithfully version-control a wrong value. If the barcode image was supplied from the wrong list, the system will approve it with a valid signature. The artwork system relocates the problem from “we do not know which version is current” to “we know exactly which version is current and it was built from the wrong source.” That is progress, but only if the brief itself is generated from the system of record rather than typed in.

Structured content authoring

Structured content authoring tools treat label content as components with metadata, assembled into SPL, EU product information, and other formats from one set of reusable elements. The strength lives in one place and every document that mentions it renders from that element. This is the closest thing to a single source of truth that the market currently offers for regulatory text, and it directly addresses failure point two: there is no superseded Word file to approve from, because there is no Word file.

Two limits. First, the reusable component has to be scoped correctly. A component reused across two strengths with a variable for the number is efficient; a component copied and edited per strength recreates the problem inside the tool. Second, structured content systems usually end at the document boundary. The artwork brief, the ERP item master, and the serialization data are still outside, and the handoff to them is where the reconciliation has to happen. The tool moves the single source of truth into a system that can actually enforce it; it does not by itself connect that source to the packaging line.

Electronic product information in the EU

The EMA’s electronic product information program is worth watching for the same reason. The EU ePI Common Standard is a semi-structured, FHIR-based format for the SmPC, package leaflet, and labeling. A one-year pilot ran from July 2023 to August 2024 with Denmark, the Netherlands, Spain, and Sweden, in which companies created and managed ePIs through the EMA’s product lifecycle management portal and published them. In March 2026 the EMA issued a draft roadmap toward phased implementation aligned with the new pharmaceutical legislation and opened user acceptance testing to industry testers through 18 September 2026.²⁸ Once the EU product information is structured, the same field-level reconciliation described above becomes possible for EU packs, and the translation problem becomes tractable because each language version is a structured document with matching elements rather than a free-text Word file.

The 12-digit NDC as a test of the whole chain

The next forcing event is already scheduled. On 5 March 2026 FDA published a final rule standardizing the NDC at 12 digits in a single 6-4-2 format, with a 6-digit labeler code, a 4-digit product code, and a 2-digit package code, and revising the barcode requirements in 21 CFR 201.25 to permit additional data carriers that can encode the full 12-digit code.²⁹ The rule takes effect on 7 March 2033, with a three-year transition through 6 March 2036 to update labeling and deplete old stock, after which drugs labeled with a 10-digit NDC may be subject to regulatory action. FDA’s guidance is to begin updating labeling now by adding leading zeros to the affected segments.³⁰

Read that as a data quality exercise and it is the whole article in one project. Every NDC in the SPL, in the ERP, in the linear barcode, in the DSCSA product identifier, and in the human-readable text of every component and every market variant changes at once. Organizations with one source of truth and a working reconciliation will run it as a query and a batch of generated artwork requests. Organizations with six copies and a spreadsheet will run it as a multi-year manual program and will discover, in the middle of it, which copies were already wrong. The transition window is long. Using the years before it to build the reconciliation is the cheapest labeling data quality program most companies will ever fund.

Conclusion

Labeling errors are not a proofreading problem that better proofreading will solve. The recall record from the FDA and from European regulators shows the same shape year after year: a batch that is chemically and physically correct is withdrawn because one rendering of its data disagrees with the others. A carton face, a blister foil, a barcode, an expiry field. The regulations in force check the printed component against the batch record and they do that well. What no regulation requires, and what almost no company documents, is a check that the batch record, the artwork, and the submitted labeling were built from the same value. That gap is why these events are tracked as labeling defects and never as data failures, and it is why the same mechanisms produce Class I recalls every year without anyone connecting them.

Sakara Digital works with pharma and biotech organizations building this kind of data control: naming an owner for each label data element, establishing one system of record, putting a field-level reconciliation between SPL, artwork, and ERP in front of component release, and running the audit that proves it holds. If you are looking at your labeling change backlog, an artwork or structured content implementation, or the 12-digit NDC transition and want an independent perspective on where the data actually diverges, we are happy to have that conversation.

For Further Reading