How FDA Device Classification Works
BetaFDA medical-device classification: candidate product codes, device class, the 21 CFR regulation behind them, the premarket pathway and cleared 510(k) predicate candidates.
Beta. This skill is live, but it has not been certified by Proofwork yet.
The approach
The skill connects to your LLM workspace (Claude, ChatGPT, or any MCP-compatible environment) as an MCP server. It returns reference data, evidence, and instructions for each step. Your LLM does the reasoning.
Proofwork is never in the inference path. The server handles deterministic work — lookups, retrieval, and evidence construction — while the host LLM makes each decision.
Workflow
- Find candidate product codes from the device description and intended use
find_fda_device_codes_payload - Pull the full record for the chosen code and read its pathway
get_fda_product_code_payload - Review the 510(k)s already cleared under that code
find_fda_predicate_candidates_payload
Available tools
Find candidate FDA product codes for a device description: class, submission type, exemption and implant flags, FDA's definition, examples
find_fda_device_codes_payload
Get the full record for one product code: 21 CFR identification text, premarket pathway, sibling codes, and what the record does not determine
get_fda_product_code_payload
List cleared 510(k)s under a product code as predicate candidates, with K-numbers and FDA database links — a research list, not an SE opinion
find_fda_predicate_candidates_payload
Where the data comes from
FDA Product Classification database (openFDA device/classification)
U.S. Food and Drug Administration
As of 2026-09-14
7,093 product codes with their device class, 21 CFR regulation and review panel.
https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfPCD/classification.cfmTerms: CC0 1.0 Universal (public domain dedication), per open.fda.gov/license
FDA 510(k) premarket notification database (openFDA device/510k)
U.S. Food and Drug Administration
As of 2026-09-14
Clearances decided on or before the corpus cutoff 2025-09-30, up to 40 per product code, newest first. Anything cleared after that date is not here: follow the FDA links for the live picture.
https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfmTerms: CC0 1.0 Universal (public domain dedication)
FDA PMA and De Novo databases (openFDA device/pma)
U.S. Food and Drug Administration
As of 2026-09-14
Original PMA approvals and granted De Novo requests up to 2025-09-30.
https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpma/pma.cfmTerms: CC0 1.0 Universal (public domain dedication)
FDA GUDID device identifier records (openFDA device/udi)
U.S. Food and Drug Administration / National Library of Medicine
As of 2026-09-15
Descriptions of marketed devices, used to build each product code's vocabulary and its example devices. GMDN term codes, names and definitions are licensed content of The GMDN Agency, are not covered by openFDA's CC0 dedication, and are not bundled in this skill.
https://accessgudid.nlm.nih.govTerms: CC0 1.0 Universal, excluding GMDN content (not bundled)
21 CFR Parts 862-892 (eCFR)
Office of the Federal Register (NARA) / Government Publishing Office
As of 2026-09-16
The identification and classification text of 2,167 regulation sections.
https://www.ecfr.gov/current/title-21Terms: A work of the U.S. Government; public domain.
21 U.S.C. 321(h) — the statutory definition of a device
Office of the Law Revision Counsel, U.S. House of Representatives
As of 2026-09-20
Returned verbatim so you can apply it yourself. The tools never decide whether a product is a medical device.
https://uscode.house.govTerms: A work of the U.S. Government; public domain.
The data ships as a point-in-time snapshot. Every tool response states the revision it answered from; confirm anything you act on against the current published source.
What this skill does not determine
- Whether a product is or is not a medical device. The statutory definition is returned verbatim for you to apply; the tools take no position.
- Substantial equivalence. The predicate list is material for your own research, not an SE argument and not an opinion on one.
- Predicate selection, De Novo strategy or PMA strategy.
- Whether a specific device actually fits a product code, or falls inside an exemption's limitations — pathway text is derived mechanically from FDA's structured fields.
- Combination products with a drug or biologic primary mode of action, SaMD risk categorisation beyond returning relevant codes, and IVD CLIA categorisation.
- Anything outside the United States: EU MDR and other jurisdictions are not covered.
What has been measured
446/515
Marketed devices whose registered product code was in the 25 candidates the find tool returns
n=515 devices described in their own GUDID records, first published 2026-03-01 to 2026-07-25 — after the 2025-09-30 corpus cutoff, and later than every case the retrieval parameters were tuned on. 86.6 per cent, 95 per cent Wilson interval 83.4 to 89.3. This measures whether the right code is in the list your assistant is shown, not whether it then picks it, and not whether FDA would agree with the labeler.
Where the code lands in the list (n=515)
Ranked first for 251 of 515 (48.7%, 95% CI 44.5-53.0). In the top 5 for 374 (72.6%), the top 10 for 408 (79.2%) and the top 25 for 446 (86.6%, 95% CI 83.4-89.3). Mean reciprocal rank 0.59. For 69 of 515 the registered code was not returned at any rank.
Every case is a real marketed device described in its own GUDID record, graded against the product code its labeler registered. Splits are temporal and disjoint: the bundled corpus stops at 2025-09-30, the dev split is GUDID publish dates up to 2026-02-28 and the test split is later still. Parameters were tuned on dev only; the test split was scored once. Dev reproduces at 88.0% in the top 25 (n=317), so there is no dev-to-test gap hiding in these numbers.
By device class (class I n=192, class II n=293)
Class II is the strong slice: 59.0% ranked first, 90.4% in the top 25 (95% CI 86.5-93.3), over 191 distinct product codes. Class I is weaker: 33.3% first, 81.8% in the top 25 (95% CI 75.7-86.6). Class I product codes are broader and their GUDID descriptions are more often catalogue copy than a description of the device.
Device class is FDA's own value from the product-code record, not the labeler's. Each slice is reported with the number of distinct product codes and companies behind it as well as its n.
Class III is NOT measured here — 7 devices, not a rate
The class III slice holds 12 cases, and no larger sample is available. The entire eligible pool for this window is 101 class III candidates from 7 product codes and 7 companies — 57 of them one company's intraocular lenses — and the 12 cases already cover all 7 codes and all 7 companies. Raising n would resample the same handful of devices and narrow the interval without adding evidence. Read it as 7 devices. If you are classifying a class III device, nothing on this page tells you how this skill performs on it.
Class III devices are rare in post-cutoff GUDID once the novelty filters have run, which is the constraint, not a sampling choice. Measuring this slice needs a different source of class III device descriptions.
Whether the company was already in the corpus (n=331 / n=184)
Devices from a company already in the pre-cutoff corpus: 50.4% first, 89.1% in the top 25. Companies absent from it: 45.6% first, 82.1% in the top 25. A new entrant is measurably harder, by about 7 points at 25.
Brand, description and company-plus-model of every case are checked against the corpus before selection, so a case is never answerable by exact match; only the company's prior presence is allowed to vary, and it is reported as a slice rather than filtered out.
Measured 2026-09-21 on cases kept back from tuning
(fda-device-gudid-retrieval,
test set).
It does not measure the assistant's final choice from the candidate list, substantial equivalence, or whether FDA would agree with the product code the labeler registered. Those need an agent-mode eval, which has not been run.
Known limitations
- Retrieval is lexical. A description using none of the vocabulary FDA, 510(k) holders or GUDID labelers use for a device type can miss the right product code; give the device type, technology and anatomical site rather than a brand name.
- The bundled 510(k), PMA and GUDID examples stop at 2025-09-30. Product codes created after that have little or no example vocabulary behind them.
- Product codes, classes and regulations change through reclassification orders and new De Novo codes. Confirm on FDA's Product Classification database before relying on a result.
- GUDID product codes are entered by labelers and are sometimes wrong or outdated: the example devices are evidence, not authority.
- GMDN terms are not included, by licence. Searching by a GMDN term name will not work.