health · family: the regulator demands evidence that cannot exist
cleared bya ghost device
The FDA's 510(k) predicate creep problem: modern devices cleared on the backs of 1970s technology
Problem statement
The FDA's 510(k) clearance pathway — used for the large majority of new medical device authorizations (82% of the devices FDA cleared or approved in 2017) — allows devices to reach market by demonstrating "substantial equivalence" to a previously cleared predicate device rather than requiring independent clinical evaluation. Over successive generations, devices accumulate significant changes in technology, materials, and intended use while maintaining a chain of equivalence back to predicates that may be decades old or even recalled. This "predicate creep" means modern devices can bear little functional resemblance to the original safety baseline, yet they may never have undergone independent clinical testing. By FDA's own account in 2018, nearly 20% of 510(k)s were cleared based on a predicate more than 10 years old, and documented predicate networks reach back decades — a published analysis of one robotic surgical system traced an ancestry network of 2,618 device instances (each a citation of a device as a predicate, among 50 unique devices) across 15 product codes, with clearances dating back to 1981.
Why this matters
Roughly 3,000 510(k) clearances are issued annually (3,173 in 2017), covering categories including surgical instruments, cardiovascular devices, and AI/ML-based diagnostic tools. When a predicate device turns out to be unsafe, the devices cleared on it may perpetuate the flaw, and the safety concern can carry across multiple generations of devices. The propagation risk is measurable: among 510(k)-cleared devices subject to Class I recalls in 2017–2021, 44.1% had been authorized using predicates that themselves had Class I recall histories, and devices cleared on such predicates were 6.4 times more likely to suffer their own Class I recall (Kadakia et al., JAMA 2023). A 2023 Lancet Digital Health study (Muehlematter, Bluethgen & Vokinger) found that more than a third of the AI/ML-based devices it examined (those cleared via 510(k) between 2019 and 2021) originated from non-AI/ML-based predicate devices in the first generation — meaning the device they were found "substantially equivalent" to contained no AI/ML technology.
What’s been tried and why it hasn’t worked
On September 7, 2023, the FDA issued three draft guidances to strengthen the 510(k) program — covering predicate selection best practices, use of clinical data in 510(k) submissions, and evidentiary expectations for implant devices. However, these remain non-binding draft guidance and do not address the fundamental structural limitation: the 510(k) pathway is designed for incremental change, not for assessing whether cumulative changes have crossed a safety threshold. The predicate-selection draft guidance recommends choosing predicates that were cleared using well-established methods, meet or exceed expected safety and performance, have no unmitigated use- or design-related safety issues, and have no design-related recall — and the FDA has encouraged use of more modern predicates since 2018 — but these are recommendations, not requirements. Devices subject to Class I recalls (the most serious category) can still be used as predicates for new submissions, propagating the safety assumptions of a recalled device into new products — the Kadakia et al. JAMA 2023 study documented descendants cleared while their predicates' safety issues remained unresolved. The predicate-selection draft guidance sets out no mechanism for evaluating cumulative technological drift across a predicate chain.
What would unlock progress
A computable "predicate distance" metric — quantifying how far a proposed device has drifted from its oldest predicate in technology, materials, and intended use — could flag submissions where cumulative change exceeds a meaningful threshold. This would require a structured, machine-readable representation of device characteristics at each step in the predicate chain, combined with a decision rule for when clinical data should be required. Adjacent models exist in software dependency tracking and version control systems.
Entry points for student teams
A student team could build a network analysis tool that maps 510(k) predicate chains from the FDA's public 510(k) database (predicates are named in the posted summary documents, which the studies cited here found are often missing for older devices — those cleared before 2000 by one account, before the mid-2000s by the other), calculates chain length and age, and identifies high-risk chains (those involving recalled predicates, technology-class changes, or extreme predicate age). A more ambitious team could prototype a "predicate distance" scoring algorithm using device classification codes, product codes, and summary documents. Relevant skills include data science, network analysis, regulatory science, and health informatics.
Genes
Structural cousins: stuck the same way, in other fields
Sources
FDA Draft Guidance, "Best Practices for Selecting a Predicate Device to Support a Premarket Notification [510(k)] Submission" (September 2023), Federal Register, "Modernizing the Food and Drug Administration's Premarket Notification Program; Draft Guidances for Industry and Food and Drug Administration Staff; Availability," September 7, 2023, Kadakia, K. T., Dhruva, S. S., Caraballo, C., Ross, J. S. & Krumholz, H. M., "Use of Recalled Devices in New Device Authorizations Under the US Food and Drug Administration's 510(k) Pathway and Risk of Subsequent Recalls," JAMA 329(2), 136–143 (2023), Muehlematter, U. J., Bluethgen, C. & Vokinger, K. N., "FDA-cleared artificial intelligence and machine learning-based medical devices and their 510(k) predicate networks," The Lancet Digital Health 5(9), e618–e626 (2023), )00126-7/fulltext; Statement from FDA Commissioner Scott Gottlieb, M.D. and Jeff Shuren, M.D., Director of CDRH, on transformative new steps to modernize FDA's 510(k) program, November 26, 2018, full text via PR Newswire, Lefkovich, C. & Rothenberg, S., "Identification of predicate creep under the 510(k) process: A case study of a robotic surgical device," PLoS ONE 18(3): e0283442 (2023), Accessed 2026-08-20. go to source 1 ↗ go to source 2 ↗ go to source 3 ↗ go to source 4 ↗ go to source 5 ↗ go to source 6 ↗
Working notes
Notes made while writing and checking this brief, shown as written. They use working shorthand.
This brief draws on the September 2023 FDA draft guidance trilogy on 510(k) modernization, a 2023 Lancet Digital Health study examining AI/ML device predicate networks, and regulatory analyses by Hogan Lovells and Covington. The predicate creep problem intersects with health-ai-device-clinical-evidence-gap (AI devices relying on non-AI predicates) and health-device-real-world-evidence-gap (postmarket evidence infrastructure). Tagged `temporal:worsening` because the number of AI/ML devices entering via 510(k) is increasing, compounding the technology-class mismatch problem. Tagged `tractability:proof-of-concept` because the predicate chain data is publicly available and amenable to computational analysis.
Reconciliation 2026-08-20: The sole cited source was the FDA guidance-search landing page, not the document, and none of the body's statistics are in that guidance. Corrections: (1) The Source line now cites the actual draft guidance ("Best Practices for Selecting a Predicate Device to Support a Premarket Notification [510(k)] Submission," September 2023) at its real fda.gov URL, plus the September 7, 2023 Federal Register availability notice for the three-guidance package. (2) "Roughly 80% of new medical device authorizations" re-sourced and sharpened to the FDA's own figure — 3,173 devices cleared via 510(k) in 2017, "82 percent of the total devices cleared or approved" — from the November 26, 2018 Gottlieb/Shuren statement (FDA's original press page has been retired; the full text was verified via PR Newswire). The same statement carries the "nearly 20 percent of current 510(k)s are cleared based on a predicate that's more than 10 years old" figure, now attributed to FDA rather than left unsourced. (3) "3,000 to 4,000 clearances annually" tightened to ~3,000/year, matching the 2017 FDA figure (also consistent with Miller, Blanks & Yagi, Journal of Medical Systems 47:93 (2023): "approximately 3,000 medical devices" cleared per year, https://pmc.ncbi.nlm.nih.gov/articles/PMC10465388/); "high-risk categories" softened, since 510(k) covers moderate-risk device classes. (4) The recalled-predicate propagation claims now cite their actual study — Kadakia et al., JAMA 2023;329(2):136–143 (doi:10.1001/jama.2022.23279): 44.1% of Class-I-recalled 510(k) devices (2017–2021) used predicates with prior Class I recalls; risk ratio 6.40 (95% CI 3.59–11.40). (5) The Lancet Digital Health claim verified against Muehlematter & Vokinger 2023 (doi:10.1016/S2589-7500(23)00126-7): "more than a third" of AI/ML devices originated from non-AI/ML predicates in the first generation; authors now named. (6) "Some chains trace back to the 1970s" could not be verified as stated; replaced with the documented case — Lefkovich & Rothenberg, PLoS ONE 2023: the da Vinci Si predicate ancestry network of 2,618 device instances across 15 product codes with clearances dating back to 1981. (7) The quoted phrase "most recent, well-characterized predicate" is not in the guidance; replaced with the guidance's actual four best-practice factors. (8) "Most significant proposed changes in decades" was law-firm editorializing; removed. Of the analyses named above, the Covington analysis was re-verified (https://www.cov.com/en/news-and-insights/insights/2023/09/fda-proposes-significant-shift-to-510k-process-in-draft-guidance-on-best-practices-for-510k-predicate-selection); the Hogan Lovells piece was not re-located.
Reconciliation 2026-09-30: Whole write-up read against every source it names, all fetched 2026-09-30. Read in full: the FDA draft guidance (page at the Source-line address, and the PDF at https://www.fda.gov/media/171838/download); the Federal Register notice (at the Source-line address); Kadakia et al. (publisher page blocked; full text read at https://pmc.ncbi.nlm.nih.gov/articles/PMC9857464/); Muehlematter et al. (publisher page blocked; full text read from the Internet Archive copy of the publisher page, https://web.archive.org/web/20240729122836/https://www.thelancet.com/journals/landig/article/PIIS2589-7500(23)00126-7/fulltext, and the citation record from Europe PMC and Crossref); the 2018 Gottlieb and Shuren statement (PR Newswire, at the Source-line address); Lefkovich and Rothenberg (https://pmc.ncbi.nlm.nih.gov/articles/PMC10047502/); Miller, Blanks and Yagi (https://pmc.ncbi.nlm.nih.gov/articles/PMC10465388/); and the Covington analysis (at the address in the note above). Changes: (1) The Source line gave the Lancet Digital Health paper two authors and no pages. The record lists three authors (Muehlematter UJ, Bluethgen C, Vokinger KN) and pages e618–e626; the Source line and the body now carry all three names, and the pages. (2) The text said "more than a third" of AI/ML devices "originated from non-AI/ML predicate devices — meaning the 'substantially equivalent' device used an entirely different technology." The paper's summary says "more than a third", but its results give 93 of 285 devices (32.6%), which it calls "a third", cleared on the basis of non-AI/ML devices in the first generation; the devices studied were cleared between 2019 and 2021. The paper says such predicates did not contain AI/ML technology; it does not say the technology was entirely different. A second reader restored the paper's own headline wording, "more than a third", since that is the finding as the paper states it in its Summary and Conclusions; the results count (93 of 285, 32.6%) is recorded here rather than in the body. The sentence now keeps "more than a third", adds the cohort (devices cleared 2019–2021) and "in the first generation", and says the predicate contained no AI/ML technology. (3) "Predicates that may be decades old, withdrawn, or even recalled": no source says a withdrawn device can serve as a predicate, and the guidance and the Covington analysis both note that a device removed from the market at FDA's initiative cannot. "Withdrawn" was removed. (4) "They have never undergone independent clinical testing" was too strong. Lefkovich and Rothenberg say devices can reach market "without undergoing independent testing", and Kadakia et al. found that 5.8% of the recalled devices they studied did have premarket clinical testing. The sentence now says "may never have". (5) The "nearly 20%" figure is from the 2018 FDA statement and describes 510(k)s at that time; the sentence now says "in 2018". (6) "2,618 device instances": the paper counts each time a device is cited as a predicate, and says the network holds 50 unique devices. That is now stated. (7) The sentence saying that when a device in a chain fails, the chain's safety assumptions "are retroactively invalidated, yet downstream devices are not automatically recalled or re-evaluated" is not in any source. It now says what the sources say: a new device may perpetuate the flaws of its predicate (Lefkovich and Rothenberg), and safety concerns may propagate across multiple generations of devices (Kadakia et al.). (8) "No mechanism exists to evaluate cumulative technological drift across a predicate chain" is not stated in any source. The predicate-selection draft guidance, read in full, contains no such mechanism, and the sentence now says only that. (9) The first entry point called the predicate chains "publicly available ... from the FDA database". The database is public, but Lefkovich and Rothenberg and Kadakia et al. both report that predicates are identified only in posted summary documents, which are often missing for older devices (before 2000 in one paper, before the mid-2000s in the other). The entry point now says so, giving each paper's date rather than a blend of the two. (10) In the paragraph above, "accelerating" became "increasing", the word the Lancet Digital Health paper uses. Confirmed in the sources and left as written: 3,173 clearances and 82 percent in 2017, and FDA encouraging more modern predicates since 2018 (2018 statement); approximately 3,000 clearances a year (Miller et al.; Kadakia et al.); 44.1% (56 of the 127 recalled devices whose predicates could be identified) and the 6.40 risk ratio, recalled devices still permitted as predicates, and descendants cleared while a predicate's recall was still open (Kadakia et al.); 15 product codes and the 1981 clearance (Lefkovich and Rothenberg, Table 1); three draft guidances made available on September 7, 2023, and their subjects (Federal Register); the four best-practice factors and their non-binding status (guidance). All three guidances were still listed as draft on fda.gov on 2026-09-30. The citations for Kadakia et al., Lefkovich and Rothenberg, the Federal Register notice and the guidance match their records. A second reader went over these repairs the same day, opened every source again, and located the Hogan Lovells analysis named above (Hogan Lovells Engage, "FDA fills 510(k) guidance gaps with drafts on predicate devices, clinical data, implantables," 12 September 2023; the firm's site returns 403 to scripts, read from the Internet Archive copy); it summarizes the three draft guidances and bears on no figure in this write-up. The statement that the draft guidances "do not address the fundamental structural limitation" is this write-up's own reading; only the predicate-selection guidance was read in full. The "predicate distance" proposal and the comparison to software dependency tracking are this write-up's own suggestions and are not drawn from the sources. The long title's "1970s Technology" was left: the 2018 FDA statement says the 1976 law "established many of the predicate devices that served as the basis for 510(k) clearances during the last 40 years", but the oldest clearance documented in the body is from 1981. Genome tags untouched.