BRENNEN

Industries / Health / Medical devices

Design and prototyping

Turning a clinical need into a device design, building and testing early versions, and keeping a record of each decision that someone else can check.

5

Milestones achieved, from 1945 to 2018.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1945

    A kidney machine from scrap

    In the occupied Netherlands, the physician Willem Kolff built a machine to clean the blood of people whose kidneys had failed, from a bathtub, cellophane sausage casing, a sewing-machine motor, a water seal from a Ford Model T engine and aluminum from a crashed plane. The 16 patients he treated from 1943 onward died; on 18 September 1945 his 17th patient, Sofia Schafstadt, recovered, the first person anywhere to survive uremic coma (poisoning by the waste the kidneys should remove) from sudden kidney failure, by the account of Wieringa and colleagues.

    LessonA design proved on patients one failure at a time is no longer accepted; Kolff's biographers doubt a modern ethics committee would have let the work go on through 16 deaths.

    pmc.ncbi.nlm.nih.gov

    Credit

    Willem Kolff and his team at the hospital in Kampen

  2. 1957

    A pacemaker in four weeks

    A pacemaker sends small electric pulses that keep a slow heart beating. The ones in use were boxes on carts plugged into wall sockets, and a three-hour power failure in Minneapolis led to the death of a baby who depended on one. The heart surgeon C. Walton Lillehei of the University of Minnesota asked Earl Bakken, an electrical engineer whose small firm repaired hospital equipment, for a battery-powered pacemaker; Bakken built one with transistors within four weeks, by Medtronic's account.

    LessonThe need came from a named failure in use, and a builder working beside the surgeon turned it into a working device in weeks.

    medtronic.compmc.ncbi.nlm.nih.gov

    Credit

    Earl Bakken (Medtronic) and C. Walton Lillehei (University of Minnesota)

  3. 1997

    Design controls become required

    An FDA study published in January 1990 looked at device recalls from October 1983 to September 1989 and found that about 44% of the quality problems behind voluntary recalls were errors designed into the devices, in products from patient chair lifts to pacemakers and ventilators. The Safe Medical Devices Act of 1990 let the FDA add design controls to its rules for device makers, and from 1 June 1997 US makers had to plan each design, write down its requirements, verify and validate it, and keep its record in a design history file.

    LessonFaults built in at the design stage caused nearly as many recalls as all other causes together.

    govinfo.gov

    Credit

    US Congress and the FDA

  4. 2016

    Testing with real users

    The FDA's final guidance on human factors, issued on 3 February 2016 and revised in August 2026, asks makers to find the critical tasks, where a user's mistake could cause serious harm, and to test the finished design in realistic simulated use with at least 15 people from each group of intended users. It followed the FDA's 2010 review of infusion pumps, which push drugs and fluids into a vein, in which the agency wrote to makers that many reported problems were associated with design deficiencies.

    LessonIn the guidance's words, many use errors "cannot be anticipated until device use is simulated and observed".

    fda.govfda.gov

    Credit

    FDA's Center for Devices and Radiological Health

  5. 2018

    A trial run in software

    FDA scientists ran VICTRE, an imaging trial done wholly by computer, with 2,986 simulated patients, simulated machines for mammography and for breast tomosynthesis (a 3D form of mammography) and a computer in place of the radiologist. It found tomosynthesis better at finding lesions, as an earlier trial in women had, and was published in JAMA Network Open in November 2018. In November 2023 the FDA issued final guidance on how a maker shows that a computer model is credible enough to count as evidence, building on the ASME V&V 40 standard of 2018.

    LessonA simulated trial is judged against the real one it copies; the authors call the results consistent, and both found a clear gain for masses, but for calcifications (small calcium deposits) the computer found a small gain where the clinical trial had found a small loss within its margin of error.

    pmc.ncbi.nlm.nih.govfda.gov

    Credit

    Aldo Badano and colleagues at the FDA; the American Society of Mechanical Engineers (ASME)

Watch for

  1. Agents that check design files

    Companies that sell software for keeping device design records say their AI agents now read a device's design and quality files, flag requirements that have no verification test and risk controls with no evidence that they work, and draft sections of technical files for people to approve. Seal calls its agent Neil; Ketryx says its agents recommend changes, with experts reviewing each step. Neither page reports a test by anyone outside the company.

    Achieved whena party independent of the vendor tests such an agent on design files with known, planted gaps and publishes how many it found and how many of its flags were wrong.

    seal.runllm.ketryx.com

    Claimed by

    Seal and Ketryx, on product pages read in November 2026

Open

  1. 01

    Planted errors in design files

    A complete, public design file for a simple device that never touches a person, such as a bench pump that moves only water, with user needs, requirements, a risk analysis, drawings, verification tests and their results, all linked to one another. Then a second copy with 50 planted errors of known kinds (a requirement with no test, a test that checks the wrong limit, a hazard with no control, a drawing at the wrong revision), kept as an answer key for scoring AI reviewers and people.

    The brief
    Credit · unclaimed
  2. 02

    Critical tasks from instructions

    A public test set of at least 30 instructions for use (IFUs) of home-use devices, as published by their makers, each with a list of critical tasks (steps where a user's mistake could cause serious harm, in the FDA's sense) agreed by two independent human-factors specialists, and a script that scores how well an AI model's list matches theirs.

    The brief
    Credit · unclaimed

Open 01 · the brief

Planted errors in design files

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A complete, public design file for a simple device that never touches a person, such as a bench pump that moves only water, with user needs, requirements, a risk analysis, drawings, verification tests and their results, all linked to one another. Then a second copy with 50 planted errors of known kinds (a requirement with no test, a test that checks the wrong limit, a hazard with no control, a drawing at the wrong revision), kept as an answer key for scoring AI reviewers and people.

Why it is still open

In 2013 Larson, Hatcliff and Chalin published an open 60-page requirements document for a patient-controlled pain pump, written because research was "hampered by the lack of realistic non-proprietary development artifacts for medical devices". Requirements are only the start of a design file. In August 2026 Gangopadhyay, Adler, Liggesmeyer and Reich proposed scoring AI help with device safety files on coverage, correctness and traceability, but against case studies that are not public or not yet built. Searches of the web, Crossref and GitHub found open templates (OpenRegulatory's), student design files, an open-source tool that uses Claude for a first check of written procedures against FDA rules, and vendor claims with no outside test (see Watch for), but no public design file with test results and planted errors. The same gap holds for drug batch records, where Enveda's AI reviewer could be measured only against Enveda's own reviewers (see the Note "Seven more flagged").

arxiv.orgpmc.ncbi.nlm.nih.govgithub.com

Built with what exists

the open pain-pump requirements as a starting point; OpenRegulatory's templates for ISO 14971 risk files and ISO 13485 records; a syringe-pump prototype built from off-the-shelf parts and run only with water; Git to version every document; Claude and Gemini as reviewers, set to cite the requirement or clause behind each finding, as the code reviewer in the Note "Findings that cite notes" does, and run five times each, because one run can miss what another catches.

  1. First step

    Link the first 30 requirements of the open pain-pump document to tests you write, plant 10 errors, and count how many one model finds.

  2. Done looks like

    A public set of at least 300 pages holding 50 planted errors, each confirmed by two quality engineers who did not plant them. Three AI models and two human reviewers are scored on the share of planted errors they find and the share of their findings that are wrong, for single runs and for five runs combined. A script lets anyone score a new model.

Open 02 · the brief

Critical tasks from instructions

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test set of at least 30 instructions for use (IFUs) of home-use devices, as published by their makers, each with a list of critical tasks (steps where a user's mistake could cause serious harm, in the FDA's sense) agreed by two independent human-factors specialists, and a script that scores how well an AI model's list matches theirs.

Why it is still open

The FDA's human factors guidance asks makers to identify critical tasks early, from analysis and known problems, before testing with users. The closest study found, published in February 2026, had ChatGPT rate one radiotherapy planning program against 14 rules of thumb for usability: of 26 issues it raised, the developers judged nine worth acting on, all minor, and there was no answer key to score against. Searches of the web, Europe PMC and GitHub found no test set for AI-written critical-task lists.

pmc.ncbi.nlm.nih.govfda.gov

Built with what exists

IFUs that makers post on their websites; AI models that read long documents with figures, such as Claude or Gemini; two human-factors specialists as labelers; a rubric built from the guidance's sections on critical task identification and known use-related problems.

  1. First step

    Pick 10 home-use devices with public IFUs, such as nebulizers, blood pressure monitors and glucose meters, write the critical-task lists with one specialist, and compare one model's lists with yours.

  2. Done looks like

    30 IFUs labeled by two specialists, with their agreement reported. At least three models scored on the share of agreed critical tasks they list and the share of their tasks the specialists reject. The IFU links, labels and scores are public, so a new model can be scored in a day.