Skip to main content

m36 · From prototype to product: FDA Class II (21 CFR 870.1425, product code DQK), 510(k), IEC 60601-1, IEC 62304, ISO 14971, EU MDR

Warning

Education only. Not for diagnosis, treatment or any clinical decision.

Important

Not legal or regulatory advice. Regulations, guidances and standards change. Every regulatory statement here was checked on 3 November 2026 against the source it cites; check the current version before relying on it.

Note

Draft, stages 2 to 4 of 6. The derivations, two interactive demos and a standards map are written. Each equation is checked by tests in tests/test_m36_verification.py, and the demos’ math in tests/test_part6_demo_math.py. The code lab is in place too. The EU route, questions and the remaining sections follow in later stages (plan: docs/outlines/m36-outline.md). Nothing here has had expert review yet.

In 60 seconds

A working prototype is the start of a product, not the end. A medical device also needs evidence that it is safe and does what it claims. That evidence includes basic electrical safety, software built under a defined life cycle, risks found and controlled, and verification and validation that trace back to the requirements. Much of that evidence is numbers from tests. How many units must pass? How sure can you be of a pass rate? How likely is a failure in the field? This module covers the route a mapping system takes to market and the arithmetic behind the evidence.

From the field

Planned in the outline (docs/outlines/m36-outline.md); written in a later stage.

Derivations

Symbols used throughout. \(R\) reliability: the probability that a unit passes a test or survives a period. \(C\) confidence level. \(n\) sample size. \(c\) number of failures allowed. \(x\) number of successes. \(\lambda\) failure rate (per hour). \(P_1\) probability that a hazardous situation occurs. \(P_2\) probability that the hazardous situation leads to harm.

Where the evidence goes

In the United States, a programmable diagnostic computer is a device that can be programmed to compute physiologic or blood-flow parameters from the output of electrodes, transducers or other measuring devices, including the programs supplied with it. It is a Class II device (U.S. Food and Drug Administration 2026a). Mapping systems are cleared under this regulation with product code DQK (TODO(source): cite primary FDA 510(k) records). Several requirements apply around it:

  • Quality system. Since 2 February 2026, FDA’s quality system rule for device makers, now called the Quality Management System Regulation, incorporates ISO 13485:2016 by reference (U.S. Food and Drug Administration 2026c).
  • Software. FDA’s 2023 guidance on software in premarket submissions asks for one of two documentation levels, Basic or Enhanced, chosen by the risk the device poses (U.S. Food and Drug Administration 2023).
  • Cybersecurity. Section 524B of the Federal Food, Drug, and Cosmetic Act sets cybersecurity requirements for “cyber devices”, including a software bill of materials (SBOM). FDA’s guidance on which devices qualify and what to submit was most recently issued on 3 February 2026 (U.S. Food and Drug Administration 2026b).
  • Changes to AI functions. For AI-enabled software functions, a predetermined change control plan (PCCP), authorized with the marketing submission, lets the manufacturer make the planned changes without a new submission for each one (U.S. Food and Drug Administration 2024).
  • Risk management. ISO 14971:2019 sets out a process to identify hazards, estimate and evaluate risks, control them and monitor the controls. It requires each manufacturer to set objective criteria for risk acceptability but does not set acceptable levels itself (International Organization for Standardization 2019).
  • Electrical equipment. IEC 60601-1 sets requirements for basic safety and essential performance that apply generally to medical electrical equipment; the current consolidated version is edition 3.2 (2020) (International Electrotechnical Commission 2020).
  • Software life cycle. IEC 62304 defines life-cycle requirements for the development and maintenance of medical device software. It does not cover validation and final release of the device, even when the device is entirely software (International Electrotechnical Commission 2015).

The EU route under the Medical Device Regulation is added in a later stage once its sources are checked.

Schematic

1 · The manufacturer
Quality management system: ISO 13485:2016, through the QMSR. Risk management over the whole life cycle: ISO 14971.

2 · Electrical equipment
Amplifiers, localization hardware, power: basic safety and essential performance, IEC 60601-1.

3 · Software
Life-cycle processes: IEC 62304. Documentation level in a US submission.

4 · Connectivity
Cybersecurity and the software bill of materials, section 524B.

5 · AI-enabled functions
Planned changes through a predetermined change control plan.

Figure 1: Which requirements bear on which part of a mapping system, in the United States. A teaching simplification, not a regulatory determination.

Assumptions

  • Schematic: no quantitative claim. Each block names only requirements whose sources were checked on 3 November 2026; the sources are cited in the list above.
  • Not a complete list. Standards for electromagnetic compatibility and usability, among others, also apply and will be added once their sources are checked; the EU route is not shown yet.

Key takeaways

  • In the US, a mapping system is a Class II programmable diagnostic computer, built under a quality system that now incorporates ISO 13485.
  • Software documentation, cybersecurity and risk management each need their own evidence, and most of it comes from tests.

Zero-failure verification

Suppose each unit (or each run of a test) passes independently with probability \(R\). The chance that all \(n\) pass is \(R^n\). If the true reliability were only \(R\), a run of \(n\) straight passes should be unlikely, at most \(1 - C\). Taking logarithms of \(R^n \le 1 - C\) (and remembering \(\ln R < 0\)) gives the smallest such sample:

\[ n = \left\lceil \frac{\ln(1 - C)}{\ln R} \right\rceil \tag{1}\]

To show 95% reliability with 95% confidence takes 59 passes in a row; 99% reliability at 95% confidence takes 299. One failure and the claim is not shown.

Key takeaways

  • Zero-failure sample size: \(\lceil \ln(1 - C)/\ln R \rceil\).
  • High reliability claims need large samples; decide \(R\) and \(C\) from the risk analysis before testing.

Allowing failures

A plan can allow up to \(c\) failures. The number of failures in \(n\) independent trials is binomial, so the plan must make passing unlikely if reliability were only \(R\):

\[ \sum_{k=0}^{c} \binom{n}{k} (1 - R)^k R^{\,n-k} \le 1 - C \tag{2}\]

Take the smallest \(n\) that satisfies it. For \(R\) = 0.95 and \(C\) = 0.95, allowing 0, 1 or 2 failures needs 59, 93 or 124 units. The plan becomes more forgiving of a stray failure, at the cost of more units.

Exact

Assumptions

  • Computes the smallest \(n\) in Equation 2; \(c = 0\) gives Equation 1. Binomial probabilities use a log-gamma function checked against SciPy in tests/test_part6_demo_math.py.
  • Assumes independent trials and one true reliability for every unit. The defaults R = C = 0.95 are example targets, which a real plan takes from its risk analysis.
  • What would change the result: variation between production lots, failures that are not independent, and test methods that can miss a failure.

Key takeaways

  • Choose the plan (\(n\), \(c\)) before testing, and write down why.
  • Allowing failures raises the sample size but makes the plan robust to a single unlucky result.

How sure is a pass rate?

After testing, you have \(x\) successes in \(n\) trials. The Clopper–Pearson interval gives exact confidence limits for the true success rate from quantiles of the beta distribution (Clopper and Pearson 1934):

\[ p_L = B^{-1}\!\left(\tfrac{\alpha}{2};\; x,\; n - x + 1\right), \qquad p_U = B^{-1}\!\left(1 - \tfrac{\alpha}{2};\; x + 1,\; n - x\right) \tag{3}\]

for a two-sided \((1 - \alpha)\) interval, with \(p_L = 0\) when \(x = 0\) and \(p_U = 1\) when \(x = n\). It covers the true rate at least \(1 - \alpha\) of the time. When all \(n\) trials pass, the one-sided lower bound at confidence \(C\) is \((1 - C)^{1/n}\): the same statement as the zero-failure formula above.

Key takeaways

  • Report a pass rate with its exact confidence limits, not just the point estimate.
  • “59 of 59 passed” means the success rate is at least 0.95 at 95% confidence, no more.

Reliability of a system

If a part fails at a constant rate \(\lambda\), the probability that it survives to time \(t\) is \(e^{-\lambda t}\). A system that fails when any of its parts fails (a series system) survives only if every part does, so the rates add:

\[ R(t) = e^{-\lambda t}, \qquad \lambda_{\text{sys}} = \sum_i \lambda_i, \qquad \text{MTBF} = \frac{1}{\lambda} \tag{4}\]

The constant-rate model ignores early failures and wear-out. It is a starting point for comparing designs, not a lifetime prediction.

Exact

Assumptions

  • Model (Equation 4): constant failure rates, parts in series, independent failures.
  • Failure rates and mission time are illustrative inputs, not data for any product.
  • What would change the result: early failures and wear-out (rates that change with time), redundancy, and repair.

Key takeaways

  • In series, failure rates add; the weakest parts dominate.
  • Mean time between failures (MTBF) is \(1/\lambda\) only under the constant-rate model.

Probability of harm

ISO 14971:2019 describes risk as a combination of the probability of occurrence of harm and the severity of that harm. It splits the probability into two steps (Annex C): a sequence of events leads to a hazardous situation with probability \(P_1\), and the hazardous situation leads to harm with probability \(P_2\) (International Organization for Standardization 2019):

\[ P(\text{harm}) = P_1\, P_2 \tag{5}\]

The split shows where a risk control acts. A control can make the hazardous situation less likely (lower \(P_1\)), make harm less likely once it occurs (lower \(P_2\)), or reduce the severity.

Key takeaways

  • Estimate \(P_1\) and \(P_2\) separately and record the basis for each.
  • Every risk control needs its own verification, which brings you back to the sample sizes above.

Clinical correlate

NotePending clinical review

Planned in the outline. Requires sign-off by an EP physician.

Visual summary

Planned in the outline.

How experts think

Planned in the outline.

Cheat sheet

Planned in the outline.

Interview questions

Planned in the outline.

Code lab

The lab rebuilds this module’s results in code, so you can see where each one comes from and break it on purpose. It runs in about 15 seconds.

Part What you do Data
A Sample sizes for several targets, with and without allowed failures, checked by simulating lots that are exactly as good as the target (Equation 1, Equation 2) synthetic
B Coverage of the Clopper–Pearson interval (Equation 3) against the normal approximation, which falls short near 100% synthetic
C Series reliability by Monte Carlo (Equation 4), and what a redundant pair of the weakest part buys synthetic
D A traceability matrix from this course’s own rules to its own tests; two requirements are verified another way, and the matrix says so this repository
E A small risk register: estimate \(P_1 P_2\) (Equation 5), apply controls, and check residual risk against an example acceptability matrix synthetic

Each part ends with checks: lines that pass with the code as given. If you change something and a check fails, that is information, not an error to hide.

Self-check

Planned in the outline.

Literature figures

Planned in the outline.

Further reading

Planned in the outline.

References

Clopper, C. J., and E. S. Pearson. 1934. “The Use of Confidence or Fiducial Limits Illustrated in the Case of the Binomial.” Biometrika 26 (4): 404–13. https://doi.org/10.1093/biomet/26.4.404.
International Electrotechnical Commission. 2015. IEC 62304:2006+AMD1:2015 CSV Medical Device Software: Software Life Cycle Processes. https://webstore.iec.ch/en/publication/22794.
International Electrotechnical Commission. 2020. IEC 60601-1:2005+AMD1:2012+AMD2:2020 CSV Medical Electrical Equipment, Part 1: General Requirements for Basic Safety and Essential Performance. 3.2. https://webstore.iec.ch/publication/67497.
International Organization for Standardization. 2019. ISO 14971:2019 Medical Devices: Application of Risk Management to Medical Devices. 3rd ed. https://www.iso.org/standard/72704.html.
U.S. Food and Drug Administration. 2023. Content of Premarket Submissions for Device Software Functions: Guidance for Industry and Food and Drug Administration Staff. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/content-premarket-submissions-device-software-functions.
U.S. Food and Drug Administration. 2024. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions; Guidance for Industry and Food and Drug Administration Staff; Availability. Federal Register 89(233):96260. https://www.govinfo.gov/content/pkg/FR-2024-12-04/pdf/2024-28361.pdf.
U.S. Food and Drug Administration. 2026a. 21 CFR 870.1425: Programmable Diagnostic Computer. Electronic Code of Federal Regulations. https://www.ecfr.gov/current/title-21/chapter-I/subchapter-H/part-870/subpart-B/section-870.1425.
U.S. Food and Drug Administration. 2026b. Cybersecurity in Medical Devices: Quality Management System Considerations and Content of Premarket Submissions. https://www.fda.gov/media/119933/download.
U.S. Food and Drug Administration. 2026c. Quality Management System Regulation (QMSR). https://www.fda.gov/medical-devices/postmarket-requirements-devices/quality-system-qs-regulationmedical-device-good-manufacturing-practices.