IEC 62304 Edition 2: What Changes in Your V&V File

Engineer regrouping software items from three coloured columns into two on a laptop beside a wired infusion pump: IEC 62304 Edition 2 changes to the V&V file
In this post
  1. Three safety classes become two
  2. The scope quietly widens
  3. The AI and machine learning lifecycle is the genuinely new work
  4. How does IEC 62304 Edition 2 line up with QMSR and CSA?
  5. What to do this quarter, and what to defer
  6. Where Nopmark fits
  7. FAQ

Correction, 22 August 2026. This article originally said Edition 2 was in the final stage of standardization with publication targeted for August 2026. Both statements were wrong. Edition 2 is at committee draft stage. The August 2026 milestone is CD2, the second committee draft, and the Johner Institute reports that an FDIS is not expected until 2028. The current stage and timeline are set out in IEC 62304 Edition 2 Is Not Publishing in 2026. Everything below describes draft content and is not implementable against final clause numbers.

IEC 62304 has run on the same three safety classes since 2006. That run is ending, and Edition 2 is the largest revision the medical device software lifecycle standard has had in two decades. The transition is not a process rewrite. It is a re-mapping of your V&V file, and the team that builds its gap list before publication sets the terms of its own audit.

Three safety classes become two

The three-class system was a genuinely good idea, and it earned its long life. Class A for software that cannot contribute to a hazardous situation, Class B where injury is possible but not serious, Class C where serious injury or death is on the table. You classified once, read the required rigor off the class, and the middle class gave every borderline system a defensible home. Auditors knew the system, suppliers knew the system, and twenty years of V&V files are organized around it.

Edition 2 collapses that into two levels. The middle home is gone.

So the first instinct, that fewer classes must mean less work, gets the mechanism backwards. A sorting exercise with three buckets produces a moderate answer for the awkward cases, and most real products have awkward cases. A single dividing line produces no moderate answers. Every borderline software system now sits on one side of that line, and its position has to be defended in writing to an auditor who has read the same standard you have, which means the classification burden does not shrink at all: it converts into a justification burden, and justification is the more expensive of the two because it has to survive interrogation.

The question to start asking now is what happens to your existing Class B items. Nobody outside the committee can answer that yet with authority. What you can do today is build the inventory: every software item, its current class, the written rationale, and how much of that rationale actually exists on paper versus living in one engineer's head.

The scope quietly widens

Edition 1 scoped itself to software within regulated medical devices. Edition 2 is written for health software generally.

That single scoping decision pulls in products that never had to think about this standard: wellness platforms, clinical support tools, software that sits near patient care without ever being cleared as a device. If you build in that gray zone, the comfortable position that "we are not a device, so 62304 is not our problem" gets weaker in August. And if you are an established manufacturer, expect your SOUP suppliers and platform partners to start receiving questions they have never been asked, because your software of unknown provenance now comes from companies that may themselves fall inside the standard's new scope.

The AI and machine learning lifecycle is the genuinely new work

Everything above is re-mapping. This part is new obligation.

Edition 2 introduces a defined development lifecycle for AI and machine learning components. It also aligns the definition of harm with ISO 14971:2020, which brings property and environmental damage into view alongside patient harm, and it is being coordinated with IEC 62366-1 on usability and IEC 81001-5-1 on health software cybersecurity.

Most medtech software teams tend to have none of the AI lifecycle documented today. The model came from a data science group. The training data has no provenance record. The verification evidence is a validation-set accuracy number in a slide deck from last spring. That was survivable while auditors had no hook to hang the question on, and Edition 2 is the hook.

So if any shipping or planned product carries an ML component, that component needs a lifecycle story: where the data came from, how the model was verified, what the monitoring plan is, and who signs off on retraining. Reconstructing that evidence takes quarters, not weeks. It belongs at the top of the gap list.

How does IEC 62304 Edition 2 line up with QMSR and CSA?

None of this timing is a coincidence. The FDA's Quality Management System Regulation took effect on February 2, 2026, aligning 21 CFR 820 with ISO 13485:2016. A day later, on February 3, the FDA published its updated Computer Software Assurance final guidance, re-anchoring assurance expectations for production and quality system software to the QMSR.

Read together, everything points one direction: risk decides rigor. CSA tells you to scale assurance effort to process risk, with scripted testing reserved for the high-risk features. ISO 14971 gives risk its vocabulary. And IEC 62304 Edition 2 restructures the software lifecycle around the same idea, with fewer and therefore harder classification calls. If your V&V file still applies uniform rigor to every artifact because uniform felt safe, all three documents are now telling you the same thing about it.

What to do this quarter, and what to defer

Do now:

  • Inventory and classification map: list every software item with its current class and locate the written rationale. Anything defended only verbally is a gap.
  • ML component census: find every AI or machine learning piece in shipping and planned products, and for each one ask whether any lifecycle documentation exists at all. This is the census most teams have never run, and it is the slowest gap to close, because the evidence usually has to be reconstructed from the data science team's memory and old experiment logs.
  • Scope check: if you ship health-adjacent software that never claimed device status, have regulatory affairs read the Edition 2 scope language the week it publishes.

Defer until the ink is dry:

  • Rewriting SOPs. Procedures written against draft text get written twice.
  • Reclassifying anything. The two-level mapping rules belong to the final standard, not to consultant previews of it.
  • Tool purchases. Vendors will sell Edition 2 readiness hard this fall, and a gap list you built yourself beats a compliance dashboard bought in a panic.

Where Nopmark fits

Nopmark's IEC 62304 course teaches the software lifecycle standard as engineering discipline: classification, lifecycle documentation, and the V&V evidence that survives an audit. If Edition 2 is about to land on your desk, the fastest preparation is knowing Edition 1 deeply, because the transition work is a comparison exercise and you can only compare against what you understand. And when Edition 2 publishes later this year, the course will be updated to cover the changes at no extra cost to anyone already enrolled, so learning Edition 1 now is not something you pay for twice. The course sits alongside the software validation and risk management courses in the medical device bundle.

FAQ

When does IEC 62304 Edition 2 publish?

No published date exists. Edition 2 is at committee draft stage, and the milestone in August 2026 is CD2, the second committee draft. The Johner Institute reports that an FDIS is not expected until 2028, with publication possibly in 2029, and IEC's forecast publication date is listed as October 2028. None of those dates is a commitment. Until the final text is out, treat every summary, including this one, as provisional, and verify specifics against the published standard before changing your quality system.

Will my Class B software need reclassification?

It depends on the final two-level mapping rules, which are not public yet, so be suspicious of anyone selling you a definite answer today. What you can do now is collect the written rationale for each Class B item, so the comparison takes days instead of months when the text lands.

Does Edition 2 apply if my product is not a regulated medical device?

The scope expands from regulated device software toward health software generally. If your product sits near patient care without device clearance, have regulatory affairs read the final scope language before assuming you are outside it.

What is the genuinely new obligation in Edition 2?

The defined lifecycle for AI and machine learning components. Classification changes are re-mapping work on evidence you already have. The AI lifecycle demands evidence most teams never collected: data provenance, model verification, monitoring, and retraining control.

How does Edition 2 relate to the FDA's CSA guidance?

They share one principle: risk decides rigor. CSA scales assurance activity to process risk for production and quality system software, and Edition 2 restructures device software classification around the same risk logic. A V&V file built on risk-tiered rigor is preparing for both at once.

Ready to walk into the Edition 2 transition with your gap list already written? The IEC 62304 course from Nopmark covers classification, lifecycle documentation, and audit-ready V&V evidence, taught by accredited instructors: https://www.nopmark.com/courses/iec62304

Keep reading