MDR Class I vs IIa/IIb: Software Classification Rules

  • MDR Annex VIII Rule 11 splits standalone software into three sub-rules: 11a diagnostic decisions, 11b physiological monitoring, 11c fallback.
  • Software providing diagnostic or therapeutic decision information defaults to Class IIa, and can escalate to IIb or III based on harm severity.
  • Under the older MDD, most standalone software was Class I and self-certified, but MDR Rule 11 removed that loophole.
  • Class I software needs only a self-written declaration of conformity, while Class IIa requires a notified body audit and an ISO 13485 QMS.
  • Keeping transport stateless, avoiding lossy transcoding, and defining strict module interfaces helps keep a data pipeline out of Class IIa scope.

Under the European Union Medical Device Regulation (MDR 2017/745), software classification depends on clinical impact. If your software provides data for diagnostic or therapeutic decisions, Rule 11 assigns Class IIa, IIb, or III. Software qualifies for Class I only when it has an intended medical purpose but does not drive diagnostic or treatment decisions. If your system merely transfers, stores, or converts imaging data without interpreting it, it does not qualify as a medical device at all.

This article provides engineering orientation, not legal advice, and a notified body or regulatory consultant decides the classification for a specific product. It explains the mechanics of MDR Annex VIII Rule 11 and how to draw clean architectural boundaries around cloud data pipelines.

MDR Class I vs IIa/IIb classification rules under Annex VIII

Annex VIII of the MDR sets the classification rules for all medical devices sold in the European Union. Chapter III, Section 6.3 contains Rule 11, which applies specifically to software.

Rule 11 divides standalone software into 4 risk tiers using a three-part structure:

  • Rule 11(a): Diagnostic and therapeutic decisions. Software intended to provide information used to take decisions with diagnosis or therapeutic purposes is Class IIa by default. If those decisions may cause death or irreversible health deterioration, it jumps to Class III. If they may cause serious deterioration or surgical intervention, it becomes Class IIb.
  • Rule 11(b): Physiological monitoring. Software intended to monitor physiological processes is Class IIa. If it monitors vital parameters where variations cause immediate danger, it is Class IIb.
  • Rule 11(c): The fallback rule. All other medical software falls into Class I.

Notice the wording in Rule 11(a). The software does not need to make the final clinical decision itself. It only needs to provide information that a clinician uses to take that decision.

Why Rule 11 upclassified legacy medical software

Under the older Medical Devices Directive (MDD 93/42/EEC), most standalone software fell into Class I. The MDD had no dedicated software classification rule. Manufacturers applied generic Rule 12 for non-invasive devices.

Class I devices allow self-certification. The manufacturer writes a declaration of conformity without an audit by an independent notified body.

The MDR removed that loophole. European regulators noticed that software errors in diagnostic tools create immediate patient risks. Rule 11 was designed to bring almost all clinical software under third-party oversight.

The shift caught many software teams by surprise. Moving from Class I to Class IIa or IIb changes your engineering overhead completely. You cannot self-certify a Class IIa device.

You must hire a notified body and complete an ISO 13485 Quality Management System (QMS) audit. You must also submit a technical dossier before placing the software on the market. In Europe, notified body review queues often take 12 to 18 months.

What pushes software from Class I into Class IIa or IIb?

The dividing line between Class I and higher classes is clinical intent. Specifically, does the software output inform, guide, or drive a doctor’s decision about a patient?

Consider what pushes an imaging application into Class IIa:

  • Automated measurements. Calculating tumor volume, cardiothoracic ratio, or bone density from DICOM pixel data.
  • Detection and segmentation. Drawing bounding boxes around lung nodules or segmenting brain lesions.
  • Clinical alerts. Generating an alarm when a laboratory value or imaging finding crosses a pathological threshold.
  • Image enhancement for diagnosis. Applying noise reduction, edge sharpening, or contrast adjustments intended to reveal pathology.

All these capabilities fall into Rule 11(a). Because they provide diagnostic information, they enter Class IIa at a minimum.

What escalates software from Class IIa to Class IIb? The severity of the patient condition and the urgency of the intervention. The MDCG 2019-11 guidance maps Rule 11 to the International Medical Device Regulators Forum (IMDRF) risk matrix.

Suppose an algorithm analyzes head CT scans to flag suspected intracranial hemorrhage. It also reorders the radiologist’s reading queue. If an erroneous result or delayed alert causes severe brain damage, that pushes the software into Class IIb. Software that calculates radiation therapy dose plans or guides surgical margins also lands in Class IIb or Class III.

Can medical software still qualify for Class I?

Yes, but the window is narrow. Under Rule 11(c), software is Class I only when it fulfills Article 2(1) without providing data for diagnosis or therapy.

MDCG 2021-24 gives concrete examples of software that remains in Class I:

  • Software that collates and formats patient data for administrative display without analyzing clinical parameters.
  • Applications that support general patient communication or appointment scheduling linked to medical records.
  • Basic rehabilitation apps that guide patients through standard physical exercises without tracking physiological response.
  • Simple calorie or nutritional logging apps intended for diabetic patients, provided they do not calculate insulin doses.

In medical imaging, a simple viewer that displays DICOM instances without processing, filtering, or measuring pixels might qualify under Class I. Even then, regulators examine your labeling closely. If your manual claims “primary diagnostic interpretation,” notified bodies argue the viewer directly informs diagnosis. That pushes the viewer into Class IIa.

When a data pipeline is not a medical device at all

Before you classify software, you must ask an earlier question: is the software a medical device at all?

Many engineers assume that any code handling medical images must be Class I or Class IIa. That is incorrect. Article 2(1) of the MDR states that a device must have a specific medical purpose, such as diagnosis, prevention, monitoring, prediction, or treatment.

MDCG 2019-11 explicitly addresses data communication systems:

“Communication systems are normally based on software for general purposes, and do not fall within the definition of a medical device.”

A cloud ingest pipeline that receives studies via DICOM C-STORE and DICOMweb STOW-RS is infrastructure. It transfers data without evaluating the image or reading clinical attributes. It behaves like a network router or file transfer utility.

Software that moves DICOM datasets from a clinic to cloud storage without altering clinical data sits entirely outside the MDR. It is non-device software. You do not need a notified body, CE mark, or Class I declaration for pure data transport.

To see how modern pipelines handle incoming transfers, read our guide on receiving STOW-RS deliveries.

How to keep transport infrastructure out of Class IIa scope

If you build an AI diagnostic platform, your machine learning models are undoubtedly Class IIa or IIb. But does your entire backend need to be classified as a medical device?

MDCG 2019-11 Section 3.3 provides the solution: software modularity.

If your product consists of multiple modules, you can separate medical device software from non-device software. To protect your data pipeline from being dragged into Class IIa, apply three architectural rules:

  1. Keep transport stateless and unopinionated. Your ingestion service should receive bytes, validate network protocols, and write to storage. Avoid adding clinical routing rules to the gateway. If your gateway parses DICOM tags to prioritize acute stroke studies over routine X-rays, auditors will treat it as part of the clinical workflow.

  2. Preserve data integrity without lossy transcoding. Transport pipelines must never alter pixel data. Applying lossy JPEG compression or downsampling to save bandwidth modifies diagnostic content. That change pulls the compression worker directly into medical device territory.

  3. Define strict interface boundaries. Isolate your core ingestion layer from downstream AI inference engines. Use dedicated queues, webhooks, or object storage boundaries. Document the interfaces and verify that failures return standard network errors rather than corrupted partial studies.

Adopting zero-retention architectures and dedicated DICOM endpoint configurations helps maintain this clean separation between communication plumbing and clinical applications. You can learn more about this distinction in our guide on medical device software qualification.

The operational trade-off: module boundaries vs integrated monoliths

Splitting a healthcare platform into medical and non-medical modules requires deliberate engineering.

A monolithic application is easier to build initially. You write a single backend that listens for DICOM, parses tags, runs an inference script, and displays results in a web portal. But when audit time comes, that monolith becomes a regulatory liability.

Because the diagnostic code is embedded in the monolith, your notified body must evaluate the entire codebase under ISO 13485 and IEC 62304. Every minor bug fix, dependency update, or schema migration in your ingestion layer requires formal change control, regression testing, and regulatory documentation.

We recommend decoupling the ingestion pipeline from your clinical engine from day one. You give up the convenience of shared memory and quick in-process function calls. In return, you keep your networking infrastructure out of the medical device boundary. Your ingestion layer can deploy updates rapidly, while your clinical modules follow formal medical device release cycles.

Dicomly builds infrastructure for healthcare software teams that receive DICOM studies from hospitals. Why we built Dicomly →

Sources

  1. Regulation (EU) 2017/745 on Medical Devices (MDR) (eur-lex.europa.eu)
  2. MDCG 2019-11 Guidance on Qualification and Classification of Software under MDR (health.ec.europa.eu)
  3. MDCG 2021-24 Guidance on Classification of Medical Devices (health.ec.europa.eu)

Start receiving DICOM today.

Early access. No credit card. First endpoint free.