Skip to main content

Software as a Medical Device (SaMD)

Regulatory basis

MDR Art. 2(1) (software included in definition); MDR Annex VIII Rule 11 (classification); MDR Annex I §17 and §21 (GSPR for software). MDCG 2019-11 and MDCG 2021-24 provide detailed guidance. For IVD software: IVDR Annex VIII Rules 1–7.

Disclaimer

This site provides general information only and does not constitute legal or regulatory advice. Always consult the official regulation text and qualified regulatory and software engineering professionals.


What qualifies as SaMD under MDR/IVDR?

Software qualifies as a medical device if it is:

  1. Intended by the manufacturer for a medical purpose (diagnosis, monitoring, treatment, etc.), and
  2. Performs an action that goes beyond storage, archival, lossless compression, or simple search

Software that only stores, displays, or transmits data without any processing, interpretation, or transformation of that data into clinical information does not qualify as a medical device.

The MDCG 2019-11 qualification checklist is the primary tool for making this determination:

Software typeQualification
EHR administrative moduleNot a medical device
EHR clinical decision support giving patient-specific treatment recommendationsMedical device (MDR)
PACS image storage and retrievalNot a medical device
AI algorithm reading CT scans for pulmonary nodule detectionMedical device (MDR)
LIS routing samples to analysersNot an IVD
Algorithm interpreting CBC to flag possible haematological malignancyIVD (IVDR)

Classification — MDR Rule 11 summary

Device output / clinical consequenceClass
Death or irreversible deterioration if decision is wrongIII
Serious (reversible) deterioration or surgical intervention requiredIIb
Minor deterioration; no serious clinical consequenceIIa
General wellness; no specific medical decisionI

See Software & SaMD classification for the full decision tree and worked examples.


Technical documentation requirements for SaMD

Software-specific elements required in the technical documentation (MDR Annex II §6 and related):

Software description document

  • Software architecture diagram
  • Identification of SOUP (Software of Unknown Provenance) / OTS components
  • Data flows, interfaces, and integration points
  • Cybersecurity architecture

Software development lifecycle (EN 62304)

All Class IIa, IIb, and III software must follow a documented lifecycle per EN 62304:2006+A1:2015. The EN 62304 software safety classification (A, B, C) must be determined and documented:

EN 62304 Software ClassInjury riskMDR Class approximate mapping
Class ANo injury or non-serious injury possibleClass I
Class BNon-serious injury possibleClass IIa
Class CSerious injury or death possibleClass IIb or III

Software verification and validation (V&V)

  • Unit, integration, and system test protocols and reports
  • Regression testing records
  • Usability testing (IEC 62366-1)
  • Performance/stress testing as applicable

Cybersecurity

MDCG 2019-16 (and its updates) requires manufacturers to:

  • Conduct a cybersecurity risk analysis
  • Define the Minimum Security Baseline
  • Implement security by design (authentication, encryption, secure update mechanisms)
  • Plan for post-market cybersecurity monitoring and vulnerability disclosure

Software lifecycle — post-market considerations

Software changes are frequent and must be managed carefully:

  • Bug fixes: typically non-significant unless they correct a safety-critical defect
  • New features: assess whether they change the intended purpose or classification
  • Algorithm updates (for AI/ML): changes to trained models must be assessed — if the change alters the output in a clinically significant way, it may be a significant change requiring NB review
  • Platform updates: OS, cloud platform, or API changes may affect software performance and safety

MDCG 2020-3 provides guidance on what constitutes a significant change for software devices.


AI and machine learning in SaMD

AI/ML-based SaMD raises specific concerns addressed by MDCG and the EU AI Act:

  • Locked algorithms: performance is fixed at CE marking; updates are change-controlled
  • Adaptive algorithms: continuously learn from new data — under current guidance, the performance at time of CE marking must be specified and the learning process bounded
  • EU AI Act: many medical AI systems will also be regulated as "high-risk AI systems" under Regulation (EU) 2024/1689, creating parallel but overlapping obligations alongside MDR/IVDR


Official references

ReferenceDescription
MDR Art. 2(1)Software in medical device definition
MDR Annex VIII Rule 11Software classification
MDR Annex I §17, §21Software GSPR
EN 62304:2006+A1:2015Software lifecycle
IEC 62366-1:2015+A1:2020Usability engineering
MDCG 2019-11Software qualification
MDCG 2021-24MDR Rule 11 classification
MDCG 2019-16Cybersecurity guidance
EU AI Act (Reg. 2024/1689)AI system obligations