ServicesAI validation and evidence

AI validation and regulatory evidence for medical devices

An AI model can perform well in development and still lack the evidence required to support its intended medical use. For an AI-enabled medical device, validation needs to connect model performance to intended purpose, patient population, clinical workflow, risk, product claims and expected use conditions.

Fabola helps manufacturers determine what evidence should be generated and how that evidence should be connected to the medical-device technical documentation.

Who this is for

  • AI-enabled medical device developers
  • Machine-learning medical device software developers
  • Clinical decision-support systems
  • Diagnostic or predictive algorithms
  • AI-enabled IVD manufacturers
  • Products using large language models or generative AI in a medical context

The problem it solves

  • What dataset should be used for validation?
  • Does the validation population represent the intended population?
  • What should the reference standard be?
  • Which performance metrics matter clinically and regulatorily?
  • Do we need subgroup analysis?
  • How should model error be linked to clinical risk?
  • How should training, test and validation datasets be documented?
  • How much independent validation is necessary?
  • How should model changes be controlled after launch?

Required inputs

  • Intended purpose
  • Product architecture
  • AI or model description
  • Model inputs and outputs
  • Target population
  • Dataset descriptions
  • Existing validation protocol
  • Existing validation results
  • Clinical or performance claims
  • Risk-management information

What Fabola does

We define the validation objectives and reference-standard strategy, and assess whether the datasets are suitable, independent and representative of the intended population.

We set out the performance metrics, subgroup analyses, acceptance criteria and statistical analysis expected. We also address robustness, clinically relevant failure modes, human-AI interaction, evidence traceability and the change-control implications.

Deliverables

  • AI Validation Evidence Plan
  • Validation gap analysis
  • Proposed acceptance criteria
  • Dataset and population assessment
  • Evidence matrix
  • Validation protocol structure
  • Regulatory traceability map
  • Recommendations for technical documentation

Price

Focused validation review

SEK 59,000

Review of an existing validation approach, dataset plan or protocol against the intended medical use and regulatory evidence needs.

Standard AI validation strategy

SEK 79,000

One model, one intended purpose and one principal validation dataset. Includes evidence plan, metrics, acceptance criteria and traceability.

Complex AI validation strategy

SEK 99,000–119,000

Multiple models, materially different patient populations, several major datasets or multiple intended uses.

All prices are in SEK and exclude VAT.

Delivery time

Typical delivery: 2–4 weeks.

Completion point

The assignment is complete when the manufacturer has a defined and documented plan for generating or closing the evidence required to support the AI component within the intended medical use.

Exclusions

  • AI model development or training
  • Dataset generation
  • Software engineering
  • Execution of statistical validation unless separately agreed
  • Clinical-site operations

If further work is needed

Depending on the outcome, no further Fabola work may be required. Where additional work is identified, common next steps include execution of the validation plan, review of the resulting evidence, its integration into the clinical or performance evaluation, and Notified Body readiness.

Frequently asked questions

Is a high AUROC or accuracy score enough to validate a medical AI system?

Not by itself. Regulatory evidence needs to show that the selected performance metrics, population, reference standard, acceptance criteria and failure analysis support the intended medical use and associated clinical risks.

Do we need an external validation dataset?

The appropriate level of independence depends on the intended use, model, evidence base and risk. The validation strategy should explicitly justify dataset independence and the relationship between development and validation data.

Do we need subgroup analysis?

Where performance may vary across clinically relevant groups, subgroup analysis can be important for demonstrating representative and safe performance. The relevant subgroups should be defined from the intended population and risk analysis.

Can you review a protocol before we run the study?

Yes. Reviewing the protocol before execution is usually more valuable than finding design problems after data collection has finished.

Primary regulatory sources

Last reviewed: September 2026

Related services

EU AI Act + MDR / IVDR

How does the EU AI Act interact with the MDR or IVDR for our product?

Clinical and performance evidence

What evidence do we actually need to support our medical device or IVD?

Notified Body readiness

Are we ready to submit to a Notified Body?

View all services

Ready to define the work?

Tell us what the product is, where the project stands and the regulatory question you need answered.

We will confirm whether this service fits, or point you to a better starting point.

Discuss your AI validation question

The first conversation is free and carries no commitment.