Skip to main content
Check a technology
Talk to us
AI RISK PROFILER METHODOLOGY

A structured, evidence-based method for assessing the risk of AI features in digital health

The AI Risk Profiler methodology (AiRP) gives evaluators, commissioners and suppliers a shared, reproducible way to work through: what an AI feature does, how it does it, which risks that activates, and what's in place to contain them.

How the method works
Be part of the validation

Informed by an international advisory group with representatives from...

APA
MHCC
NHS
APTA1
Mind
Ulster
UCL
IBM
DNV
APA Labs
HOW THE METHOD WORKS

Six steps, one feature at a time

The method assesses each discrete AI feature in a product as a unit. Select a step to see what it covers.

1

Clinical
context

What's at stake
How serious the consequences are if something goes wrong, and who receives the AI output. This shapes the entire risk profile.
2

Computational function

What the AI does
Whether it delivers information, generates content, classifies an input or makes a recommendation. A product can have several functions, and each is assessed.
3

Inference technique

How it does it
A large language model, a deep-learning classifier, a statistical model or a rules engine. Suppliers declare this through a structured self-declaration; an incomplete declaration is itself a finding.
4

Risk activation

Which risks occur
Which of the ten defined AI risks the combination of function and technique activates, and how significant each is given the clinical context.
5

Mitigations

What contains them
What the architecture does to reduce risk (for example, whether a professional must review output before it reaches the user), and what the governance programme does, from performance testing to post-market monitoring.
6

Gap analysis

What remains
Where unmitigated risks remain, and what they mean clinically and for data and security.
10 RISKS PROFILED

A defined set of AI risks relevant to health

Naming the risks explicitly is what lets different assessors reason about the same product in the same way. Select a risk to read what it means.
Omission: The AI leaves out information that matters to a safe or accurate outcome.
Hallucination: The AI generates content that is fluent and plausible but factually wrong or fabricated.
Overconfidence: The AI presents outputs with more certainty than the underlying evidence supports.
Misclassification: The AI assigns an input to the wrong category, for example mislabelling a symptom or image.
Bias: The AI performs unevenly across population groups, disadvantaging some patients.
Drift: Performance degrades over time as real-world data moves away from what the model was built on.
Distortion: The AI subtly alters meaning or emphasis, changing how information is understood.
Opacity: The reasoning behind an output cannot be interpreted or explained.
Propagation error: An early mistake carries downstream, compounding through later steps or systems.
Persuasive influence: The AI shapes a person’s decision or behaviour beyond simply informing them.
Pink Glass Icon Health Heart Chart

Clinical context first

Risk severity is always anchored in who receives the AI output and what decision it shapes.
Pink Glass Icon People Group

Participatory development

Developed through a formal multi-stakeholder evaluation programme, with the people who'll actually use it.
Pink Glass Icon Chat Communicate Cycle

Empirically grounded

Its taxonomies and risk logic have been tested against a dataset of real AI feature assessments from live health products.

Designed to support compliance across global AI frameworks - not replace them

Regulatory & Standards Alignment
EU AI ACT
EU AI Act
Articles 9–17 risk management & transparency obligations
European Union
DCB 0129
DCB0129
Clinical risk management for health IT systems
United Kingdom
NICE ESF
NICE ESF
Evidence Standards Framework for digital health technologies
United Kingdom
GMLP
GMLP
Good Machine Learning Practice for medical AI
US · UK · Canada
FDA SaMD
FDA SaMD
AI/ML Software as a Medical Device guidance
United States
ISO 14971
ISO 14971
Medical device risk management standard
International
NHS AI Strategy
NHS AI Strategy
NHS England AI adoption & governance framework
United Kingdom
Pink Glass Icon Search List
RESEARCH FOUNDATION

"Operationalising Trust in Health AI: A Participatory Framework for Contextual Risk Assessment in Digital Health"

A working paper co-authored with Ulster University and the American Psychological Association has been submitted to HHAI2026 - the Hybrid Human-Artificial Intelligence Conference in Brussels, July 2026. A peer-reviewed journal paper is planned for H2 2026.

Learn more

Developed and validated in an open research programme guided by global experts

An Advisory Steering Group of clinical, technical and governance experts and a Supplier User Group of digital health developers guide and stress-test the method.

Read more about the commencement of the research group here.  

DH AI group logos

Why this matters now

35%

of people globally now use AI to manage their health, making AI health products a mainstream clinical touchpoint, not a fringe use case.

64%

of AI-fluent consumers believe they can perform at least one medical task as well as a trained professional. Perceived AI competence is outpacing governance.

-10

year-on-year drop in public confidence in finding reliable health information. Trust in the health information environment is deteriorating fast.

Source: 2026 Edelman Trust Barometer Special Report: Trust and Health. 16,009 respondents across 16 markets.

Engage with the AiRP method and its validation

Open to developers, health systems, regulators and researchers who want to shape how AI risk in health is assessed.