Tick these items before you move to development. Each group matches a chapter in the design section.
Problem definition¶
Clinical problem documented with measurable success criteria, set before seeing any data
AI tool scoped: what it will and will not do
Clinical need confirmed with at least one clinician from the target setting
Reasons for choosing AI over a simpler rule or score documented
Stakeholders¶
Stakeholder map completed (clinicians, nurses, patients, managers, data protection officer, IT, legal, ethics)
Patient engagement plan in place, with compensation and accessibility considered
Interdisciplinary team assembled, and expertise consulted from outside the team recorded
Intended use¶
Clinical settings defined, including the resources each setting has
User groups defined, with their role and technical literacy documented
Explainability needs documented for each user group
AI inputs and outputs specified
Community-defined standards chosen for inputs and labels (for example SNOMED CT, LOINC, HL7 FHIR)
Out-of-scope uses and settings documented
Ethics and legal¶
Type of review determined (accredited METC for WMO research, local review for non-WMO research, hospital governance for deployment) and started
Application-specific ethical issues listed and discussed with the team
Data governance agreements in place or in progress
Regulatory route determined (MDR/IVDR class, whether the EU AI Act treats the tool as high-risk, in-house exemption or CE marking)
GDPR Art. 6 lawful basis and Art. 9(2) condition for health data documented
Data protection impact assessment (DPIA) started
Data¶
Data sources identified
Sources of data variation documented (equipment, protocols, operators, sites, time)
Annotation and labelling plan written
Attributes for bias monitoring (for example age, sex, ethnicity, socioeconomic status) identified
Legal basis and safeguards for processing these attributes documented
Plan for later external evaluation on data from another site
Risk management¶
Risk management file opened
Initial risk register completed with likelihood and consequence ratings
Mitigation measures planned for high-priority risks
Bias and social impact¶
Likely sources of bias in data, labels and problem framing recorded
Effects on staff workload and skills considered
Effects on health equity considered
Environmental cost of training and running the tool estimated
Plan for telling patients that AI is used in their care