When you finish the design phase, you should have a set of documents that development can build on. The table lists them by topic.
| Topic | What you should have |
|---|---|
| Problem and scope | A clinical problem with measurable success criteria; the limits of what the tool will do; your reasons for choosing AI |
| Stakeholders | A stakeholder map; records of engagement activities and what changed because of them; patient representation |
| Intended use | Clinical settings and the resources in each; user groups and their technical literacy; explainability needs per group; model inputs and outputs |
| Ethics and law | The right type of review (METC, local review or hospital governance); data governance agreements; the MDR/IVDR class and EU AI Act status; a GDPR lawful basis and Art. 9 condition; a DPIA in progress |
| Data | Data sources and access agreements in progress; documented sources of data variation; attributes chosen for bias monitoring, with a legal basis |
| Risk | A risk management file; an initial risk register with likelihood and consequence; planned mitigations |
| Bias and society | Likely sources of bias; expected effects on staff, equity and the environment |
In the fictional sepsis running case used throughout this section, this would mean, for example, a problem statement agreed with ward nurses and internists, a decision on whether the hospital relies on the MDR in-house exemption, and a register that already lists alert fatigue as a risk.