Clinical AI still needs
a defensible safety case.
Dr Chiho Song provides Clinical Safety Officer support for AI, SaMD and model-driven health IT products entering the NHS: intended-use review, AI hazard analysis, workflow risk, DCB 0129 evidence, DTAC clinical safety mapping and signed CSO release where appropriate.
If your AI product is heading for the NHS.
The common late blocker is not whether the model is impressive. It is whether the product team can show a named Clinical Safety Officer, a hazard log, a clinical risk management file, a Clinical Safety Case Report and a defensible route for model change after go-live.
This page is for clinical AI and SaMD suppliers who need the clinical safety stream handled alongside validation, DTAC and medical-device questions. If the immediate wording from the buyer is DCB 0129, use the DCB 0129 consultant page; if it is external CSO, use independent Clinical Safety Officer support.
Hazards that ordinary software checklists miss.
AI clinical safety is not just model validation. The hazard analysis has to cover the workflow, the user, the display, the escalation route, and what happens when the model changes or drifts.
Unsafe over-reliance
Users defer to model output even when clinical context should override it, or the interface makes disagreement harder than acceptance.
Performance changes in deployment
The training or validation cohort does not match the NHS site, patient mix, pathway, equipment, referral pattern or documentation style.
Risk surfaced badly
Confidence, uncertainty, urgency or caveats are hidden, over-simplified or routed to the wrong person at the wrong point in the workflow.
Safety case goes stale
A retrain, threshold change, prompt change or workflow change alters the hazard profile but no one updates the hazard log or safety case.
Use outside intended purpose
A model built for triage becomes used for diagnosis, prioritisation or reassurance without the controls needed for that new claim.
Clinical workarounds
Real users create shortcuts when the product slows them down, producing hazards no offline validation dataset can reveal.
Clinical safety evidence for the AI workflow.
Clinical safety, medical device regulation and validation must align.
Clinical risk in the NHS workflow.
Supplier-side clinical risk management: hazard log, clinical risk management file, Clinical Safety Case Report, safety requirements and CSO release memo.
Regulated intended purpose.
UKCA, CE, MHRA classification and technical documentation are separate from DCB 0129. They often need to be aligned, but one does not replace the other.
Evidence that supports the claim.
Reference standard, cohort, acceptance thresholds and performance limits need to match the intended use that appears in the safety case.
Send the intended use first.
For AI and SaMD, the first useful artefact is a clear intended-use paragraph. Send that with the current validation evidence, the target NHS setting and what the buyer or reviewer is asking for.
What needs reviewing?
Choose the route that matches the blocker.
DCB 0129 or medical device?
How NHS clinical safety and medical-device regulation answer different questions.
Read → ProcurementDTAC clinical safety
Evidence mapping for named CSO, hazard log, CRMF, CSCR and release memo.
Read → Ongoing supportContract Clinical Safety Officer
Named CSO retainer and post-go-live safety governance after the first release.
Read →