About
A resource built by a clinician, for the CSOs doing the work.
Clinical AI Safety is independent, NHS-focused, and written from frontline experience of the gap between certified theory and real clinical deployment.
The author
Dr Doju Cheriachan
MBBS, GMC registered · Internal Medicine Trainee, Sheffield Teaching Hospitals NHS Foundation Trust · CSO certified
PubMed-indexed author. LLM evaluator. Clinical AI collaborator.
The decision to build this resource came from a pattern I kept seeing on both sides of NHS AI deployment: certified Clinical Safety Officers who understand DCB0129 and DCB0160 in principle, and suppliers producing Safety Case documentation that technically meets the standard but does not actually describe how the system behaves in a real clinical workflow.
The gap between those two things is where patients get hurt. And it is a gap that existing training, by design, does not fill — the certification is necessarily generic, and the application is necessarily local.
Clinical AI Safety is an attempt to fill that gap honestly. Not as a replacement for CSO certification, and not as a substitute for a Trust's own clinical safety processes, but as the practical companion most CSOs say they wanted when they were first asked to sign off an AI deployment.
What to expect
Clinician-led. Framework-anchored. Scenario-driven.
30 modules
A structured curriculum across hazard identification, safety case authorship and review, post-deployment monitoring, AI-specific failure modes, and incident response.
10 simulators
Interactive scenarios. Hazard log authorship, safety case evaluation, DPIA review, PSIRF response exercises. Feedback modelled on how a senior CSO would challenge your reasoning.
12 frameworks
Every module is explicitly anchored to the relevant standards — DCB0129, DCB0160, DTAC, PSIRF, ISO 14971, BS AAMI 34971, MHRA AIaMD, UK GDPR, NICE ESF, EU AI Act, AMLAS, ECSF.
AI CSO Assistant
An agentic assistant, under active development, designed to help CSOs draft hazard entries, review supplier documentation, and stress-test a safety case against AI-specific failure modes.