NHS · Clinical AI · RAG
Clinical teams waste minutes hunting through NICE guidelines and BNF drug protocols mid-consultation. This chatbot has read all of them — and answers with a citation, a SNOMED CT code, and a confidence score in seconds, not minutes. Human in the loop, always.
The Problem
Clinical teams navigating NICE guidelines, BNF drug protocols, and SNOMED terminology face the same problem every day: the knowledge exists, but it is scattered across PDFs, portals, and systems — none of them queryable in the time a consultation allows. Existing search tools return lists of documents. Clinicians need answers, with citations, in under thirty seconds.
The Approach
Seven NICE and BNF guidelines — Type 2 Diabetes (NG28), Cancer referral (NG12), UTI (NG109), Chronic Kidney Disease (NG136), Sepsis recognition (NG51), and Metformin (BNF) — are structured into a knowledge base. Gemma 4, an open-weight model self-hosted on Azure Container Apps, retrieves the relevant section and returns a plain-English answer. Every single response includes a NICE guideline citation, a SNOMED CT code, a confidence score, and a mandatory "requires clinical review" flag. No autonomous decisions. Every output is explainable and auditable.
Use Cases
The RAG pipeline, LLM, and knowledge base are all replaceable. The same architecture ships to a UK Primary Care Network, a US Epic shop, or a pharma regulatory team — with a different knowledge base loaded and a different compliance layer on top.
Problem: A GP registrar needs to check the NICE threshold for metformin dose escalation mid-consultation. Searching NICE.org.uk takes 3–5 minutes they don't have.
How to use it: Ask the assistant: 'What does NICE say about metformin dosing for T2 diabetes?' — get the answer, confidence score, and NG28 citation in under 10 seconds. No tab-switching, no PDF.
Problem: Commissioning a new frailty pathway requires checking eligibility criteria across 4 NICE guidelines. Doing this manually takes a full analyst day and produces inconsistent outputs.
How to use it: Query each guideline in sequence, get structured outputs with NICE citations and SNOMED codes, and export the conversation as an audit trail for governance review. Replace a day's work with an hour.
Problem: You are building a clinician-facing portal and need a guideline query layer. Building a RAG pipeline from scratch takes 3–6 months of ML engineering time.
How to use it: Fork this architecture: swap Gemma 4 for Azure OpenAI GPT-4o, point the knowledge base at your full NICE corpus in Weaviate, wrap the API behind NHS Login, and deploy to Azure UK South in 4–6 weeks.
Problem: Physicians spend 15–20 minutes per shift searching for clinical policy documents, prior-authorisation criteria, or formulary restrictions — outside the EHR.
How to use it: Embed this RAG layer as a CDS Hook inside the EHR sidebar. Swap the knowledge base for your institution's own clinical policies + CMS National Coverage Decisions. Use Epic OAuth for auth. Runs inside your AWS GovCloud VPC.
Problem: Prior-authorisation review requires a clinician to manually check whether a requested treatment meets clinical criteria. The backlog delays care and costs staff time.
How to use it: Use the RAG pipeline as a first-pass decision support layer: query the clinical criteria for the requested code, generate a structured recommendation with confidence score, and route only borderline cases to a human reviewer.
Problem: A regulatory submission needs to reference NICE (UK), FDA (US), and EMA (EU) guidance for the same indication. Navigating three separate corpora across teams takes weeks.
How to use it: Load all three jurisdictions into a unified vector store (Pinecone or Qdrant). Query across them in a single session. The LLM cites the source guideline and jurisdiction for every answer — making cross-market regulatory analysis auditable.
Problem: Your clinical decision tool needs a natural-language query layer over proprietary clinical knowledge — but building and maintaining a RAG pipeline is not your core product.
How to use it: Licence or fork this pipeline as a microservice. Replace the NICE knowledge base with your own clinical graph (e.g., Causaly's biomedical ontology or Verisian's trial data). Wrap with your brand. Ship under your compliance layer.
Interoperability
Every layer is replaceable. The cards below map each component to its UK NHS, US, and international equivalent — so a procurement team or engineering lead can assess integration lift at a glance.
UK NHS
Azure OpenAI GPT-4o (NHS-compliant region)
US
AWS Bedrock Claude / Amazon Titan
International
GCP Vertex AI Gemini
UK NHS
SNOMED CT browser API + ClinicalKnowledge.org
US
FDA guidance documents + CMS National Coverage Decisions
International
WHO ICD-11 + local national formulary
UK NHS
Pinecone / Weaviate over full NICE corpus
US
ChromaDB / pgvector (self-hosted, HIPAA-eligible)
International
Qdrant / Milvus (open-source, on-premise)
UK NHS
FHIR R4 (NHS England API, GP Connect)
US
Epic SMART on FHIR / Cerner CDS Hooks
International
HL7 FHIR R4 (region-specific endpoint)
UK NHS
Azure UK South (NHS-approved, ISO 27001)
US
AWS GovCloud / AWS Bedrock (HIPAA BAA)
International
GCP EU / Azure EU (data residency)
UK NHS
DTAC, DSP Toolkit, NHS AI Lab assurance
US
HIPAA, SOC 2 Type II, FDA SaMD guidance
International
ISO 27001, MDR (EU), regional health regs
UK NHS
NHS Login / Azure AD (ADFS)
US
Epic OAuth 2.0 / Auth0 (HIPAA-eligible)
International
SAML 2.0 / Okta Enterprise
This is a portfolio demonstration. No real patient data is stored or processed. Every response from the clinical assistant requires human review before any clinical action is taken. UK GDPR Art. 22 — Automated Decision-Making — applies. This system never makes autonomous clinical decisions.