HealthTech NHS Radiology AI Clinical Decision Support — NHS Trust

MHRA AI classification

AI CDS that aids radiologist diagnosis = Software as a Medical Device (SaMD) Class IIa (MHRA MDR 2002 — UK MHRA post-Brexit classification).

Class IIa requirements

• 1UKCA marking (UK Conformity Assessed — CE mark no longer valid in GB from July 2024),
• 2UK Responsible Person registration with MHRA,
• 3Notified Body review (MHRA-approved UK Notified Body — for Class IIa, requires Notified Body certification),
• 4clinical performance data (sensitivity and specificity vs radiology gold standard — NHS DTAC requires peer-reviewed evidence),
• 5post-market surveillance (ongoing accuracy monitoring in production — MHRA requirement for AI medical devices).

AI CDS workflow

CT/X-ray image acquired → DICOM image sent to AI service (PACS/RIS integration via DICOM C-STORE) → AI model processes image → structured finding returned (anatomy, finding, confidence score, priority flag) → finding overlaid on PACS viewer (OHIF Viewer integration) → radiologist reviews AI finding + image → radiologist reports (AI finding is a prompt, not a decision).

Critical design

AI is a second reader that prompts, not a decision maker.

AI alert

• urgent finding flag (pneumothorax, PE — pulmonary embolism) → immediate radiology consultant alert (GOV.UK Notify).
• Radiologist can accept, modify, or reject AI finding — all decisions logged for audit.
• DCB0129 AI Clinical Safety —

AI clinical safety hazards

(1) false negative — AI misses significant pathology (radiologist over-relies on AI normal result → missed diagnosis).

Mitigation

display confidence score prominently, training for radiologists on AI limitations, audit of false negatives monthly. (2) False positive — AI flags non-existent pathology → unnecessary patient workup.

Mitigation

AI finding shown as "suggested" not "confirmed" — requires radiologist verification. (3) Demographic bias — AI performs differently across patient demographics (age, sex, ethnicity — documented in AI bias testing).

Mitigation

model validation across Trust's patient demographics, bias monitoring dashboard. (4) Distribution shift — CT scanner model changes → AI trained on old scanner → degraded performance.

Mitigation

AI performance monitoring (monthly accuracy check vs radiologist gold standard).

NHS AI Lab 14 Evidence Standards

• 1intended purpose clearly defined,
• 2training data documented (source, size, demographics),
• 3model architecture documented,
• 4performance metrics (sensitivity, specificity, AUC-ROC — vs radiologist gold standard),
• 5external validation dataset (different Trust population),
• 6bias testing (performance across demographic groups),
• 7deployment context documented (workflow, user training),
• 8monitoring plan (post-deployment performance tracking),
• 9human oversight mechanism (radiologist always final decision),
• 10explainability (attention maps — where AI is looking on image),
• 11safety validation (DCB0129 hazard log),
• 12benefit assessment (reduction in backlog × radiologist time saving),
• 13equity assessment (access for all patient groups),
• 14governance (NHS Trust Caldicott Guardian sign-off).