NHS AI Radiology Triage Platform -- Chest X-Ray Priority

MHRA Class IIa SaMD

AI that aids radiologist triage decision is Class IIa.

Regulatory pathway

UK Notified Body (BSI Assurance UK) technical file review.

Key IEC 62304 Class B requirements

• 1software requirements specification (SRS -- what AI detects, sensitivity/specificity targets, workflow),
• 2software design (model architecture -- DenseNet-121 chest X-ray classification),
• 3software testing (unit tests + integration tests + V&V study),
• 4anomaly resolution (bug tracking -- all safety-critical bugs closed before release).

ISO 14971 risk management

primary hazard -- AI fails to detect urgent finding (false negative) -- patient delay in treatment.

Mitigation

• AI triage assists, never replaces -- all X-rays reported by radiologist.
• AI negative (no urgent finding) moves to standard worklist, not discarded.
• AI Model for Chest X-

Ray Triage

Chest X-ray

AI model

DenseNet-121 (dense convolutional network -- pre-trained on CheXpert and fine-tuned on UK NHS chest X-ray dataset).

Conditions detected

• 1pneumothorax (collapsed lung -- urgent -- AI sensitivity target 95%),
• 2consolidation (pneumonia, lung cancer -- urgent -- AI sensitivity target 90%),
• 3pleural effusion (fluid -- semi-urgent),
• 4cardiomegaly (enlarged heart -- semi-urgent),
• 5normal (low priority).

Uncertainty quantification

Monte Carlo Dropout -- AI returns confidence interval per finding (high confidence urgent = immediate escalation, low confidence = report with standard priority + clinical review flag).

Training data

42,000 UK NHS chest X-rays (multi-centre, demographic diversity confirmed -- NHS AI Lab Standard 2).

PACS Workflow Integration

PACS (Sectra IDS7) integration: (1) DICOM C-STORE receiver (X-ray arrives from DR/CR room -- DICOM C-STORE to AI service -- AI processes within 60 seconds), (2) AI priority score (urgent: 1-5, high: 6-10 -- sent back to PACS as DICOM SR), (3) PACS worklist reorder (Sectra IDS7 accepts DICOM SR worklist priority -- urgent studies moved to top of reporting queue), (4) radiologist notification (urgent finding -- PagerDuty alert to on-call radiologist), (5) AI performance monitoring (weekly -- AI sensitivity vs radiologist confirmed urgency -- DCB0129 ongoing monitoring). 60-second

AI inference

GPU inference (AWS g4dn.xlarge -- NVIDIA T4 GPU) -- DenseNet-121 inference in 8 seconds, 60-second SLA with PACS DICOM round-trip included.

DCB0129 Clinical Safety

DCB0129 for chest X-ray

AI

clinical safety case argument. (1) AI is triage assistance only (AI never discards study, never reports study -- it only reorders worklist), (2) all urgent studies reviewed within 1 hour of AI urgent flag (radiologist acknowledges alert), (3) AI sensitivity monitoring (weekly sensitivity check -- if sensitivity drops below 92% on live data, AI is taken offline for retraining), (4) audit log (every AI triage decision logged -- AI score, radiologist report, any discordance), (5) discordance review (monthly review of cases where AI was urgent but radiologist normal -- learning log).

Clinical Safety Officer

Consultant Radiologist.

MHRA CE marking note

UKCA (post-Brexit) -- not CE for UK market.