This project investigates whether LLMs can automate review of NHS patient safety learning responses using the HSSIB Learning Response Review and Improvement Tool (LRRIT). It compares LLM outputs with expert reviews, identifies risks, and develops a framework for qualification and risk classification in non-diagnostic governance roles.

The project prepares anonymised PSIRF reports, configures LLMs for review, and develops an evaluation framework comparing AI outputs to expert scoring. Analysis identifies error patterns and governance risks, with outputs translated into regulatory recommendations through stakeholder consultations and policy briefs. Data access, report variability, and model risks are mitigated through secure NHS sharing, standalone LLM setups, and experienced research teams.

This project uses advanced generative AI to create highly realistic synthetic medical images for stress testing AI models, focusing on chest radiograph disease detection and mammographic density prediction. It aims to inform MHRA guidance on AI safety and provide tools for independent performance evaluation.

Causal generative AI models are retrained on diverse datasets, used to stress-test image classification systems, and analysed statistically to identify vulnerabilities. Outputs and insights are shared in workshops with regulators, clinicians, and patient representatives to shape future regulatory frameworks. Mitigations include multiple generative AI approaches, curated datasets, and experienced research staff.

ARIES-MED evaluates a generative AI tool that produces radiology-style reports across trauma CT, chest X-ray, and CT head in emergency departments. AI-generated reports are compared with radiologist and teleradiology outputs to assess accuracy, sensitivity, specificity, and clinical acceptability.

The project extracts and de-identifies imaging studies, conducts blinded expert review and evaluation by ED clinicians, performs statistical analysis to validate accuracy and reliability, and translates findings into regulatory insights through white papers and stakeholder briefings. Risks such as reviewer bias or data quality issues are mitigated via blinded randomisation, rater training, Safe Haven approvals, and structured scoring protocols.

This project evaluates two AI tools—qXR and Annalise Enterprise CXR—using 110,000 chest X-rays from NHS Greater Glasgow and Clyde. It assesses whether these tools can accurately identify normal scans without missing clinically important findings, potentially reducing reporting delays caused by radiologist shortages.

The project engages patients and the public in study design, securely prepares and manages the large CXR dataset, and ensures robust governance through ethics approvals and collaboration agreements. AI performance is evaluated against radiologist reviews, including bias analyses, and findings are disseminated through reports, publications, regulatory white papers, and NHS guidance.

The CLIO (Clinical LLM Intelligent Oversight) extension project expands an InnovateUK study that monitors risks of AI Ambient Scribes—tools that generate clinical notes from doctor–patient conversations. The extension will include three additional NHS Trusts to evaluate audio quality, AI output drift, and clinician reliance, aiming to detect early risks like poor performance or over-reliance and ensure safe adoption of AI scribing tools.

• The programme uses a coordinated, multi-site approach to enable safe, effective and scalable evaluation of AI technologies in real clinical settings. Sites are first supported to establish the necessary contracting, ethics and governance foundations, ensuring each location is ready to participate securely.
• Once onboarded, sites connect to a shared monitoring platform that integrates local clinical data and captures insights from over 500 patient encounters per site. Independent clinical evaluators apply a standardised review process to generate robust evidence on performance, usability and clinical impact.
• Insights from all sites are brought together through shared dashboards and facilitated workshops, enabling cross-site learning, comparison and the identification of best practices. This ensures that local findings contribute to national understanding and support the wider adoption of safe, effective AI tools.
• Strong governance underpins delivery, supported by flexible contracting, protected evaluator time, strict data anonymisation and oversight from a dedicated steering group.