STAR-TRE is developing practical methods, tools and guidance to help Trusted Research Environments (TREs) assess privacy risks in free text more consistently. To achieve this, they are developing approaches to identify both direct and contextual (known as “quasi”) identifiers while preserving the information that makes free text valuable for research.
Every day, public services generate vast amounts of free-text information — from clinical notes and radiology reports to education records, social care assessments and justice documentation. These narratives provide the context, reasoning and detail that structured (i.e. coded) data cannot capture, helping explain not just what happened, but why.
Yet, despite representing some of the richest information available, free text is rarely accessible for research. Unlike structured datasets, narrative text can contain names, locations, family relationships, occupations, significant life events and other information that could identify an individual. Assessing and managing these privacy risks is complex, making free text one of the most underused sources of research data.
Led by DataLoch (University of Edinburgh and NHS Lothian), STAR-TRE (Safe and Trustworthy Assessment of Risk in TREs for Sensitive Free-Text Access) brings together the Scottish Safe Haven Network and the University of Sussex to develop evidence-based approaches that combine artificial intelligence, privacy risk assessment, governance and public involvement. Although the project is focused on health data, the methods, governance approaches and evaluation frameworks are designed to be transferable across other forms of sensitive free text held within TREs.
What STAR-TRE aims to achieve
By the end of the project, the STAR-TRE team aims to develop and pilot the use of a toolkit and guidance to support the wider assessment of privacy risks caused by direct and quasi-identifiers within free text by TRE operators. The team’s goals include:
- Exploring public perspectives on the use of Large Language Models (LLMs) within a de-identification tool – develop and deliver a series of discussion workshops with members of the public across the UK focused on the possible use of LLMs to identify privacy risks in free-text.
- Developing a de-identification process – identifying possible quasi-identifiers – e.g. family relationships, ages of children, current and past jobs – that would underpin a privacy-risk assessment tool;
- Securing feedback from the public and TRE operators – gathering input on a prototype privacy-risk assessment tool to inform further development; and
- Test de-identification tools – conduct pilot tests of the de-identification process and privacy-risk assessment tool.
Progress so far
Within the first six months of delivery, the STAR-TRE team has:
- Embedded public contributors as members of the project team from the outset.
- Commissioned Ipsos to run sessions on LLMs with public members and report their findings which highlighted support for the idea alongside the need for a human role in implementation.
- Explored – with TRE operators and public members – the STAR-TRE prototype privacy-risk assessment tool and generated a prioritised list of next steps for improvement .
- Developed a new taxonomy (classification) of privacy-risk types, covering direct and quasi-identifiers, that may contribute to re-identification.
- Developed the first components of a de-identification approach combining rule-based methods with AI-assisted identification of contextual privacy risks.
- Evaluated the ability of LLMs to identify contextual privacy risks within representative health free-text examples.
- Secured access to pilot datasets to evaluate the complete toolkit within real-world TRE workflows.
During the remainder of the project, the team will refine the privacy risk assessment toolkit, undertake further user evaluation with both public contributors and TRE professionals, and pilot the complete approach using realistic research scenarios. The project will conclude by producing practical guidance to support wider adoption across TREs.
How STAR-TRE supports DARE UK’s vision
STAR-TRE contributes to DARE UK’s vision by developing the evidence, methods and governance needed to support informed decisions about the use of sensitive free text.
Through combining artificial intelligence, transparent privacy risk assessment, public involvement and practical implementation guidance, the project aims to provide TREs with the confidence to unlock one of the richest, and currently least accessible, sources of research data.
Although demonstrated initially using healthcare data, the principles, tools and governance approaches developed through STAR-TRE are intended to support the safe use of sensitive free text across a wide range of public sector data, creating a foundation for more trustworthy and consistent access to narrative information for research.
About the DARE UK Next-Gen Catalysts
The DARE UK Next-Gen Catalysts are early-stage, exploratory projects developing and testing the next generation of capabilities for Trusted Research Environments (TREs) across the UK. Bringing together researchers, technology specialists, public contributors and other partners, these projects explore bold new ideas that could expand how sensitive data is accessed, analysed and governed while maintaining privacy, security and public trust. Their role is to generate evidence, prototypes and practical learning that can help shape the future of secure, trustworthy data research for public benefit.
