At DataLoch, we are collaborating on two unique projects that will enhance the possibilities for novel research conducted through Trusted Research Environments (TREs) in the UK.
TransPECT will develop methods to assess the confidentiality risks posed by new Artificial Intelligence (AI) models developed using sensitive research data, while HEAL-Scot will produce a blueprint for TRE operators to more efficiently and robustly link health with other public-sector data.
TransPECT
AI language models (e.g. ChatGPT and Gemini are two of the most commonly known) are of increasing interest to the research community, as they can quickly analyse and learn from large amounts of data. Models that are hosted locally (with no data being shared externally) can be safely trained within TREs. However, models can memorise information from their training data, creating a risk that sensitive information may be exposed when a trained model leaves the TRE. This is a particular issue when health data from real patients is part of the model training process.
To address this problem, the TransPECT team – a University of Edinburgh collaboration led by Arlene Casey bringing expertise from DataLoch and working with Health Informatics Centre, King’s College London, and Public Health Scotland – is investigating how, when and why information is memorised by language models, what this means for confidentiality, and how these risks can be systematically measured and assessed.
We will translate this work into methods, tools and guidance for TRE operators, integrating our final outputs with the SACRO-ML toolkit. Our work will open up the possibility for governance specialists within TREs to make informed decisions about new AI models for free-text data and establish a pathway for these models to be safely deployed in frontline services.
TransPECT introduction page
HEAL-Scot
We know that a person's health is influenced by a wide range of factors, including where they live, their social situation, and their access to education and health care. However, linking data from multiple sources can be time-consuming and complex, which can limit the scope and depth of research and has impacted on the number of studies that have been completed to date.
To resolve this issue, we are working with colleagues from the Health Informatics Centre to establish a reproducible methodology to link health datasets with housing, environment, and geographic data. We will test this methodology in support of an example research project – led by Laura Ward from the University of Dundee – exploring the impact of social inequalities on health.
Our goal is to develop a tried-and-tested blueprint for TRE operators that will improve the speed and consistency in preparing data for multi-sector, multi-region research projects, while retaining the highest levels of confidentiality.
HEAL-Scot introduction page
Ways of working
Collaboration is at the heart of both TransPECT and HEAL-Scot. We're working closely with a range of partners, bringing together the knowledge and expertise that will lead to important improvements for research that relies on TREs.
Regular project meetings ensure our public collaborators can share their feedback, offer new ideas, and help shape the projects. This approach allows us to remain aware of societal expectations and ensures that our work remains relevant and trustworthy.
Both projects will conclude in March 2027.
Acknowledgement
Both TransPECT [grant number: UKRI4081] and HEAL-Scot [grant number: UKRI4075] have been funding by DARE UK as part of the Research Exemplars projects that aim to demonstrate how emerging TRE capabilities can enable real-world sensitive data research.
