FORTRESS-TeHR: Federated, Open and Reliable TREs for Synthetic Textual Healthcare Records

Current trusted research environments (TREs) and their capabilities are designed for secure and efficient handling of real-world structured data. While this often supports rich data analysis, the focus on structured data types restricts the research space and questions we can answer. For example, often missing from the data landscape is free-text narrative, which is a key data source in healthcare. Valuable clinical information is often recorded only in notes and reports, with some specialties and services almost completely documenting care in text (e.g. mental health). Research in these domains may therefore be limited to an incomplete structured data environment, which in turn may result in lack of understanding and progress, and exasperation of inequities (e.g. patients with certain diseases being less likely to benefit from healthcare data analytics). A key reason for free-text data not being hosted in TREs is its inherent sensitivity and unexpected risks to patient privacy. Manual text de-identification is impossible to scale to accommodate different settings and numbers needed for most analytics purposes. Several efforts have been invested in automated anonymisation of free text data, but there are still no accepted protocols, techniques or metrics to characterise and quantify privacy preservation of free-text data. The associated, unquantified risk has consequently resulted in real-world free-text data rarely, if ever, being hosted within TREs. Synthetic text collections are emerging as an alternative and trustworthy solution, where data is generated through an AI process that ensures both task-specific data relevance and privacy. In this pilot we will explore embedding of synthetic healthcare text generation and its use into existing TRE infrastructures developed in DARE UK Phase 2. This will expand the current TRE capabilities with a new modality in healthcare. To ensure fidelity and representativeness of synthetic text, we will use de-identified real-world data from different settings as a guide for federated text generation. Built in this process will be the use of differential privacy to ensure patient confidentiality. A key outcome of the project will be a methodology and prototype that provides a trusted and reliable approach to generation and validation of clinical free-text data. We will co-develop with clinicians, patients and regulators an early technical validation framework consisting of privacy validation, text quality assessment (with reference to a given clinical or cohort specification) and early testing of the usefulness of generated data for analytics tasks. As a demonstrator use case, we will apply the framework to assess the quality and the value of generated text for federated learning for downstream language modelling tasks in the domain of cardiology. We will organise a series of “validatathons” with patients, clinicians, data scientists and regulators to explore the challenges, opportunities and acceptance for synthetic free-text data. We will use existing, publicly available real-world de-identified datasets to generate different collections of synthetic data. We will then use these collections to train models to optimise patient cardio-workflows and compare the outcomes with the workflows generated from the original dataset. This way, we will use “synthetic data as a risk management approach for testing different analytical capabilities within TREs”, mitigating patient privacy risks. Generated data will be represented using adapted common data models for sensitive data (e.g. OMOP), with RO-Crate used to represent additional meta-data describing the synthetic generation process and data provenance.