Sovereign health AI means that the data used to train Health AI models, remain under the Trust's control from start to finish. All patient health data must stay within the Trust's domain, whether it is identifiable or de-identified. De-identification reduces risk but does not remove it. Linkage, re-identification and inappropriate secondary use remain real possibilities once data leaves the Trust's information governance structure. Our legal and regulatory obligations also go beyond confidentiality. We must be able to monitor, audit and account for how patient data is used, and we cannot do that for data held outside our domain.

The Health Data Institute (HDI), a research directorate within the Trust, puts this principle into practice. Trust staff collaborate with external service users, typically academics, who work only with the Trust's de-identified patient health data. Data access is confined to the Analytics Environment, a dedicated area within our Secure Data Environment. Access is role-based and time-limited, so each research team sees only what its approved project requires, and only for as long as the project runs.

Before any work begins, the HDI Data Access Committee (DAC) screens each proposal for its likely impact on patient care, patient safety and health inequalities. The DAC is mandated by UK legal and regulatory data frameworks. In the HDI, the DAC is split into two parts - the Project Selection Group and the Steering Group.

Research teams cannot egress trained models from the Analytics Environment. The HDI transfers approved models directly to its inference server. There, models run in shadow deployment at the 'bedside', generating outputs alongside routine care without influencing clinical decisions. This allows real-world evaluation of performance, safety and equity in our own patient population before any model is considered for clinical use.

This approach gives academic partners access to rich NHS data and a realistic route to clinical evaluation. At the same time, the Trust keeps full custody of patient data and the models derived from it. Patients can be confident that their data never leaves the organisation entrusted with their care, and that its use is scrutinised at every stage.