NHS Data for AI Training

The NHS holds one of the world's most valuable clinical datasets. Almost none of it is accessible for AI training.

The UK National Health Service provides universal healthcare to 67 million people. Every GP visit, hospital admission, A&E attendance, diagnostic scan, lab test, and prescription generates data. Aggregated across decades, this represents one of the largest, most comprehensive, and most longitudinal clinical datasets on the planet.

For AI researchers, this data would be transformative. Models trained on NHS data could understand disease progression across entire populations, identify rare disease patterns invisible in smaller datasets, and capture real-world treatment outcomes across diverse demographic groups.

But NHS data is governed by some of the strictest data protection frameworks in the world: UK GDPR, the Data Protection Act 2018, the NHS Data Security and Protection Toolkit (DSPT), Caldicott Principles, and the Common Law Duty of Confidentiality. Getting access through traditional channels — NHS Digital data releases, research ethics committee approval, data sharing agreements — takes years and typically only grants access to de-identified extracts for specific, pre-defined research questions.

NHS data governance: the acronyms that block AI training

• UK GDPR (2018) — The UK's implementation of the General Data Protection Regulation. Article 5(1)(c) requires data minimisation: "adequate, relevant and limited to what is necessary." Exporting millions of patient records to train an AI model arguably violates this principle. Rapha Protocol's compute-to-data architecture aligns with it: only model weights leave, not data.
• Data Protection Act 2018 — Supplements UK GDPR with specific provisions for health data (Schedule 1, Part 2). Processing of health data requires a specific condition. Research is a permitted condition — but only with appropriate safeguards. Compute-to-data's hardware enclave, kernel air-gap, and policy engine constitute those safeguards.
• NHS DSPT (Data Security and Protection Toolkit) — Mandatory annual self-assessment for all NHS organisations handling patient data. Covers 10 data security standards across people, process, and technology. Rapha Protocol maps 34 OPA policy controls directly to DSPT standards.
• Caldicott Principles — Eight principles governing use of patient-identifiable information in the NHS. Principle 4: "Access to patient identifiable information should be on a strict need-to-know basis." Principle 7: "The duty to share information can be as important as the duty to protect patient confidentiality." Compute-to-data satisfies both: the AI model needs the data (Principle 7), but only during training and never exported (Principle 4).
• NHS Research Ethics Committee (REC) — Required for research involving NHS patients or their data. REC approval is separate from data access — you need both. Rapha Protocol is the technical infrastructure for data access; the research protocol and REC application are the researcher's responsibility.

NHS data access pathways compared

UK-specific compliance advantage of compute-to-data

The UK Information Commissioner's Office (ICO) has repeatedly signalled that data minimisation and purpose limitation are central to lawful AI development. The ICO's guidance on AI and data protection emphasises that organisations should "consider whether you can achieve your purpose by using less data, or by using a different technique that is less intrusive."

Rapha Protocol's compute-to-data architecture directly implements this guidance. By training models at the data source without exporting raw records, the data processing is both more minimised (only weights leave) and more auditable (every job produces a cryptographically signed proof receipt) than the alternative of exporting de-identified NHS data to cloud infrastructure.