⚡ Quick Summary
Produced as part of the Trustworthy and Ethical Assurance of Digital Health and Healthcare (TEA-DH) project, funded by the Assuring Autonomy International Programme, a partnership between Lloyd's Register Foundation and the University of York, this report introduces Trustworthy and Ethical Assurance (TEA) and applies it to digital health and healthcare. It is addressed to a broad set of actors in the assurance ecosystem, from regulators, accreditation and standards bodies to practitioners and affected individuals.
TEA is presented as argument-based assurance: a structured, accessible argument made of a top-level goal claim, supporting property claims and the evidence that grounds them, documented as an assurance case. The report situates this in the UK's AI assurance ecosystem, lists complementary mechanisms such as risk assessment, bias audit, compliance audit, conformity assessment and formal verification, and groups ethical principles under SAFE-D — Sustainability, Fairness, Data Stewardship, Accountability and Explainability.
Fairness and health equity are the goals examined. Two models are offered: the 'Unvirtuous Circle', a four-quadrant map of World, Data, Design and Ecosystem, and a project lifecycle model spanning project design, model development and system deployment. Two case studies — an AI-enabled clinical diagnostic support system that predicts hypertension risk in patients with Type 2 diabetes, and the CemrgApp Scar Quantification Tool — show how fairness requirements and claims would be identified. The report ends with communities of practice and making assurance cases adhere to the FAIR principles.
🧩 What’s Covered
The report is organised into four chapters, references and two appendices.
- Introduction and context: reviews risk-based approaches, including the EU's risk-categorised regulatory framework and NIST's AI Risk Management Framework with its map, measure, manage and govern stages, and lists limitations of risk-based approaches such as missed hazards, dynamic risks, incomplete data, under-regulation of low-risk systems and limited stakeholder engagement.
- Trustworthy and Ethical Assurance: defines TEA as a tool and a framework, sets out the three basic elements of an assurance case (goal claim, property claims, evidence) with a toy example for an explainable AI system, and adds Box 1.1 on the AI assurance ecosystem, its mechanisms and the roles of assurance providers, regulators, standards bodies, accreditation bodies, government, research bodies, civil society and professional bodies.
- From safety to SAFE-D: argues that safety assurance alone is insufficient, using a hypothetical diagnostic tool with 98% average accuracy to show how accuracy claims can conceal unfair distribution of error, and introduces the SAFE-D principles.
- Fairness and health equity: gives the WHO definition of health equity, sets out social determinants of health (Dahlgren-Whitehead model, the Black Report, Marmot's Whitehall studies and the social gradient) and presents the two models with a worked influenza-forecasting scenario.
- Case studies (Chapter 3): an AI-enabled clinical diagnostic support system trained on 42,000 Connected Bradford patient records with four algorithms combined in an ensemble and evaluated by accuracy and Cohen's Kappa, structured by a fairness considerations map of lifecycle questions; and the CemrgApp Scar Quantification Tool, whose draft case uses four core attributes — bias mitigation, diversity and inclusivity, non-discrimination and equitable impact — to generate six claims, shown as a partial argument.
- Communities of practice and public reason (Chapter 4): defines communities of practice, reports three cross-cutting engagement themes, lists four design properties of the platform (shareable cases, open source, flexible, collaborative) and discusses bias in assurance cases and the 'court of public reason'.
- FAIR principles and next steps: applies findability, accessibility, interoperability and reusability to assurance cases, covering metadata and tagging, APIs and access control, JSON versus SACM/XMI formats, modular design and documentation.
- Appendices: describe the TEA-DH project's objectives, its workshops with regulators, practitioners and researchers, and access to the TEA platform.
💡 Why it matters?
Assurance of data-driven health technologies often stops at accuracy or safety claims, and the report shows why that is insufficient: a system can meet an average performance target while concentrating errors in one patient group. TEA gives teams a way to turn principles such as fairness into explicit claims and evidence and to document them stage by stage across a project's lifecycle, making decisions reviewable by clinicians, regulators, accreditation bodies and affected patients. The report connects this to UK policy — the pro-innovation approach to AI regulation and the Introduction to AI Assurance — and notes that assurance mechanisms also support interoperability with regimes such as the EU AI Act.
❓ What’s Missing
Both case studies are works in progress: the report explicitly offers no complete assurance cases, and some claims have no evidence at all (the claim about supporting patient engagement) or unresolved metrics (computational efficiency). The authors acknowledge difficulty obtaining systematic user feedback and diverse validation datasets, and note that serving a broad audience trades targeted, actionable recommendations for breadth. Whether the lifecycle model and the core-attribute approach can be integrated is left open. Only 80 of the 89 supplied pages were extracted, stopping part-way through Appendix 1, so Appendix 2 on the TEA platform was not reviewed.
👥 Best For
Clinical and research teams developing digital health or AI systems who need to structure and document fairness arguments; assurance, compliance and audit practitioners preparing evidence for regulators, accreditation bodies and NHS adopters; data scientists, engineers and product managers unfamiliar with argument-based assurance; and policy or governance staff building assurance capability and communities of practice.
📄 Source Details
Trustworthy and Ethical Assurance of Digital Health and Healthcare, by Burr, Arana, Gould Van Praag, Habli, Kaas, Katell, Laher, Leslie, Niederer, Ozturk, Polo, Porter, Ryan, Sharan, Solis Lemus, Strocchi and Westerling, dated 2024 in the recommended citation printed on page 2 together with the DOI https://doi.org/10.5281/zenodo.10532573. The supplied PDF runs to 89 pages; text was available for 80, ending part-way through Appendix 1. No edition, version or series number is printed.
