Answers are grounded in this paper and cite their sources.
Objective : A persistent gap separates the trustworthy AI principles articulated by consensus guidelines and regulatory instruments from the deterministic, verifiable specifications required to develop clinical decision-support systems. While frameworks such as FUTURE-AI operationalize trustworthiness through detailed sub-principles and lifecycle recommendations, the translation of these principles into formally structured engineering requirements remains under-specified. This paper presents a requirements engineering framework that systematically translates stakeholder needs, preferences, concerns, and regulatory constraints into formal, verifiable system specifications for clinical AI. Methods : The framework follows a three-phase methodology: (i) hybrid multi-stakeholder elicitation combining bottom-up clinical, patient, and ethical inputs with top-down regulatory and contractual constraints, followed by scoping to define the operational system boundary; (ii) formalization of scoped inputs into structured, verifiable requirements via the ISO/IEC/IEEE 29148 standard, applying its syntactic template, obligation levels (shall/should/may), functional and non-functional typing, and per-requirement means of verification (MoV); and (iii) systematic mapping of the formalized requirements to the FUTURE-AI taxonomy as a normative coverage criterion rather than a descriptive classification scheme. Results : Applied to the AI4HF project, a Horizon Europe initiative for personalized risk prediction in chronic heart failure, the framework yielded a fully specified, verifiable requirement corpus, demonstrating that it produces standards-compliant specifications with exhaustive FUTURE-AI coverage and per-requirement verification. The instantiation generated 173 requirements, each paired with a defined MoV; inter-rater validation of the FUTURE-AI mapping on a stratified subset yielded Cohen's κ = 0.72 (substantial agreement). The resulting corpus is publicly available, constituting, to our knowledge, the first openly available FUTURE-AI-aligned, ISO/IEC/IEEE 29148-compliant requirements dataset for a high-risk clinical AI system. Conclusion : The methodology is designed to be reproducible across clinical AI projects beyond the AI4HF instantiation and is directly applicable to multi-stakeholder development efforts subject to high-risk regulatory classification (EU AI Act, MDR/IVDR), supporting trustworthy-by-design specification of healthcare AI systems.
No takes yet. Share an insight, caveat, or question.
Angi et al. (2026) studied this question.