PREFER-IT frames inclusive health AI as a design obligation, not an afterthought
Artificial intelligence in healthcare often arrives wrapped in efficiency claims, but the equity consequences can be immediate if inclusion is treated as optional. PREFER-IT addresses that problem head-on by laying out a transdisciplinary framework for building inclusive health AI across the full lifecycle, from design and data collection to validation, monitoring, and governance. The work emerged from a five-day co-creation workshop in the Netherlands involving 37 experts drawn from healthcare, ethics, law, social sciences, AI, and patient advocacy.
The central message is simple and clinically relevant: an algorithm can only be as fair as the process that shaped it. If the voices of patients, clinicians, regulators, and communities are absent early, the resulting system can harden existing disparities rather than reduce them.
Eight themes emerged from a workshop built around participation
The framework rests on design thinking and participatory methods rather than a narrow technical audit. Participants organized the discussion around three broad challenges: how to integrate lived experience and stakeholder voices across the AI lifecycle, how to design data collection practices that promote fairness and avoid inequality, and how to strengthen regulatory structures that protect human rights while supporting inclusivity.
From that process came eight thematic clusters that form the PREFER-IT acronym: Participatory and co-design approaches, Representative and diverse data, Education and digital literacy, Fairness, Ethical and legal accountability, Real-world validation and feedback, Inclusive communication, and Technical interoperability. Each cluster maps onto a different layer of AI development, including humans, data, process, system, and governance.
One sentence captures the practical burden on health systems: inclusion has to be built into the machinery of AI, not appended to it after deployment.
Data, not just code, determines who benefits from AI
The framework gives unusual weight to representative and diverse data, and for good reason. Bias in healthcare AI often begins long before a model makes its first prediction. It can enter through imbalanced datasets, narrow definitions of disease, underrepresentation of marginalized groups, and collection practices that ignore social context or patient experience. Once those patterns are embedded in training data, they may be reproduced at scale with an appearance of precision.
PREFER-IT treats fairness as a structural problem rather than a purely statistical one. That approach is clinically important because real-world care unfolds in populations with different languages, health literacy levels, digital access, comorbidity profiles, and trust in medical institutions. A model that performs well in one subgroup but poorly in another may deepen disparities in triage, diagnosis, risk prediction, or treatment allocation.
Regulation and accountability sit inside the lifecycle, not beside it
The framework also places strong emphasis on ethical and legal accountability. In practice, this means governance cannot wait until an algorithm is already in circulation. Systems that affect patient care need clear oversight during development, testing, implementation, and post-deployment review. The framework links inclusivity to legal and regulatory obligations, arguing that human rights protections should shape how health AI is built and monitored.
This matters because many failures in digital health are not purely technical failures. They are failures of accountability, communication, and institutional design. If patients cannot understand how a system is used, if clinicians cannot interrogate its outputs, or if regulators lack mechanisms to monitor downstream effects, then even a technically sophisticated model may be unsafe in practice.
Real-world validation depends on feedback loops that keep patients visible
PREFER-IT explicitly includes real-world validation and feedback as one of its core themes. That is more than a post-launch quality check. AI tools in healthcare operate inside shifting clinical workflows, uneven data environments, and diverse patient populations. A model that seems acceptable in development may behave differently once it meets the complexity of routine practice.
Feedback loops are therefore essential. They allow developers and health systems to detect whether an AI tool is helping one group while underperforming for another, whether its outputs are intelligible to clinicians, and whether communication practices are respectful and accessible. In a clinical setting, these details determine whether technology supports care or quietly complicates it.
The framework also recognizes inclusive communication and education and digital literacy as separate requirements. Patients and clinicians do not engage AI on equal footing. Some users will need clearer explanations, cultural adaptation, or support navigating digital interfaces. Others will need training to interpret outputs without overreliance or mistrust. Inclusion fails when systems assume uniform digital competence.
PREFER-IT turns inclusion into a systems-level responsibility
By mapping its themes across humans, data, process, system, and governance, PREFER-IT moves beyond abstract calls for fairness. It offers a structured way to ask who is represented, who decides, who is accountable, and who may be left out. The framework is transdisciplinary by design, and that choice reflects the reality of health AI: no single discipline can solve bias, legitimacy, usability, and regulation at once.
The workshop format also matters. Co-creation creates a record of how diverse expertise can be brought into one process, which is especially relevant for AI products that will later be used in high-stakes medical settings. Participatory and co-design approaches place people with lived experience alongside technical experts at the point where decisions are still malleable. Once a model is entrenched in clinical infrastructure, changing it becomes much harder.
PREFER-IT is best read as both a framework and a warning. Health AI that ignores inclusion can amplify inequity at scale; health AI that embeds participation, fairness, and accountability from the start may be more likely to serve the people it is intended to help. The remaining challenge is implementation: how health systems, regulators, and developers operationalize these eight themes in real procurement, validation, and governance pathways will determine whether the framework changes practice or remains a conceptual map.