Guide for design - The ELSA Way

Before anyone writes code or requests data, a team has to decide which clinical problem it is solving, for whom, with what data, and under which rules. This section covers those decisions. A wrong choice here (the wrong outcome to predict, a missing user group, an unplanned legal basis) is expensive to undo once a model exists.

Illustration showing the components of project design, including problem definition, stakeholder engagement, and planning stages. — Figure 1:The Turing Way Community. This illustration is created by Scriberia with The Turing Way community, used under a CC-BY 4.0 licence. DOI: 10.5281/zenodo.3332807.

Running case: sepsis early warning

The design chapters follow one fictional project. A regional hospital in the Northern Netherlands wants to detect sepsis earlier on its adult wards. Its data science team, working with a university research partner, builds a model that estimates each patient’s sepsis risk every hour from vital signs, lab results and nursing observations in the hospital’s electronic health record (EHR). When the score crosses a threshold, the ward nurse and the physician on call receive an alert. The clinicians decide what to do.

The people with a stake include ward nurses, internists and intensivists, the rapid response team, patients and a patient advisory panel, the hospital’s data protection officer (in Dutch, the FG), the clinical physics and medical technology department, IT, and the board. Each chapter returns to this case to show what its advice looks like in practice.

FUTURE-AI Lekadir et al. (2025) places ten of its recommendations in the design phase: G1 (engage interdisciplinary stakeholders), Us1 (intended use and user requirements), Un1 (intended clinical settings and cross-setting variations), Un2 (community-defined standards), R1 (sources of data variation), F1 (potential sources of bias), E1 (explainability needs and requirements), T1 (risk management process), G6 (application-specific ethical issues) and G7 (social and societal issues). The chapter on legal and regulatory questions also starts work on G5 (regulatory requirements). FUTURE-AI places G5 in deployment, but the regulatory route shapes design choices long before go-live.

Chapters in this section¶

Chapter	What you will learn
Problem definition and scoping	How to state a clinical problem, scope the AI tool and set success criteria before you see any data
Stakeholder engagement	Who to involve, when, and how, including patients
Intended use and user requirements	How to describe settings, user groups, interaction and explainability needs
Ethical review and approval	Which review your project needs in the Netherlands, and which ethical questions to work through
Data strategy	How to plan representative data and labels before collection starts
Legal and regulatory considerations	How the EU AI Act, MDR/IVDR, GDPR and Dutch law apply to your tool
Risk management planning	How to open a risk management file and keep it up to date
Identifying sources of bias	Where bias enters through data, labels and problem framing
Social and societal impact	Effects on staff, equity, the environment and the patient-clinician relationship

References¶

Lekadir, K., Frangi, A. F., Porras, A. R., & others. (2025). FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare. BMJ, 388, e081554. 10.1136/bmj-2024-081554