General recommendations

Seven FUTURE-AI recommendations apply across all six principles Lekadir et al. (2025).

G1: Engage interdisciplinary stakeholders throughout the lifecycle¶

Research ++, deployable ++. Involve healthcare professionals, patients, ethicists, data managers and legal experts from design to deployment, through working groups, advisory boards, interviews, co-creation meetings or surveys.

→ See: Stakeholder engagement

G2: Protect data privacy and security¶

Research ++, deployable ++. Use privacy-enhancing techniques, carry out a data protection impact assessment (DPIA, the risk assessment the GDPR European Parliament and Council (2016) requires before high-risk processing such as AI on health data), control access after deployment and protect the tool against attacks. In Dutch healthcare, NEN 7510 is the standard for information security.

→ See: Privacy and security

G3: Address the risks you identified¶

Research ++, deployable ++. Turn the risks found during design into measures in your modelling plan, such as data augmentation for robustness, transfer learning for new settings, or resampling to correct bias between groups.

→ See: Risk management planning, Addressing AI risks

G4: Define an adequate evaluation plan¶

Research ++, deployable ++. Before you evaluate, fix the test data (kept strictly apart from the training data), the metrics, and a reference to compare against, such as current practice.

→ See: Evaluation planning

G5: Identify and comply with regulatory requirements¶

Research +, deployable ++. Identify early which rules apply, because they shape design choices. In the EU these include the GDPR, the MDR European Parliament and Council (2017) or IVDR European Parliament and Council (2017) and the EU AI Act European Parliament and Council (2024). The paper says the AI Act classes all healthcare AI as high-risk; the Act itself is narrower (see the FUTURE-AI overview).

→ See: Legal and regulatory considerations, Regulatory compliance

G6: Investigate and address application-specific ethical issues¶

Research +, deployable ++. Beyond privacy, transparency, equity and autonomy, ask developers, clinicians and ethicists which ethical issues your particular application raises.

→ See: Ethical review and approval

G7: Investigate and address social and societal issues¶

Research +, deployable +. Consider effects on working conditions, professional skills (including their loss), relations between patients and professionals, and the tool’s carbon footprint.

→ See: Social and societal impact