AI-PROGNOSIS_D2.6 / Trustworthy AI development and evaluation framework (updated version)

Executive summary

This deliverable presents the updated version of the Trustworthy Artificial Intelligence (AI) Development and Evaluation Framework for the AI-PROGNOSIS project. It builds upon the foundational version (D2.2 ‘Trustworthy AI development and evaluation framework (fundamental version)’) by incorporating new dimensions, updated practices, and consolidated technical reporting from all development partners. The framework continues to define core dimensions of trustworthy AI—such as robustness, generalisation, explainability, transparency, reproducibility, fairness, privacy, and accountability—while also introducing environmental sustainability as a new dimension.

The updated version places significant emphasis on alignment with existing and emerging regulatory frameworks. It highlights adherence to the recently approved EU AI Act, incorporates best practices outlined in FDA guidance on AI-enabled medical devices, integrates the FUTURE-AI international consensus guidelines, and highlights ISO/IEC standards that are relevant to trustworthy AI development. Additionally, it promotes the use of the TRIPOD-AI checklist as a primary standard for transparent and comprehensive reporting of AI models in healthcare.

In this updated version of the deliverable, we further elaborate the practices to be followed throughout the entire lifecycle of the AI-PROGNOSIS project. We specifically focus on best practices for clinician-facing explainable AI and the introduction of LLM-based explanation pathways. A new component on risk management has also been included, ensuring that risks are systematically identified, monitored, and addressed throughout the AI lifecycle. Also, this deliverable consolidates compliance reporting across model and application development partners, capturing implementation progress and future plan. Together, these updates provide a comprehensive, actionable, and forward-looking framework to guide the trustworthy development, evaluation, and deployment of AI tools in the AI-PROGNOSIS ecosystem and similar healthcare applications.