AIM-SAFE

AIM-SAFE develops a European multi-agent framework for trustworthy Generative AI in clinical decision support, tackling patient safety challenges in polypharmacy management and oncology treatment optimization. Its modular, locally deployed architecture leverages containerised SLM-powered agents that collaborate on bounded clinical functions, such as adverse drug event prediction, deprescribing support, pharmacogenomic reasoning, and phenoconversion modeling, while operating within hospital environments under robust privacy controls, continuous federated learning workflows, and explainable intelligence, fully aligned with European data governance laws (e.g., GDPR, AI Act, EHDS).

Beyond model development, AIM-SAFE equips researchers and clinicians with the knowledge, tools, and governance frameworks to apply Generative AI responsibly and effectively in practice. AIM-SAFE delivers open-source toolkits, FAIR datasets, and Safety Co-Pilot modules to advance trustworthy healthcare AI, ensuring ethics, transparency, and patient safety. Its key components include the CAFEIN federated learning platform, explainable AI engines enhanced with Reinforcement Learning from Human Feedback loops, and integrated safety governance through the RUC2 framework.

This approach enables privacy-preserving data harmonization across clinical partners, ensuring sensitive data remains under local control while contributing to shared models. The system integrates multimodal biomedical data from EHRs, pharmacogenomic profiles, imaging metadata, and European knowledge databases. By reinforcing European digital sovereignty, reducing reliance on non-EU AI providers, and ensuring full compliance with EU ethical, legal, and regulatory principles, AIM-SAFE transforms clinical decision support into a safe, explainable, and human-centred approach that reduces adverse drug events, improves patient care, and contributes to lower mortality and healthcare costs.

More information on the project is available in u:cris.

 

Experts of the Department working on this project: