The project
Off-the-shelf Generative AI creates synthetic data, but it is not ready for clinical use.
Identified Gaps
This research identifies key gaps in developing trustworthy generative models and ML methods for healthcare:
- Absence of standardized synthetic data evaluation frameworks
- Trust deficits in healthcare generative models
- Resource intensiveness of current approaches
- Design-induced opaqueness in AI decision-making
THUNDER Vision & Core Pillars
This framework illustrates a structured approach to Trustworthy & Responsible Generative AI, visualized as a classical temple supported by three core pillars and a solid foundation. The structure is Anchored in Expert Medical Knowledge, which serves as the base for the entire system. At the center of the architecture is the middle pillar, Knowledge-Guided Generation (RI-2: The Engine), which utilizes RAG, LLM, and expert-defined knowledge to drive the AI's output. To its left stands the pillar of Standardised Evaluation (RI-1: The Ruler), focusing on the metrics of Fidelity, Utility, and Privacy. To the right, the third pillar emphasizes Frugal & Interpretable by-design learning (RI-3: The Vehicle), incorporating technical approaches like NAS, SNNs, and Explainability. Together, these three Research Initiatives support the overarching roof of the structure, representing the ultimate goal of achieving a safe and reliable generative AI system.
Our Approach
Knowledge-Guided Generative Models
Developing advanced generative AI models that incorporate domain knowledge to produce trustworthy and clinically relevant synthetic data.
Frugal & Interpretable Design
Creating resource-efficient machine learning models that are interpretable by design, enabling transparent decision-making in healthcare.
Standardized Evaluation Metrics
Defining comprehensive metrics for evaluating synthetic healthcare data quality, including semantic validity, fairness, and privacy preservation.
Sepsis as Target Application
Focusing on sepsis—a WHO-designated global health priority—to demonstrate and validate the framework's real-world clinical impact.
Work packages
-
WP1Management
-
WP2Training, communication, & dissemination
-
WP3Foundations of metrics for synthetic data evaluation
-
WP4Pioneering knowledge-guided generative models for trustworthy synthetic data
-
WP5Advanced frugal-by-design ML for actionable insights
-
WP6Iterative clinical validation and refinement of AI models
