Fostering Trust in AI driven Healthcare: SecUre and uNbiased knowleDge guided gEneRative AI

forthcoming

Machine learning (ML) offers transformative opportunities for healthcare, with applications ranging from precision medicine to operational optimization. However, progress is constrained by limited access to diverse, high-quality datasets, exacerbated by fragmentation, data scarcity, and stringent privacy regulations. Traditional data augmentation methods fail to fully capture the complexity and heterogeneity of healthcare data. Generative AI, particularly large language models (LLMs), offers a promising alternative by synthesizing realistic datasets while addressing data scarcity. Yet, their adoption in healthcare is hindered by critical concerns about trustworthiness, including semantic validity, fairness, bias mitigation, fidelity, privacy preservation, and real-world utility. This research identifies key gaps in developing trustworthy generative models and ML methods for healthcare. These include the absence of standardized synthetic data evaluation frameworks, trust deficits in healthcare generative models, resource intensiveness, and design-induced opaqueness. The overall objective of the THUNDER project is to forge a comprehensive framework for trustworthy and responsible generative AI in healthcare. This will be achieved by defining (i) standardized evaluation metrics, (ii) developing advanced knowledge-guided generative models, and (iii) creating a fully frugal-by-design and interpretable-by-design learning models. These efforts, driven by interdisciplinary and intersectoral mobility and knowledge exchange, will establish a new paradigm for AI-driven healthcare. We will target Sepsis as a global health priority identified by the World Health Organization (WHO).