In rare disease studies, researchers have consistently encountered issues like data shortage and diagnostic uncertainty—challenges largely stemming from these diseases’ low prevalence, varied symptoms, and complex diagnostic procedures. Traditional research methods frequently produce unreliable results, largely because of persistent issues such as limited data and clinical complexity. In recent years, the advancement of artificial intelligence, particularly foundation models, has opened up new possibilities to address these challenges. This paper reviews recent progress in the application of AI to rare disease research and, based on current literature and technological trends, proposes a comprehensive conceptual framework. The proposed framework consists of four key components: multimodal pretraining, dynamic prompting, efficient retrieval, and lightweight deployment. Although this framework remains at a theoretical stage, it aims to offer inspiration and guidance for the development of future diagnostic models. The paper seeks to provide an integrated perspective on AI research in the context of rare diseases, highlighting the potential of cross-modal integration and large model adaptation in enabling more efficient, accurate diagnostics and personalized treatment strategies).
Chen et al. (Thu,) studied this question.