中国医疗卫生领域多模态数据集构建标准现状与展望OA

High-quality datasets serve as a cornerstone for advancing artificial intelligence(AI)applications in healthcare,while dataset standards provide the technical foundation for ensuring data quality,enabling cross-institutional sharing and facilitating a multimodal integration,and support the transition to Digital Intelli-gent Evidence-Based decisionmaking paradigm.To systematically examine the current landscape of medical multimodal dataset standards in China,identify key challenges and propose optimization pathways for the AI era,we conducted a systematic search of national public service platforms and the websites of standardization organizations,ultimately including 55 medical dataset standards published between 2019 and 2025.A struc-tured analysis was performed across six dimensions:standard type,publication year,disease coverage,modali-ty composition,integration of traditional Chinese medicine(TCM)elements,and the landscape of standard-de-veloping organizations.The results revealed that:(1)group standards predominate(63.2%),with only one in-dustry standard found,reflecting a structural pattern characterized by"group standards leading,industry stan-dards lagging";(2)disease coverage is disproportionately concentrated in a few areas,such as musculoskeletal system or connective tissue diseases(n=10)and neoplasms(n=6),with over half of disease categories covered by fewer than three standards,indicating a substantial gap relative to national strategic priorities for chronic disease prevention and control;(3)although most standards integrate multiple data types,they exhibit weak cross-modal alignment mechanisms and lack in the normative capacity to define inter-modal relationships;(4)TCM elements have been incorporated into 20 standards but in a fragmented manner,with a notable ab-sence of a collaborative Chinese-Western medicine annotation framework.Accordingly,we proposed three key optimization pathways tailored to AI applications:a knowledge graph-centric multimodal fusion standards system,a"treatment unit"-based collaborative annotation framework for integrated Chinese-Western medi-cine,and full-cycle control specifications for data element quality and privacy protection.