大数据驱动的婴幼儿视力智能评估系统的研发和应用

The development of intelligent medical system driven by big data is the core power of the new generation of medical revolution. With the support of the major R&D program (91546101), we developed an intelligent diagnosis and treatment decision-making system (Nature Biomedical Engineering, 2017) for congenital cataract, completing a series of digital diagnosis and treatment work. However, the existing medical large data intelligent analysis technology is limited to formatting electronic medical record data and static medical images, and also difficult to learn the dynamic changes of disease representation effectively, which has become an important challenge for the reform of large data-driven diagnosis and treatment model. This project aims to focus on the problem of visual impairment assessment in infants and young children. Based on the large-scale video data of 4196 infants and young children's behavioral phenotypes accumulated in the previous period, the core techniques of medical artificial intelligence model with video clips as the basic unit are created around three medical signs (strabismus, nystagmus and compensatory head position) , which are highly related to visual function. In order to improve the level of early prevention and treatment of infantile eye diseases, we here developed an intelligent evaluation system for infant vision, and applied it to a smart phone for large-scale screening and promotion.

大数据驱动的智能化医疗系统研发是新一代医学革命的核心动力。本团队前期在重大研发计划培育项目（91546101）支持下，研发了先天性白内障智能诊疗决策系统（Nature Biomedical Engineering，2017），并完成一系列的数字诊疗工作。然而，现有医学大数据智能分析技术，局限于格式化电子病历数据及静态医学图像，难以有效学习不断动态变化的疾病表征，这成为了大数据驱动诊治模式变革面对的重要挑战。本项目拟聚焦婴幼儿视觉损伤评估这一难题，以前期积累的4196例婴幼儿的行为表型视频大数据为基础，围绕三个视功能高相关的医学体征（斜视、眼球震颤、代偿头位），创建以视频片段为基本单位的医学人工智能模型核心技术，研发婴幼儿视力智能评估系统，并应用在智能手机以进行大规模筛查推广，全面提高婴幼儿眼病早防早治水平，同时为医学视频大数据分析与人工智能转化应用提供参考。

大数据驱动的智能化医疗系统研发是新一代医学革命的核心动力。然而，现有医学大数据智能分析技术，局限于格式化电子病历数据及静态医学图像，难以有效学习不断动态变化的疾病表征，这成为了大数据驱动诊治模式变革面对的重要挑战。.本项目聚焦婴幼儿视觉损伤评估这一难题，前期以积累的4196例婴幼儿的行为表型视频大数据为资源基础，围绕三个视功能高相关的医学体征（斜视、眼球震颤、代偿头位），创建以视频片段为基本单位的医学人工智能模型核心技术。.本项目定量对比了不同视功能群体4大类、13个行为特征的发生频率及严重程度，研究结果显示，婴幼儿视功能正常与否，确实会带来行为模式上的差异。在进一步研究中，明确了斜视、眼球震颤、代偿头位等11个标志性的医学行为体征与婴幼儿视觉损伤的量化关系。依托已挖掘出的量化关系，本项目运用了全长度视频训练的深度学习算法（时序分割网络），将标记好的视频片段进行随机分割以供机器学习，研发婴幼儿视功能智能评估系统，判断准确率达到了80%以上，相关研究成果发表于《自然生物医学工程》（Nature Biomedical Engineering，2019）。.本项目为建立基于手机客户端的婴幼儿智能视力筛查应用场景提供了技术基础，将面向社会提供免费的婴幼儿视功能评估服务以进行大规模筛查推广，全面提高婴幼儿眼病早防早治水平，同时为医学视频大数据分析与人工智能转化应用提供参考，具有重要的科学价值和社会意义。