【作者】 彭伟；

【导师】 孙怡宁； 孙少明；

【作者基本信息】 中国科学技术大学 ， 生物物理学， 2023， 博士

【摘要】 当前,近视已经成为我国重大的公共卫生问题之一,其严重危害儿童青少年的体质健康。面对居高不下的近视患病率,近视防控已经上升到国家战略层面,并纳入地方政府绩效考核,防控工作迫在眉睫。就目前的医学发展水平来看,近视尚无疗效显著的治疗方法,更无法生理治愈。因此,如何预防近视发生、控制近视进展一直是研究的核心热点。本文以群体、个体两个维度的近视生活化防控为目标,以视觉行为/环境风险辨识引导健康生活育成作为核心思想,采用传统生物医学统计分析和新一代可解释机器学习建模相结合的技术路径,重点解决了风险评估建模依赖生物参数、传统数理分析与建模能力局限、个体预测难以指导干预实施等痛点问题,最终构建起了涵盖群体影响因素评估、个体风险预测解析、健康生活育成方案制定以及视力健康数据服务等功能的闭环式近视防控解决方案。具体的工作内容和创新点如下:1.近视的流行现状调查与相关因素分析围绕近视流行的群体性防控,本文开展了流行现状与相关因素的统计分析工作。研究设计了多阶段整群随机抽样程序,获取了中小学生样本的横断面数据,重点包括近距离屈光、睡眠和户外活动等生物行为信息。近视流行相关因素的分析采用二元逻辑回归,并按照学校、地区对样本进行逐步分层,讨论不同群体之间相关因素的差异性。研究结果显示,有效样本14588人,近视患者6636人,中小学生整体近视率为45.5%,三年级、六年级、八年级为三个近视高发的年级段,同比增长分别为12.4、10.4和11.7个百分点。总体而言,睡眠时间、白天的户外时间、家庭作业是最重要的视觉健康影响因素。此外,不同群体之间近视流行相关因素的分布和权重存在显著差异。通过大样本分层的逻辑回归分析,本研究识别出了特定人群的可调控近视风险因素,特别是与近视相关的生物行为,实现了近视群体防控策略的细化。2.近视发生预测与风险解释围绕近视流行个体防控中的“防”,本文开展了近视发生风险预警预测的研究工作。研究采用前瞻性纵向队列的实验方法,在基线时,纳入未近视的样本,并通过问卷收集样本的日常生活信息,包括人口统计学特征、父母学历与近视、学业环境与负担、生活行为与习惯,以及户外与运动等五个方面;1年随访期后,观察样本的近视发生情况。基于五个维度的生活信息以及近视次年是否发生的标签,本文采用逻辑回归和多种机器学习算法构建近视发生预测的可解释模型。研究结果显示,Catboost算法的全特征模型和节约模型AUC值分别为0.91和0.89,成功实现了未病状态下近视未来发生的准确预测。与标准方法逻辑回归以及类似的预测研究相比,模型具有预测精度高、应用成本低、操作难度小等优势。与此同时,基于可解释机器学习技术,模型建立了个体视觉行为/环境与近视发生风险之间的因果关系,从而驱动精准化、个性化生活方式管理方案的制定。3.屈光进展预测与风险解释围绕近视流行个体防控中的“控”,本文继而进行了近视人群屈光进展预测的研究。研究同样采用前瞻性纵向队列的基本思路,在基线时,测试样本的屈光度、眼轴长度、眼压等眼部生物参数,并收集初始屈光度数据及各项日常生活信息,1年随访期后,再次测试样本的屈光度。针对屈光度的年度变化,本研究将样本分为5个进展等级,并以此作为分类标签。模型的构建同样采用机器学习算法。结果显示,样本屈光度年平均进展为-0.60±0.53D,最优预测模型的平均精度为0.87。相对于眼部生物特征,日常佩戴眼镜习惯、定期视力检查等行为数据对屈光进展具有更强的预测效应。同样,应用可解释模型,本研究精准辨识出了个体层面上可调控的环境/行为风险。4.近视防控信息化平台建设针对近视防控多模态的数据服务以及上述理论研究成果的应用实施,本研究开展了信息化数据平台的建设工作。首先,基于经典的3级预防体系,设计了近视防控的生活方式管理技术内容和干预路径,明确了平台的多角色功能,并提出了信息采集-风险评估-干预执行-过程监管-优化调整的闭环式生活方式管理服务模式;随后,以多层级的生活方式管理服务为核心,开发了近视防控数据平台,具体包括服务器端、Web前端和移动端三个部分,共同实现了学校近视报告生成、学生视力健康数据管理、视觉风险评估、离线数据上传、生活方式管理方案推送与服务等功能;最后,本文依托该数据平台,在示范校的学生群体中,开展了开放式的视力健康生活育成,初步验证了生活方式管理在近视预防中的有效性和可实施性。信息化平台的应用,为生活方式管理提供了具体的技术途径,解决了近视防控生活化策略落地实施的难题。更多还原

【Abstract】 Currently,myopia has emerged as a significant public health issue in China,posing a serious threat to the physical well-being of children and adolescents.In response to the high prevalence of myopia,myopia prevention and control have risen to the national strategic level and integrated into local government performance evaluations,emphasizing the urgency of addressing this issue.The current level of medical development does not provide an effective therapy for myopia,let alone a physiological cure.Consequently,the focus of research has been on how to effectively prevent myopia and control its progression.This study aimed to address the prevention and control of myopia in daily life from both group and individual dimensions.The core idea was to use visual behavior/environmental risk identification to guide healthy life cultivation.The study employed a technical approach that combined traditional biomedical statistical analysis with new generation interpretable machine learning modeling.The focus was on addressing several challenges,including the dependence on professional physiological indicators for risk assessment modeling,limitations in traditional mathematical analysis and modeling capabilities,and the difficulty in guiding intervention implementation through individual prediction.As a result,a closed-loop myopia prevention and control technology system and service mode were developed,which included group influencing factors assessment,individual risk prediction analysis,formulation of healthy life cultivation plans,and visual health data services.The specific research contents and innovations are as follows:1.Investigation on the prevalence of myopia and analysis of related factorsThis study conducted a statistical analysis of the epidemic status and related factors concerning the group prevention and control of myopia.A multi-stage cluster random sampling procedure was designed to obtain cross-sectional data from samples of primary and secondary school students,with a focus on biological behavior information such as near refraction,sleep and outdoor activities.The associated factors with myopia were analyzed by using binary logistic regression,and the samples were stratified based on schools and regions to examine the variations in these factors across different groups.The research results revealed a total of 14,588 valid samples,of which 6,636 were myopia patients.The overall myopia rate was found to be 45.5%.The third,sixth,and eighth,grades exhibited a high incidence of myopia,with a year-on-year increase of 12.4,10.4,and 11.7 percentage points,respectively.Overall,sleep time,outdoor time during the day and homework were identified as the most significant influencing factors of visual health.Furthermore,there were variations in the distribution and impact of factors associated with myopia among different groups.Through this large-sample stratified logistic regression analysis,the study successfully identified modifiable risk factors for myopia in specific groups,particularly biological behaviors,and contributed to the refinement of group prevention and control strategies for myopia.2.Prediction and interpretation of myopia occurrence riskThis study focused on the."prevention" of individual prevention and control of myopia,and carried out the research work on the prediction of myopia occurrence.The study adopted a prospective longitudinal cohort experimental method.At baseline,non-myopia samples were recruited,and their daily life information was collected through questionnaires,including five aspects:demographics characteristics,parents’ educational background and myopia,academic environment and burden,life behaviors and habits,and outdoor and sports.After a one-year follow-up visit,the incidence of myopia in the sample was evaluated.Based on the five dimensions of life information and the label of whether myopia occurs in the next year,this study developed a interpretable prediction model for myopia occurrence using logistic regression and several machine learning algorithms.The results showed that the AUC values of the full feature model and the compact model of the Catboost algorithm were 0.91 and 0.89,respectively,successfully achieving accurate prediction of the future occurrence of myopia in a pre-disease state.In comparison to the standard method of logistic regression and similar prediction studies,the model exhibited high prediction accuracy,low application cost,and low operational difficulty.Furthermore,the model utilized interpretable machine learning technology to establish the causal relationship between individual visual behaviors/environments and myopia occurrence,driving the formulation of precise and personalized lifestyle management plans.3.Prediction and interpretation of refractive progression riskFocusing on the "control" of individual prevention and control of myopia,this study carried out the research work on predicting the progression of myopia refraction.The study also adopted a prospective longitudinal cohort.At baseline,the sample’s refraction,axial length,intraocular pressure and other ocular biological parameters were tested,and initial refractive data as well as daily life information were collected.After a one-year follow-up visit,the sample’s refraction was measured again.To consider the annual changes in refraction,this study categorized the sample into 5 progression levels and assigned them as classification labels.The prediciton model was also constructed using machine learning algorithms.The results indicated that the average annual progress of the sample refraction was-0.60±0.53D.The average accuracy of the optimization model was 0.87.Compared to biological characteristics,behavioral data such as glasses wearing habits and regular visual examinations showed a stronger predictive effect on refractive progression.Moreover,applying interpretable model,the study accurately identified modifiable environmental/behavioral risks at the individual level.4.Development of myopia prevention and control information platformIn response to the multimodal data services for myopia prevention and control,as well as the application of the research results discussed in this paper,this study focused on constructing an information data platform.Firstly,the lifestyle management technology content and intervention path for myopia prevention and control were designed based on the classic three-level prevention strategy,with the multi role functions.A closed-loop lifestyle management service mode integrating information collection,risk assessment,intervention execution,process supervision,and plan adjustment was proposed.Subsequently,with the service applied to lifestyle management as the core,a myopia prevention and control data platform was developed,consisting of three parts:the server,the web application and the mobile application,which collectively realized school myopia report generation,student vision health data management,visual risk assessment,offline data upload,and lifestyle management plans recommendation.Finally,this study utilized this data platform to carry out open life cultivation of visual health,and conducted initial verification of the effectiveness and feasibility of lifestyle management in preventing myopia.This information platform offers the specific technical approaches for managing lifestyles,and addresses the challenge of implementing life-oriented strategies for preventing and controlling myopia.更多还原