【作者】 于洋；

【导师】 李先能；

【作者基本信息】 大连理工大学 ， 系统工程， 2020， 硕士

【摘要】 近年来,中国在健康医疗领域发起了一场“供给侧”改革,成立了卫生与健康委员会,出台了“健康中国2030”、家庭医生签约模式等指导性政策法规,努力建立更广泛的人健康服务体系。健康体格检查是面向社会全人群的基础性健康服务,目前基于健康体检数据的研究主要面临两个问题,一是随着健康体检服务的普及,健康体检数据虽然覆盖面越来越广、数据量越来越大,但是缺乏与疾病的直接关联,并且数据专业性不强、项目差异较大,采用机器学习方法开展疾病风险预测通常难以得到较好的精度;二是随着疾病筛查技术的不断更新,健康体检项目组合越来越丰富、价格差异越来越大,通常情况下依赖于医疗经验的健康体检项目设计方法已经不适用,难以甄别出最具信息价值和全面反映身体状况的体检项目,降低了健康体检的效益。针对上述问题,本文从健康体检数据的病史记录出发,通过对受检者病史的特点和分布进行总结梳理,建立了健康体检数据与6类慢性疾病的特征关联;针对健康体检数据的数据量大、阳性样本率低、特征稀疏特点,本文应用轻型梯度提升(LightGBM)算法以及单边梯度采样(GOSS)和互斥特征绑定(EFB)技术,建立了基于健康体检数据的慢性疾病风险预测模型,并将预测结果与可拓展梯度提升(XGBoost)算法进行对比;提出基于特征重要性和价格的体检项目排序方法,提取与慢性疾病风险预测关联程度最高的100项特征,整理和归类为3类、42个体检项目,实现了健康体检项目的优化;总结健康体检项目的特点规律,提出健康体检项目选择的一般原则,针对家庭医生体检、职工体检和慢病体检开展体检套餐设计。本文实现了适用于健康人群的慢性疾病风险预测方法,通过与XGBoost算法进行对比,证明了该方法具有更好的预测精度,在解读异常数据之外提出了一种新的、有效的健康体检数据分析方法。针对健康体检套餐组合依赖于专家经验和调查问卷等主观认识,缺乏系统性的问题,提出了基于特征重要性和价格影响因素的健康体检项目排序方法,实现了数据驱动的健康体检项目优化,完成了适用于不同群体的健康体检套餐设计。更多还原

【Abstract】 In recent years,China has initiated a “supply-side” reform in the field of health care,established the Health and Wellness Committee,and introduced guiding policies and regulations such as “Healthy China 2030” and the family doctor contract model in an effort to establish a wider range of human health service system.Health checkups are basic health services for the entire population of society.At present,there are two main problems faced by research based on health checkup data.First,with the popularization of health checkup services,the coverage and amount of the health checkup data have become increasingly large,but it still lacks a direct correlation with the disease and has a low positive sample rate and sparse features.It is often difficult to obtain good accuracy when using machine learning methods to predict disease risk.Second,as disease screening technologies continue to update,the health portfolio of medical examination items is becoming more and more abundant,and the price difference is getting widening.Generally,the design methods of health medical examination items that rely on medical experience are no longer applicable.It is difficult to identify the most informative and comprehensive physical examination items,which reduces the benefits of health checkups.To solve the above problems,this study focuses on the medical history records of the health examination data and the characteristics and distribution of the patient’s medical history are summarized to establish the association between the health examination data and the characteristics of 6 chronic diseases;considering the large quantity,low positive sample rate,and sparse feature characteristics of the health examination data,the LightGBM algorithm,Gradient-based One-Side Sampling(GOSS),and Exclusive Feature Binding(EFB)technology are applied to establish a chronic disease risk prediction model based on health examination data,and the results were compared with those of the XGBoost algorithm;this study further proposes a method for ranking physical examination items based on feature importance and price,and 100 features most associated with the prediction of chronic disease risk are extracted and classified into 3 groups and 42 examination items,which have achieved the optimization of health examination items;the characteristics and laws of health checkup items are finally summarized and the general principles of health checkup items selection are provided,which can assist the design of checkup packages for family doctors,staff checkups,and chronic disease checkups.In this study,a method for predicting the risk of chronic diseases suitable for healthy people is established.By comparing with XGBoost algorithm,it has been proved that the method proposed by this study has a better prediction accuracy,and besides interpreting abnormal data,it is also a novel effective method for analyzing physical examination data.As for the issues that health checkup package relies on subjective knowledge such as expert experience and questionnaires,and lacks of systemic,a method for sequencing health examination items based on feature importance and price influencing factors has been proposed to optimize the health checkup items and accomplish the application designed of health checkup packages for different groups.更多还原