面向复杂因素的心血管疾病预测研究

发布时间：2018-04-14 03:30

本文选题：心血管疾病 + CVD　；参考：《郑州大学》2017年硕士论文

【摘要】：心血管疾病(Cardiovascular disease,CVD)作为城乡居民健康的头号杀手,在基层社区对CVD进行初筛需求强烈。通过CVD风险预测制定针对性的干预措施,能有效降低发病率,其中核心的技术环节是建立准确性高的预测模型。经济社会发展推动人们生活方式和环境的快速变化,使得CVD危险因素变得复杂,建立新的面向复杂因素的CVD预测模型具有重要应用价值和现实意义。传统的CVD预测用Logistic回归、Cox回归等数理预测法建立公式化的预测模型,受公式的线性拟合能力约束,纳入的因素均为连续型和二分类变量,以非线性数据表述的多分类复杂因素不能适用。而复杂因素中可能包含了重要的潜在信息,对准确地预测CVD贡献较大。神经网络在非线性数据处理中的学习能力、适应能力较强,为解决上述问题提供了思路。以实现面向复杂因素的CVD预测为目标,首先分析回归模型无法面向复杂因素进行CVD预测的具体原因,利用哑变量将复杂因素线性化解决该问题;其次利用浅层神经网络将非线性的复杂因素映射到高维空间进行线性的回归拟合,以实现面向复杂因素的CVD预测,并研究提高模型的AUC(The area under ROC curve)值;最后设计基于深度学习的CVD预测,利用无监督学习改善浅层神经网络参数的初始化策略,降低预测结果的方差值。主要研究工作及结果如下:(1)从模型机理上分析回归模型对复杂因素不适用的原因,采用对复杂因素设置哑变量的方法应用回归模型。实验以Logistic回归为例,包含复杂因素条件下,改进模型的AUC值由Logistic的0.7634、0.6700提升到了0.8784、0.7999,符合传统回归模型的0.78~0.86,表明回归模型不能直接纳入复杂因素。(2)针对CVD复杂因素数据的非线性特征,建立基于浅层神经网络的CVD预测模型,并通过改进网络参数的初始值提高模型预测的准确性。实验结果显示,模型的平均AUC值提高到了0.9024、0.8423。(3)利用深度学习逐层提高CVD数据特征的表达能力,以学习到的最优参数初始化神经网络,解决神经网络参数随机初始化导致的预测不稳定问题,降低预测结果的方差值,提高模型的预测稳定性。实验结果显示,模型的预测结果方差由12.665、9.051降为5.723、4.642,AUC值进一步提高到0.9198、0.8959。
[Abstract]:Cardiovascular disease (CVD), as the leading killer of urban and rural residents' health, has a strong demand for primary screening of CVD in grass-roots communities.It is effective to reduce the incidence of CVD risk by making relevant intervention measures, among which the key technique is to set up accurate prediction model.The development of economy and society promotes the rapid change of people's life style and environment, and makes the risk factors of CVD become more complicated. It is of great value and practical significance to establish a new CVD prediction model for complex factors.The traditional CVD prediction model is formulated by Logistic regression and Cox regression, which is constrained by the linear fitting ability of the formula, and the factors included are continuous and binary variables.Multi-classification complex factors expressed by nonlinear data cannot be applied.Complex factors may contain important potential information and contribute greatly to accurate prediction of CVD.The learning ability and adaptability of neural network in nonlinear data processing is strong, which provides a way to solve the above problems.In order to realize the CVD prediction for complex factors, firstly, the specific reasons why the regression model can not predict CVD for complex factors are analyzed, and the complex factors are linearized by dummy variables to solve the problem.Secondly, the shallow neural network is used to map the nonlinear complex factors to the high-dimensional space for linear regression fitting to realize the CVD prediction for complex factors, and to improve the AUC(The area under ROC curvevalue of the model.Finally, CVD prediction based on deep learning is designed to improve the initialization strategy of shallow neural network parameters by using unsupervised learning to reduce the square difference of prediction results.The main research work and results are as follows: (1) the reason why the regression model is not applicable to complex factors is analyzed from the model mechanism, and the method of setting mute variables for complex factors is used to apply the regression model.The experiment takes Logistic regression as an example, under the condition of complex factors,The AUC value of the improved model was raised from 0.7634 / 0.6700 of Logistic to 0.8784 / 0.7999, which is in line with the traditional regression model's 0.780.86, which indicates that the regression model can not directly incorporate the complex factors. (2) aiming at the nonlinear characteristics of CVD complex factor data, a CVD prediction model based on shallow neural network is established.The accuracy of model prediction is improved by improving the initial value of network parameters.The experimental results show that the average AUC value of the model is increased to 0.9024 and 0.8423.3) the expression ability of CVD data features is improved layer by layer by depth learning, and the neural network is initialized with the optimal parameters.The problem of prediction instability caused by random initialization of neural network parameters is solved, the square difference of prediction results is reduced, and the prediction stability of the model is improved.The experimental results show that the variance of the predicted results of the model decreased from 12.665U 9.051 to 5.723N4.642AUC further increased to 0.9198N0.8959.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R54;O212.1

【相似文献】