基于多标签体检数据的疾病风险分析方法研究

发布时间：2019-03-15 21:01

【摘要】：健康体检是疾病预防很重要的环节。医生可以根据个人的健康体检结果及时分析潜在的病症,进而对其进行健康指导。针对健康体检结果的分析,传统的处理方式为有经验的医生针对身体各部分的体检结果给出整体的健康状况和疾病风险分析,随着数据的日益增多,以及医生经验的良莠不齐等现状,人工的分析方法在效率和精度方面无法满足日益增多的体检需求。随着数据挖掘技术的发展,人工智能、机器学习方法已被广泛用于医疗辅助诊断和疾病风险分析。数据预处理是机器学习的重要环节之一,在医疗体检数据中,体检结果往往存在个体性差异。体现在对于某一个特征,整个人群的特征数值分布的标准差相对较大,而且在均值以下的数量远超在均值以上的数量,表现为数据分布极为不平稳。然而,传统化的数据归一化方法并不能很好的规避这一问题。通过数学变换可以较好地解决这一问题并在一定程度上提高模型的收敛速度以及精度。本文主要工作包括:1、提出FN(Fusion normalization)方法来进行特征的平稳化处理,并将特征值归一化至(0,1)区间;2、针对多标签问题,本文分别建立以SVM、GBDT、LR为基础分类器的三种组合模型SVMs、GBDTs、LRs来处理医学多标签数据;3、针对医学体检指标值正常人群数量大于异常人群数量,而造成的数据不平衡问题,本文根据标签数据的比值采用对不同的标签设置不同的惩罚因子的方法来处理。本文的数据集包含性别、空腹血糖等62个特征,高血压、糖尿病、脂肪肝3个标签。数据集中数据类型有字符型和数值型。实验结果表明:FN(Fusion normalization)方法处理过后的体检数据相比于不做归一化的的数据,Max_min归一化以及标准归一化方法,在组合模型SVMs、GBDTs、LRs上的准确率均有不同程度的提高。
[Abstract]:Health check-up is a very important part of disease prevention. Doctors can analyze the underlying symptoms on the basis of individual health check-up results, and then provide health guidance to them. According to the analysis of the health examination results, the traditional treatment method is to give the whole health condition and disease risk analysis for the experienced doctors according to the physical examination results of each part of the body. With the increasing of the data, As well as the mixed experience of doctors and so on, the artificial analysis method can not meet the increasing demand for physical examination in terms of efficiency and accuracy. With the development of data mining technology, artificial intelligence and machine learning methods have been widely used in medical assistant diagnosis and disease risk analysis. Data preprocessing is one of the important links in machine learning. In medical physical examination data, there are often individual differences in the results of physical examination. For a certain feature, the standard deviation of the distribution of the characteristic values of the whole population is relatively large, and the number below the mean value is far higher than the number above the mean value, which shows that the distribution of the data is extremely uneven. However, the traditional method of data normalization is not a good way to avoid this problem. This problem can be solved by mathematical transformation and the convergence speed and precision of the model can be improved to a certain extent. The main work of this paper is as follows: (1) the FN (Fusion normalization) method is proposed to stabilize the features and normalize the eigenvalues to (0,1); 2. Aiming at the multi-label problem, this paper establishes three combination models based on SVM,GBDT,LR classifier, SVMs,GBDTs,LRs, to deal with medical multi-label data. 3. In view of the imbalance of data caused by the number of normal population is larger than that of abnormal population, according to the ratio of label data, the method of setting different punishment factors for different labels is adopted to deal with the problem. This data set contains 62 features such as gender, fasting blood glucose, hypertension, diabetes, and fatty liver. The data types in the dataset are character type and numeric type. The experimental results show that the accuracy of the: FN (Fusion normalization) method in combination model SVMs,GBDTs,LRs is improved to some extent compared with the non-normalized data, the Max_min normalization method and the standard normalization method.
【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】：R194.3;TP18

【参考文献】