支持向量机在糖尿病数据中的应用研究

发布时间：2018-04-11 19:34

本文选题：支持向量机算法 + 糖尿病肾病　；参考：《北京邮电大学》2017年硕士论文

【摘要】：随着医疗行业信息化程度的不断加深,医疗数据收集与分析的手段越来越多,也越来越智能。对于快速增长的数据规模,我们需要采用更有效的手段对其进行分析处理,挖掘出数据内涵的宝贵价值。数据挖掘技术及支持向量机相关理论是实现医疗数据价值的一种重要方法。对于具体的应用场景,本论文选取了糖尿病这一现代社会中常见的慢性疾病作为算法应用研究对象。糖尿病作为一种慢性疾病,长期的经济投入和身体健康压力给个人、家庭带来了沉重的经济与身心负担。因此,在糖尿病患病早期即对其重要并发症进行筛查预测,早发现早避免,不仅可以提高糖尿病患者的生活质量,也能够降低后期肾病治疗带来的健康风险与经济压力。本文基于以上几点,展开糖尿病及糖尿病并发肾病的相关研究,构建糖尿病肾病筛查分类模型,对于糖尿病患者并发肾病的前期筛查具有较大意义。研究对象为包括北京在内的十余城市的人群体检数据,收集了健康人数据2612例,糖尿病患者数据994例,其中糖尿病并发肾病109例,糖尿病无并发肾病885例,作为原始数据集进行接下来的糖尿病肾病筛查分类模型构建。主要内容如下:一、调研已有的医疗数据研究方法、诊断指标及高危发病因素。对原始数据进行预处理,通过原始记录文字进行文本匹配标注正负样本。进行统计分析,对比糖尿病患者和健康人的各项指标数据分布,糖尿病有无并发肾病的指标数据分布,发现各组的指标均存在一定交叉,不能仅根据一项指标的值来判断是否有糖尿病并发肾病的可能。本文使用研究数据绘制ROC曲线,选择与糖尿病并发肾病具有紧密联系的12项指标,构建糖尿病并发肾病筛查模型的输入向量,用于后续糖尿病肾病筛查分类模型的建立。二、使用支持向量机分类算法构建糖尿病肾病筛查分类模型,使用训练数据训练分类超平面,对比归一化数据与未归一化数据的分类准确率,实验结果表明对数据进行归一化处理后可以有效的提高分类准确率。使用网格优化法对糖尿病肾病筛查分类模型的准确率进行进一步提高,结果表明利用网格优化得到的分类准确率都优于随机参数选择,但其搜索范围大,步长设置需根据具体数据情况确定,需引入更加智能的参数优化方法。三、使用粒子群算法对支持向量机算法惩罚参数C和核函数参数σ进行优化,对比不同核函数对于粒子群算法的优化敏感度。实验表明粒子群算法对于多项式核函数和RBF核函数的优化效果较为明显。其中PSO-RBF-SVM的交叉验证准确率达到了 94.89%,具有一定的实际应用价值与前景。
[Abstract]:For the rapid growth of data scale, we need to use more effective means to analyze and process it, and to excavate the valuable value of data connotation.Data mining technology and support vector machine theory is an important method to realize the value of medical data.For specific application scenarios, this paper selects diabetes mellitus, a common chronic disease in modern society, as the research object of algorithmic application.Therefore, early detection and prevention of important complications in the early stage of diabetic disease can not only improve the quality of life of diabetic patients, but also reduce the health risk and economic pressure brought about by late nephropathy treatment.Based on the above points, the related studies of diabetes mellitus and diabetic complicated nephropathy are carried out, and the classification model of diabetic nephropathy screening is constructed, which is of great significance for the pre-screening of diabetic nephropathy.The subjects were physical examination data of more than ten cities, including Beijing. 2612 healthy people and 994 diabetic patients were collected. Among them, 109 cases were diabetic complicated with nephropathy, 885 cases were diabetes mellitus without complicated nephropathy.The next diabetic nephropathy screening classification model was constructed as the original data set.The main contents are as follows: 1. Investigate the existing medical data research methods, diagnostic indicators and high-risk factors.The original data is preprocessed and the positive and negative samples are marked by text matching.Statistical analysis was carried out to compare the distribution of index data between diabetic patients and healthy persons, and the distribution of indexes of diabetes mellitus complicated with nephropathy.The possibility of diabetic nephropathy cannot be judged solely by the value of one indicator.In this paper, the ROC curve was drawn by using the research data, and 12 indexes closely related to diabetic nephropathy were selected to construct the input vector of diabetic nephropathy screening model, which was used to establish the following diabetic nephropathy screening classification model.Secondly, the classification model of diabetic nephropathy screening is constructed by using support vector machine classification algorithm, and the hyperplane is trained by training data, and the classification accuracy of normalized data and non-normalized data is compared.Experimental results show that the normalized data processing can effectively improve the accuracy of classification.The accuracy of diabetic nephropathy screening classification model was further improved by using mesh optimization method. The results show that the classification accuracy obtained by mesh optimization is better than that of random parameter selection, but its search scope is large.The step size setting needs to be determined according to the specific data, and a more intelligent parameter optimization method should be introduced.Thirdly, the particle swarm optimization algorithm is used to optimize the penalty parameter C and the kernel function parameter 蟽 of the support vector machine algorithm, and the sensitivity of different kernel functions to the optimization of the particle swarm optimization algorithm is compared.The experimental results show that the particle swarm optimization algorithm is more effective for polynomial kernel function and RBF kernel function optimization.The accuracy of cross-verification of PSO-RBF-SVM is 94. 89, which has some practical application value and prospect.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R587.1;TP18

【相似文献】