基于组MCP和复合MCP的人脑功能超网络分析及抑郁症分类研究  

Research on Analysis of Human Brain Functional Hyper-network and Classification of Depression Based on Group MCP and Composite MCP

作　　者：薛晓倩 李瑶 梁家瑞 Ibegbu Nnamdi Julian 孙超 郭浩 XUE Xiao-qian;LI Yao;LIANG Jia-rui;Ibegbu Nnamdi Julian;SUN Chao;GUO Hao(College of Information and Computer,Taiyuan University of Technology,Jinzhong 030600,China;College of Software,Taiyuan University of Technology,Jinzhong 030600,China)

机构地区：[1]太原理工大学信息与计算机学院,山西晋中030600 [2]太原理工大学软件学院,山西晋中030600

出　　处：《小型微型计算机系统》2022年第1期210-217,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61472270,61672374,61741212,61876124,61976150)资助;山西省重点研发计划项目(201803D31043)资助;山西省科技厅应用基础研究项目青年面上项目(201801D121135,201803D31043)资助;山西省教育厅高等学校科技创新研究项目(2016139)资助;教育部赛尔网络下一代互联网技术创新项目(NGII20170712)资助。

摘　　要：近年来,脑功能超网络模型在脑疾病诊断中多有应用.传统的脑功能超网络大多通过LASSO方法进行构建,然而由于脑区间存在组效应问题,在过去的几年里,对LASSO方法进行延伸以进一步改善超网络成为主要研究内容,由此出现各种分组模型方法.但这些方法均存在同样的问题,即惩罚函数对系数的过强压缩,导致模型中目标变量回归系数的有偏估计,使得噪声变量在压缩的同时,目标变量也进行了一定程度的压缩.因此,本文考虑到该问题,并在组效应的基础上,提出两种基于Minimax Concave Penalty(MCP)的无偏稀疏模型用以改进原有方法:组MCP方法和复合MCP方法.实验结果表明,两种方法均优于传统方法,而两种方法由于对变量是否进入模型采取了不同解决方式,因而构建的超网络结构差异较大,复合MCP方法构建的超网络的超边分布范围较为集中,而组MCP方法则较为分散;此外,复合MCP方法得到较好的分类表现和较高的分类权重.本文提出的方法所构建的脑功能超网络可以更好地表达抑郁症患者与正常对照的结构差异,具有重要的理论意义和临床价值.In recent years, the brain functional hyper-network model has been widely used in the diagnosis of brain diseases.Traditional brain functional hyper-networks are mostly constructed by the LASSO method.However, due to the problem of group effects in brain areas, in the past few years, the extension of the LASSO method to further improve the hyper-network has become the main research content, and various grouping models have appeared.But these methods all have the same problem, that is, the penalty function compresses the coefficients too strongly, which leads to the biased estimation of the regression coefficients of the target variables in the model, so that while the noise variables are compressed, the target variables are also compressed to a certain extent.Therefore, this paper takes this problem into account and proposes two unbiased sparse models based on Minimax Concave Penalty(MCP)and also based on the group effect to improve the original method: group MCP method and composite MCP method.The experimental results show that the two methods are better than the traditional methods, and because the two methods adopt different solutions to whether the variables enter the model, the structure of the hyper-network constructed is quite different.The hyper-edge distribution of the hyper-network constructed by the composite MCP method is more concentrated, while the group MCP method is more scattered;in addition, The composite MCP method has better classification performance and higher classification weight.The brain functional hyper-network constructed by the method proposed in this paper can better express the structural difference between depression patients and normal controls, which has important theoretical and clinical value.

关 键 词：无偏稀疏模型 组效应 组MCP 复合MCP 分类 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]