基于组选择的近似无偏稀疏脑功能超网络模型构建与分类研究

针对LASSO方法构建脑功能超网络模型缺乏组效应解释能力和网络有偏性问题,提出了两种基于组变量选择的近似无偏稀疏脑功能超网络模型来改善超网络的构建,分别为组最小最大凹惩罚方法和组平滑剪裁的绝对值偏差方法,并将其分别应用于抑郁症的分类研究中。分类结果显示,两种方法的分类表现均优于传统超网络模型,且组最小最大凹惩罚方法的分类准确率最高,达到86.36%。结果表明若想构建有效的脑功能超网络模型,不仅需要考虑脑区间组效应的解释能力,还需考虑模型变量选择的有偏性问题。而且在考虑到超网络有偏性的基础上,选取较为宽松的惩罚方式来选取目标变量,则可更精确地表征人脑的复杂高阶多元交互信息。

Construction and classification research of approximately unbiased sparse brain functional hypernetwork model based on group selection

Aiming at the problem of the lack of group effecting explanation ability and network bias in the brain functional hyper-network model based on the LASSO method,this paper proposed two approximate unbiased sparse brain function hypernetwork models based on group variable selection to improve the construction of hypernetworks,such as the group minimax con-cave penalty(gMCP)method and the group smoothly clipped absolute deviation method(gSCAD).Then it applied them to the classification research of depression.Classification results show that the classification performance of the two proposed methods are better than the traditional hyper network model,and the classification accuracy of gMCP is the highest,reaching 86.36%.These results indicate that in order to build an effective brain functional hypernetwork model,not only the group effecting should be considered,the bias in the selection of model variables should also be considered.Moreover,taking into account the bias of the hypernetwork,it can more accurately represent the complex and high-order multivariate interactive information of the human brain that choose a more relaxed punishment method for the selection the target variable.