基于非凸稀疏优化的影像遗传学特征筛选方法  

Image Genetic Feature Screening Method Based on Non-convex Sparse Optimization

作　　者：赵迎利 朱旭 ZHAO Ying-li;ZHU Xu(Mathematics Teaching Department,Xi’an Jiaotong University City College,Xi’an 710018,China;School of Mathematics and Statistics,Xi’an Jiaotong University,Xi’an 710049,China)

机构地区：[1]西安交通大学城市学院数学教学部,陕西西安710018 [2]西安交通大学数学与统计学院,陕西西安710049

出　　处：《计算机技术与发展》2025年第4期135-140,共6页Computer Technology and Development

基　　金：西安交通大学城市学院2024年度校级科研项目(2024Y02);2024年陕西省体育局常规课题(20240184);中国民办教育协会规划课题(CANFZG22137)。

摘　　要：为了从海量的影像学数据与遗传学数据筛选出与精神类疾病紧密相关的生物标志物特征,传统方法在处理此类多模态高维数据时往往过度惩罚系数导致过稀疏,进而引入估计偏差。针对此问题,提出一种有效的影像遗传学特征筛选方法,该方法通过将改进的基于双曲正切函数的近似l 0范数作为正则项惩罚典型相关分析,将最大化两类数据间的相关性作为优化目标,从而避免过度惩罚导致关键特征丢失,最终实现无偏估计和稳定特征筛选的双重目标。为了高效求解此优化问题,采用基于局部二次逼近技术和交替凸搜索方法的优化算法。模拟数据实验结果表明,该方法展现了更高的特征筛选的全面性和相关系数估计的准确性,精神分裂症数据实验结果表明,该方法成功地识别出了包括IL10在内的8种潜在候选基因,以及丘脑在内的7个风险脑区。In order to screen for biomarker features closely related to mental disorders from massive imaging and genetic data,traditional methods often excessively penalize coefficients when processing such multi-modal high-dimensional data,resulting in sparsity and introducing estimation bias.To address this issue,an effective image genetics feature screening method is proposed.This method penalizes canonical correlation analysis by using an improved approximate l 0 norm based on hyperbolic tangent function as the regularization term,and maximizes the correlation between two types of data as the optimization objective,thereby avoiding the loss of key features caused by excessive punishment.Ultimately,the dual objectives of unbiased estimation and stable feature screening are achieved.In order to efficiently solve this optimization problem,an optimization algorithm based on local quadratic approximation technique and alternating convex search method is adopted.The simulation data experiment results show that the proposed method exhibits higher comprehensiveness in feature screening and accuracy in estimating correlation coefficients.The schizophrenia data experiment results indicate that the proposed method successfully identified 8 potential candidate genes,including IL10,as well as 7 risk brain regions within the thalamus.

关 键 词：非凸惩罚 改进的双曲正切函数 典型相关分析 特征筛选 精神分裂症 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]