基于自适应结构稀疏回归的异常脑电识别方法

PDF下载HTML

Title:

Abnormal electroencephalography recognition via adaptive structured sparse regression

作者:

吴涛¹, 刘夏², 孔祥增²

1. 福建农林大学 计算机与信息学院, 福建 福州 350117;
2. 福建农林大学 未来技术学院, 福建 福州 350117

Author(s):

WU Tao¹, LIU Xia², KONG Xiangzeng²

1. College of Computer and Information Sciences, Fujian Agriculture and Forestry University, Fuzhou 350117, China;
2. College of Future Technology, Fujian Agriculture and Forestry University, Fuzhou 350117, China

关键词:

异常脑电识别; 聚合; 结构稀疏; 统计特征; 离散小波变换; 正则化; 组特征选择; 自适应聚合

Keywords:

abnormal EEG recognition; aggregation; sparse structure; statistical features; discrete wavelet transform; regularization; group feature selection; adaptive aggregation

分类号:

TP391

DOI:

10.11992/tis.202411006

摘要:

特征约简是提升脑电信号病理解码精度的一种重要手段。然而，目前的异常脑电识别方法通常采用单尺度聚合实现特征降维，并未考虑不同尺度聚合特征之间的互补性，容易导致识别精度不高的问题；此外，现有方法在学习过程中常忽略了脑电数据特征内在的相关结构信息。为此，提出一种基于自适应结构稀疏回归的异常脑电识别模型及其两阶段构造方法。采用自适应局部和全局特征聚合机制来有效融合从原始数据提取的小波统计特征，为高精度脑电信号病理解码提供更具代表性的特征。定义了一种新的正则化稀疏模型，它通过加权L₁范数约束剔除非重要特征，同时利用加权成对结构正则化实现对高度相关特征的组选择。在实际异常脑电数据集上的实验结果表明，新方法大幅度提高了分类精度和分类结果的稳定性。

Abstract:

Feature reduction is a critical technique for enhancing the decoding accuracy of electroencephalography (EEG) pathology. However, conventional abnormal EEG identification methods generally employ single-scale aggregation for dimension reduction, overlooking the complementary nature of multiscale aggregated features. This oversight frequently leads to suboptimal classification performance. Furthermore, current approaches often overlook the inherent correlation structure present among EEG features. To this end, an abnormal EEG recognition model based on adaptive structured sparse regression and its two-stage construction method is proposed. First, a novel adaptive local-global aggregation mechanism is employed to integrate wavelet statistical features extracted from the original data. This mechanism aims to provide representative features for high-precision EEG pathology decoding. Second, a novel sparse regularization model is defined, which can automatically eliminate unimportant characteristics by the L₁-norm constraint term. Concurrently, weighted piecewise structural regularization is employed to enable the group selection of highly correlated features. The experimental findings based on real-world abnormal EEG datasets demonstrate that the proposed method significantly enhances the accuracy and stability of the classification results.