摘要：

在基于脑电的卒中分类诊断任务中, 以卷积神经网络为基础的深度模型得到广泛应用, 但由于卒中类别病患样本数量少, 导致数据集类别不平衡, 降低了分类精度。现有的少数类数据增强方法大多采用生成对抗网络(GAN), 生成效果一般, 虽然可通过引入缩放点乘注意力改善样本生成质量, 但存储及运算代价往往较大。针对此问题, 构建一种基于线性有效注意力的渐进式数据增强算法LESA-CGAN。首先, 算法采用双层自编码条件生成对抗网络架构, 分别进行脑电标签特征提取及脑电样本生成, 并使生成过程逐层精细化; 其次, 通过在编码部分引入线性有效自注意力(LESA)模块, 加强脑电的标签隐层特征提取, 并降低网络整体的运算复杂度。消融与对比实验结果表明, 在合理的编码层数与生成数据比例下, LESA-CGAN与其他基准方法相比计算资源占用较少, 且在样本生成质量指标上实现了10%的性能提升, 各频段生成的脑电特征样本均更加自然, 同时将病患分类的准确率和敏感度提高到了98.85%和98.79%。

关键词: 脑卒中, 脑电, 生成对抗网络, 自注意力机制, 线性有效自注意力

Abstract:

Electroencephalogram(EEG)-based Convolutional Neural Network(CNN) models are widely used to classify and diagnose stroke. However, owing to the small number of stroke patient samples, the category imbalance of the dataset reduces the classification accuracy. Most existing minority-class data enhancement methods utilize Generative Adversarial Network(GAN), and the generation effect is average. Although the quality of data synthesis can be improved by applying dot product attention, storage and computing costs are often high. To address this problem, a progressive data-enhancement algorithm, LESA-CGAN, based on linear effective attention is constructed. First, the algorithm adopts a two-layer autoencoding conditional GAN architecture to extract EEG label features, generate EEG samples, and refine the generation process layer by layer. Second, by introducing a Linear Effective Self-Attention(LESA) module in the encoding part, it enhances the extracted hidden layer features of the EEG labels and reduces the overall computational complexity of the network. The results of the ablation and comparison experiments indicate that under the conditions of a reasonable number of coding layers and proportion of generated data, LESA-CGAN requires fewer computing resources compared with other benchmark methods and achieves a 10% performance improvement in sample generation quality indicators. The EEG feature samples generated in each frequency band are more natural, and the accuracy and sensitivity of patient classification increase to 98.85% and 98.79%, respectively.

Key words: stroke, Electroencephalogram(EEG), Generative Adversarial Network(GAN), self-attention mechanism, Linear Effective Self-Attention(LESA)