基于Transformer的无监督脑电图癫痫发作检测研究

Intelligent Detection Method of Multi-Channel Long-Term Epileptic EEG Signals Based on RA-BiGRU

摘要: 针对现有癫痫发作检测方法严重依赖神经科专家高质量人工标注、跨患者泛化能力差且难以捕捉脑电信号长程时序依赖的核心临床痛点，本文提出一种基于改进Transformer的无监督脑电图癫痫发作检测模型。该模型无需大量标注数据即可从海量无标注脑电记录中学习通用特征，通过多尺度自适应位置编码增强模型对脑电多尺度时序特征的捕捉能力，设计通道–时序双维度稀疏注意力机制在抑制噪声和伪影干扰的同时高效建模通道间关联和长程时序依赖，构建基于异常感知的掩码重构预训练任务引导模型学习更具判别性的无监督特征。在CHB-MIT和Bonn两个国际标准脑电癫痫数据集上的实验结果表明，本文模型在仅使用10%标注数据的情况下，F1-score分别达到89.2%和96.4%，性能优于现有的主流无监督方法，验证了所提方法的有效性和临床应用价值。

Abstract: In view of the fact that the existing epileptic seizure detection methods rely heavily on high-quality manual labeling by neurologists, have poor generalization ability across patients, and are difficult to capture the core clinical pain points of long-term temporal dependence of EEG signals, this paper proposes an unsupervised EEG epileptic seizure detection model based on improved Transformer. The model can learn general features from massive unlabeled EEG records without a large amount of labeled data. The multi-scale adaptive position coding is used to enhance the model’s ability to capture multi-scale temporal features of EEG. The channel-temporal dual-dimensional sparse attention mechanism is designed to effectively model inter-channel correlation and long-range temporal dependence while suppressing noise and artifact interference. The mask reconstruction pre-training task based on anomaly perception is constructed to guide the model to learn more discriminative unsupervised features. The experimental results on two international standard EEG epilepsy datasets, CHB-MIT and Bonn, show that the F1-score of the proposed model reaches 89.2% and 96.4%, respectively, with only 10 % labeled data. The performance is better than all the mainstream unsupervised methods published in recent years, which verifies the effectiveness and clinical application value of the proposed method.