基于PAMFC-InfoNet网络的癫痫脑电检测方法
Epileptic EEG Detection Method Based on PAMFC-InfoNet Network

为解决现有癫痫脑电(EEG)检测方法中，传统机器学习依赖手工特征、深度学习单分支模型难以兼顾局部细微特征与长程时序依赖的问题，文章提出了一种双分支并行网络架构PAMFC-InfoNet (Parallel Attention-Based Multiscale Feature Correlation-Informer Network)。该架构包含AMFC分支与Informer分支：AMFC分支通过通道注意力、多尺度卷积及特征相关模块，增强对癫痫尖波/棘波等局部特征的捕捉能力；Informer分支采用轻量化ProbSparse注意力机制，高效建模脑电信号的长程时序关联；同时引入动态加权融合策略，根据样本特征分布自适应调整双分支权重。在波恩大学与CHB-MIT两个公开的EEG数据集上验证，5折交叉验证结果显示：波恩大学数据集平均准确率达99.43%、灵敏度100%、AUROC 99.43%；CHB-MIT数据集平均准确率99.61%、灵敏度99.89%、AUROC 99.62%。消融实验证明，AMFC分支、Informer分支及动态融合策略均对模型性能有显著贡献。结果表明，PAMFC-InfoNet能全面提取脑电特征，在癫痫检测中具备高准确性与鲁棒性，为临床癫痫自动检测提供有效解决方案。
To address the issues in existing epileptic electroencephalogram (EEG) detection methods—where traditional machine learning relies on manual features, and single-branch deep learning models struggle to balance local fine-grained features and long-range temporal dependencies—this paper proposes a dual-branch parallel network architecture named PAMFC-InfoNet (Parallel Attention-based Multiscale Feature Correlation-Informer Network). This architecture comprises two branches: the AMFC branch and the Informer branch. The AMFC branch enhances the ability to capture local features such as epileptic spikes/sharp waves through channel attention, multi-scale convolution, and a feature correlation extraction Module. The Informer branch employs a lightweight ProbSparse attention mechanism to efficiently model the long-range temporal correlations of EEG signals. Additionally, a dynamic weighted fusion strategy is introduced to adaptively adjust the weights of the two branches according to the feature distribution of samples. The model was validated on two public EEG datasets: the University of Bonn dataset and the CHB-MIT dataset. Results of 5-fold cross-validation show that on the University of Bonn dataset, the average accuracy reaches 99.43%, sensitivity 100%, and AUROC 99.43%; on the CHB-MIT dataset, the average accuracy is 99.61%, sensitivity 99.89%, and AUROC 99.62%. Ablation experiments confirm that the AMFC branch, Informer branch, and dynamic fusion strategy all contribute significantly to the model performance. These results indicate that PAMFC-InfoNet can comprehensively extract EEG features, exhibits high accuracy and robustness in epilepsy detection, and provides an effective solution for clinical automatic epilepsy detection.