基于跨层全连接神经网络的癫痫发作期识别

Epileptic EEG identification with cross layer fully connected neural network

在缺乏足够先验知识下,自适应癫痫发作期识别异常困难。提出一种新的度量通道之间的同步特征计算方法(聚类划分互信息),以相关矩阵方式组织单窗口内全局同步特征模式,进而设计一种跨层全连接神经网络分类器,对非平稳同步特征模式实现自适应分类。实验表明该方法可获得[98.19%±0.24%]精确度,[98.27%±0.51%]敏感度和[98.11%±0.36%]特异度,超过了现有大部分方法的分类性能。另外,该方法无须去噪和去伪迹等预处理过程;而且其仅需设置一个超参数(时间窗),避免了过多的潜在错误参数设置而导致的分类性能的降低。

Under the circumstance of insufficient prior knowledge,it becomes even more important to adaptively classify the synchronization dynamics to accurately characterize the intrinsic nature of seizure activities represented by the EEG.This paper first measured the global synchronization by calculating clustering partition mutual information(MI) of all EEG data channels.Then it designed a cross layer fully connected net to adaptively characterize the synchronization dynamics captured correlation matrices and automatically identify the seizure states of the EEG.It also performed experiments over the CHB-MIT scalp EEG dataset to evaluate the proposed approach.It identified seizure states with an accuracy,sensitivity and specificity of [98.19%± 0.24%],[98.27%±0.51%],and[98.11%±0.36%],respectively.The resulted performance was superior to those of most existing methods over the same dataset.The approach alleviated the need for strictly denoising and artifact removing based on the EEG prior knowledge that is mandatory for existing methods.Only one hyper-parameter need be set manually to avoid getting worse performance because of complex parameter setting.