Wang Yueming, Qi Yu, Wang Yiwen, Lei Zhen, Zheng Xiaoxiang, Pan Gang
Qiushi Academy for Advanced Studies, Zhejiang University, Hangzhou 310027, People's Republic of China. Department of Computer Science, Zhejiang University, Hangzhou 310027, People's Republic of China.
J Neural Eng. 2016 Oct;13(5):056009. doi: 10.1088/1741-2560/13/5/056009. Epub 2016 Aug 22.
There is serious noise in EEG caused by eye blink and muscle activities. The noise exhibits similar morphologies to epileptic seizure signals, leading to relatively high false alarms in most existing seizure detection methods. The objective in this paper is to develop an effective noise suppression method in seizure detection and explore the reason why it works.
Based on a state-space model containing a non-linear observation function and multiple features as the observations, this paper delves deeply into the effect of the α-stable distribution in the noise suppression for seizure detection from scalp EEG. Compared with the Gaussian distribution, the α-stable distribution is asymmetric and has relatively heavy tails. These properties make it more powerful in modeling impulsive noise in EEG, which usually can not be handled by the Gaussian distribution. Specially, we give a detailed analysis in the state estimation process to show the reason why the α-stable distribution can suppress the impulsive noise.
To justify each component in our model, we compare our method with 4 different models with different settings on a collected 331-hour epileptic EEG data. To show the superiority of our method, we compare it with the existing approaches on both our 331-hour data and 892-hour public data. The results demonstrate that our method is most effective in both the detection rate and the false alarm.
This is the first attempt to incorporate the α-stable distribution to a state-space model for noise suppression in seizure detection and achieves the state-of-the-art performance.
脑电图（EEG）中由眨眼和肌肉活动引起的噪声严重。该噪声的形态与癫痫发作信号相似，导致大多数现有癫痫发作检测方法的误报率相对较高。本文的目的是开发一种有效的癫痫发作检测中的噪声抑制方法，并探究其工作原理。
基于一个包含非线性观测函数和多个特征作为观测值的状态空间模型，本文深入研究了α稳定分布在从头皮脑电图中检测癫痫发作的噪声抑制中的作用。与高斯分布相比，α稳定分布是不对称的，并且具有相对较重的尾部。这些特性使其在对脑电图中的脉冲噪声进行建模时更具优势，而高斯分布通常无法处理这种噪声。特别地，我们在状态估计过程中进行了详细分析，以说明α稳定分布能够抑制脉冲噪声的原因。
为了验证我们模型中的各个组件，我们在收集的331小时癫痫脑电图数据上，将我们的方法与4种不同设置的不同模型进行了比较。为了展示我们方法的优越性，我们在我们的331小时数据和892小时公共数据上，将其与现有方法进行了比较。结果表明，我们的方法在检测率和误报率方面都是最有效的。
这是首次尝试将α稳定分布纳入状态空间模型以用于癫痫发作检测中的噪声抑制，并取得了当前最优的性能。
