基于脑电信号的癫痫发作预测特征及识别

解码癫痫发作前脑电信号的神经元集群异常痫样放电活动，对癫痫发作进行有效预测并实施病前干预，可显著减少疾病病损，是癫痫防治的研究热点之一。基于脑电信号的癫痫发作预测研究关键在于发作间期和前期的异常状态识别，研究上述两状态间的神经动力学特征差异对明确癫痫发病机制、选取高分辨特征，进而有效识别该渐进性疾病所处的发作阶段具有重要价值。目前，研究者已对当前主流特征提取及模式识别方法进行了充分的调研梳理，但忽视了神经动态特征变化对于癫痫发作预测的重要意义。基于此，本文归纳总结了5类典型的发作预测特征分析方法及其优缺点，重点剖析了发作间期至前期神经生理特征的动态变化及其动力学特性，类比分析了当前该领域主流的机器学习和深度学习特征识别方法，以期为进一步建立精准、高效的癫痫发作预测技术提供新思路。

Features and Recognition of Epileptic Seizure Prediction Based on Electroencephalogram Signals

The method that decoding the electroencephalogram (EEG) signal from abnormal epileptiform discharge activity of neuron clusters in the preictal states can significantly decrease the lesions by predicting epileptic seizures effectively and implementing interventions in patients before the onset of seizures, and thus is considered the hotspot of the current research in epilepsy prevention and treatment. The key to epileptic seizure prediction based on EEG signals lies in the identification of abnormal states in the inter-ictal and pre-ictal states. Studying the differences in neurodynamic characteristics between the above two states contributes greatly to clarifying the pathogenesis of epilepsy, and is of great value for the prevention and prognosis of patients. By extracting the high-resolution features from the neurodynamic characteristics, the onset of this progressive disease can be effectively identified. Despite the prevailing feature extraction and pattern recognition methods have been investigated sufficiently, it appears that the existed research ignores the importance of identifying changes in neurodynamic characteristics for seizure prediction. Pointing at the deficiency aforementioned, this paper summarizes five typical analysis methods of seizure prediction in neurodynamics, including time domain, frequency domain, time-frequency domain, nonlinear dynamics and global synchronization analysis, as well as their specific characteristics. Since multiple properties of EEG before epileptic seizures, such as amplitude, phase, transient frequency, band power, brain area energy, system and dimensional complexity, and global synchronization level, will change correspondingly with the abnormal activity of brain neuron clusters, the dynamic changes of neurophysiological features are analyzed with emphasis to research neurodynamic properties from inter-ictal to pre-ictal. In addition, the prevailing machine learning and deep learning methods of feature recognition are compared. Facing the current challenges, this study finally synthesizes the latest findings in this field, aiming at providing new insights for establishing accurate and efficient technology for epileptic seizure prediction.