癫痫（Epilepsy），是一种神经系统疾病，通常是脑部病变造成的脑细胞突然异常的过度放电引发的脑功能失调的一种慢性疾病。慢性脑深部电刺激术(Deep Brain Stimulation, DBS)是一种很有前景的治疗癫痫的外科手术方法。为减低刺激剂量，研究闭环、按需启动的刺激方法，具有重要的临床意义。本文以颞叶癫痫模型大鼠作为研究对象，通过对临床多通道脑电信号的记录以及活动视频监视研究核团深部局部场电位信号(Local Field Potential, LFP)同癫痫发作之间的关系。通过使用癫痫大鼠深部信号的几种功率谱及特征信号对癫痫预测GOFA(Generic Osorio-Frei Algorithm)算法进行了改进。同时尝试使用平滑伪Wigner-Ville分布以及改进后的GOFA算法对实验得到的数据，包括大鼠的脑电信号以及实际临床癫痫病人的深部脑电信号进行处理并预测癫痫的发作。研究结果表明：一.大鼠在癫痫发作时ANT的LFP会有类似皮层的同步化放电现象出现，因此可以通过分析该核团的LFP来预测大鼠癫痫的发作，深部核团信号是闭环DBS治疗方法中理想的信号源，不需要为DBS装置增加更多的侵入式植入部件，能减少病人的手术损伤和感染。二.平滑伪Wigner-Ville分布能够直观的给出大鼠深部脑电和病人深部脑电信号能量同时间-频率的关系，从而能从图像上判断癫痫发作的时间，但是由于其计算量较大，计算时间长以及缺乏一定的评判标准，并不适合于实时嵌入式系统。三.改进后的GOFA算法通过设定合适的参数，能够很好的从大鼠深部核团信号的棘波检测中预测癫痫的发作。本文用改进后的GOFA算法对7段24小时连续采集的大鼠丘脑前核深部信号进行了临床发作预测，预测的准确率超过85%，每小时误报率在7次以下，平均提前预测时间在1秒以上。对于临床癫痫病人深部核团信号的GOFA算法还需要进一步研究和改进。最后，本文对改进的GOFA算法在嵌入式系统中的应用进行了讨论，对于算法的存储空间要求以及计算步数作出了分析和计算。为将来在嵌入式系统中实现脑电监测-分析-刺激打下了基础。
Epilepsy is a common chronic neurological disorder characterized by recurrent unprovoked seizures. Deep brain stimulation (DBS) is a promising surgical treatment for the therapy of epilepsy. Since the effect of stimulation will decrease with the time of electrical stimulation, a strategy of stimulation of closed-loop approach, which consists of signal acquirement, signal processing and seizure prediction, is necessary. In this paper, we conduct experiment on kainic acid induced epileptic rats, recording their local field potential (LFP) and their corresponding behavior. We try to build relationship of the signal and behavior. Several kind of power spectrum are utilized to improve the performance of a seizure detection algorithm—Generic Osorio-Frei Algorithm (GOFA). The smoothed pseudo Wigner-Ville distribution method and improved GOFA are utilized to process the deep brain signal, including the rats’ data and some data from real patients, in order to predict seizures onset automatically. The result of the research show that, Firstly, the LFP of ANT presents related features with seizure onset and it is an ideal signal source for the closed-loop DBS epilepsy treatment system for its decreasing the risk of blooding and infection without extra intracranial devices. Secondly, the smoothed pseudo Wigner-Ville distribution method can give the power distribution in the time-frequency domain thus it is convenient to find the onset time in the image visually. However, the huge computing task and indefinite standard of the detection makes it difficult to use in the embedded system. Finally, the improved GOFA, with the benefit of order statistic filter’s edge preservation feature, is suit for the seizure detection and clinical onset prediction from the LFP of ANT after setting some parameters. 7×24 hour data are processed, above 85% of the clinical onset are predicted more than 1 second before and the false positive rate of prediction is less than 7/h. More research is needed for the improved GOFA to predict seizure of real patients’ deep brain signal.The possibility of the improved GOFA in the embedded micro-system has been discussed in the last chapter and the complexity, besides the data structure, of this algorithm has been analyzed, building solid foundation for the future work of developing better closed-loop deep brain stimulation devices.
