癫痫是常见的脑部系统神经疾病，严重的患者生活不能自理。大部分癫痫患者可通过药物或手术进行治疗，对于少部分具有耐药性癫痫同时不能手术的患者可以通过脑深部电刺激的方法进行治疗。脑深部电刺激疗法可分为开环和闭环两种模式，闭环模式需要实时采集患者的脑电信号并根据可靠的癫痫检测算法根据脑电信号判断患者状态，检测到患者正在发作及时给予电刺激。 基于以上背景，本文首先基于苏黎世联邦理工学院和伯尔尼大学睡眠觉醒中心公开的癫痫患者颅内脑电数据集对患者进行分析，对比癫痫发作和非发作状态下患者的脑电信号特征差异以及不同分类算法的差异，同时考虑到本文所使用的脑起搏器的硬件条件，确定了以1至60赫兹的功率谱密度和Katz分形维数、Hjorth参数作为癫痫检测算法的特征，以线性判别分析和支持向量机为分类算法的癫痫检测算法，两种算法在数据集中的平均准确率分别为0.94和0.93。其次，本文在构建多种检测算法的基础上，开展闭环脑深部电刺激治疗耐药性癫痫患者的临床试验。在临床试验中对构建的癫痫检测算法的临床实用性进行了验证，两例患者使用的检测算法的准确率分别为0.96和0.86。同时，本文对检测算法在临床试验中所遇到的问题进行解决，增加了高通滤波的环节来避免刺激伪迹触发刺激的问题。本文还对算法的假阳刺激较多问题进行探究，发现大部分假阳刺激是由IEDs导致，构建了一个以随机森林模型为分类算法的二次判别算法，两例患者假阳性分别降低约80%和50%。本文也对通用的癫痫发作检测算法进行了探索，以数据集中的数据做训练，以我们在临床试验中采集到患者的局部场电位数据做测试，两例临床患者应用通用线性判别分析算法的准确率分别为0.95、0.86，应用通用支持向量机算法的准确率分别为0.92、0.80。 综上，本文的工作围绕闭环脑深部电刺激治疗药物难治性癫痫的算法展开，构建了两种能够用于临床的癫痫发作检测算法，依此开展了相关临床试验，并进行了降低算法假阳性的研究和通用癫痫发作检测算法的探索。本文的研究为相关临床试验提供了算法前提，为闭环脑深部电刺激治疗耐药性癫痫患者的大规模应用贡献了思路。关键词：癫痫；闭环脑深部电刺激；脑电信号；癫痫检测算法

Epilepsy is a common neurological disorder in the brain system, and severe patients cannot take care of themselves. Most epilepsy patients can be treated through medication or surgery, while a small number of patients with drug-resistant epilepsy who cannot undergo surgery can be treated through deep brain electrical stimulation. Deep brain electrical stimulation therapy can be divided into two modes: open loop and closed loop. The closed loop mode requires real-time collection of the patient's EEG signals and the use of reliable epilepsy detection algorithms to determine the patient's status based on the EEG signals. If the patient is experiencing an attack, timely electrical stimulation should be given. Based on the above background, this article first analyzes the intracranial EEG datasets of epilepsy patients publicly available at the Federal Institute of Technology Zurich and the Sleep Awakening Center of the University of Bern. The differences in EEG signal characteristics between patients with seizures and those without seizures, as well as the differences in different classification algorithms, are compared. At the same time, considering the hardware conditions of the brain pacemaker used in this article, a power spectral density of 1 to 60 Hz, Katz fractal dimension, and Hjordh parameter are determined as the features of the epilepsy detection algorithm. Linear discriminant analysis and support vector machine are used as the classification algorithms for epilepsy detection. The average accuracy of the two algorithms in the dataset is 0.94 and 0.93, respectively. Secondly, based on the construction of multiple detection algorithms, this article conducts clinical trials of closed-loop deep brain stimulation therapy for drug-resistant epilepsy patients. The clinical practicality of the constructed epilepsy detection algorithm was validated in clinical trials, and the accuracy of the detection algorithms used in the two cases was 0.96 and 0.86, respectively. At the same time, this article solves the problems encountered by detection algorithms in clinical trials by adding a high pass filtering step to avoid the problem of stimulus artifacts triggering stimuli. This article also explores the problem of many false positive stimuli in the algorithm, and finds that most of the false positive stimuli are caused by epileptic discharges during the interictal period. A secondary discrimination algorithm based on the random forest model classification algorithm was constructed, and the false positives in the two cases were reduced by about 80% and 50%, respectively. This article also explores universal seizure detection algorithms, using data from datasets for training and testing local field potential data collected from clinical trials. The accuracy rates of applying the universal linear discriminant analysis algorithm to two clinical cases were 0.95 and 0.86, respectively, and the accuracy rates of applying the universal support vector machine algorithm were 0.92 and 0.80, respectively.In summary, the work of this article revolves around the algorithm of closed-loop deep brain electrical stimulation for the treatment of drug-resistant epilepsy. Two seizure detection algorithms that can be used in clinical practice were constructed, and relevant clinical trials were conducted based on this. Research on reducing algorithm false positives and exploring universal seizure detection algorithms were also conducted. This study provides algorithmic prerequisites for relevant clinical trials and contributes ideas for the large-scale application of closed-loop deep brain electrical stimulation in the treatment of drug-resistant epilepsy patients.Keywords: Epilepsy; Closed-loop deep brain stimulation; Electroencephalogram signals; Seizure detection algorithm