癫痫是一种神经系统疾病，是由于各种原因造成脑病变导致脑细胞产生异常过度放电或/和全身强直-阵挛的慢性疾病。它影响全球1%的人口，在我国约有九百万患者，其中1/3为药物难治性癫痫，需要进行手术。但如多致痫灶或致痫灶在功能区的患者因风险大、效果不好不适合做手术。深部脑刺激被证明对此类癫痫有一定疗效，且具有可逆性，是一种极具前景的治疗手段。传统DBS采用固定参数持续刺激，刺激剂量大对脑有潜在伤害，功耗较大。研究具有脑电癫痫检测功能的闭环DBS，能够克服这些问题，具有重要价值。我们开展动物实验用商用放大器获取海仁酸造模癫痫大鼠皮层和丘脑前核脑电。找到三种能满足植入式深部脑刺激器低功耗限制的癫痫检测算法：高法、线长和半波。优化后用4只大鼠分布于20多小时中的144次癫痫发作测试，用计算速度、漏检、误检、检测延迟（DO-EO）等指标评价其效果，并形成一套三种算法联合决策的策略。结果显示：三种算法单独检测效果较好，其中线长算法除误检率最高外其他指标优于另两种算法，联合决策则优于三者单独使用；同一大鼠同一次癫痫发作ANT脑电与皮层脑电的算法检测结果没有显著差异，表明以ANT作为检测反馈源在癫痫动物模型上具有可行性。参考三种算法的思想，开发了一种新型的实时癫痫检测算法，模拟医生观察脑电的过程，提取脑电局部极值和半波的3种特征分析脑电的幅度能量、频率和节律性。用301医院和德国Freiburg大学分别包括15和78次癫痫发作共计80余小时的两组颅内脑电进行测试，正确检出率分别为100%、94%，误检率0、0.8次/小时，平均延迟1.9、4.4秒。Freiburg数据库是常用的癫痫检测数据库，此结果正确检出率和误检率达到同类研究水平，延迟低于多数同类研究。开发了一套准在线嵌入式实时癫痫检测演示系统，其中脑电波由电脑控制声卡输出，脑电采集及癫痫检测电路主要由MSP430低功耗单片机和TI的生物电模拟放大前端芯片搭建，计算效率高度优化的高法算法被移植其中。用大鼠癫痫脑电测试该系统，能够和Matlab获得同样检测效果，实际电路功耗没有进行严格控制，理论上整机电流消耗225uA，其中单片机45uA。
Epilepsy is a chronic neurological disorder characterized by unprovoked overcharge and tonic-clonic seizure caused by different kinds of brain pathological changes. Epilepsy influences about 1% of all human beings. There are 10 million epilepsy patients in China, one third of which has pharmaco-resistant epilepsy. Surgery is their only hope. However destructive surgery is unsuitable or too risky for those with multi seizure focus or seizure focus in functional brain area. DBS has been proved to have therapeutic effect for pharmaco-resistant epilepsy, and it is reversible, showing a promising surgical therapy. Traditional DBS delivers constant stimulation with preset mode. Since the therapeutic effect will decrease and adverse effect will increase with the time of stimulation, constant stimulation consumes unnecessary power. The research of closed-loop DBS that detects electrographic seizure onset and delivers stimulation accordingly is a valuable research. In this dissertation, we conducted experiment on kainic acid induced epileptic Wistar rats, recording their Electroencephalography both on the cortex and anterior neucleus of thalamus. Three seizure detection algorithms: linelength, halfwave, generic Osorio-Frei algorithm（GOFA）, are implemented and optimized to meet the need of real-time low power embedded system that has very limited storage and calculation speed. Calculation speed, false positive (FP), false negative (FN), detection delay, are criterion to evaluate detection performance of the algorithms. Then the three algorithms are combined to form a united strategy which has better detection performance. 144 seizures from more than 100 hours of electrograph recorded from 4 rats are analyzed. The results showed: 1. each of the three algorithms alone can detect most seizures, and linelength is obvious better than other two algorithms in most cases except for higher FP rate. And the united strategy has better performance than any of the three algorithms. 2. Detection of ANT electrograph and cortex showed no difference, proving that using ANT, the common used DBS stimulation target for refractory epilepsy, as the feedback source of closed-loop DBS is feasible.Reference with the three classic algorithms, a novel real-time seizure detection algorithm is developed, which simulate how doctors inspect epilepsy Electroencephalography. It used three features based on local extremum and half wave to measure the amplitude energy, frequency and rhythmicity of signals. 15 seizures recorded from 301 hospital and 78 seizures from the Freiburg EEG database, more than 80 hours in total, are used to validate the novel algorithm. The detection rate is 100% and 94%, FP rate is zero and 0.8 FP/hour, average delay is 1.9 and 4.4 second. Validated with common used Freiburg epilepsy database, the novel algorithm proves an average FN and FP rate and a shorter delay comparing with similar researches in the world,A real-time seizure detection demonstration system consisted of two parts was built. An analog signal generator restreamed out the recorded Electroencephalography through the audio board of a laptop. And a seizure detection circuit built with MSP430 microcontroller and a bio-potential amplifier frontend IC acquisited and analyzed the data with the highly calculation-optimized version of GOFA and linelength. The embedded system detected seizures with the same performance as Matlab algorithm did. It has a theoretical current consumption of 225uA, and the microcontroller consumes 45uA.
