Research On Key Technologies Of Electroencephalogram-based Seizure Detection And Detection-Suppression Epilepsy Therapy System

Epilepsy is a common and serious brain disorder, and epileptic seizures usually cause convulsions, loss of consciousness, and muscle spasms, that overshadow the patients’lives. There are 50 million people suffered from epilepsy worldwide, while only eighty percent of them could be treated by traditional methods such as pharmacotherapy and surgical resection. There are still about 10 million patients remain untreated and thus has brought great mental and economic burden to the family and society. Researches and developments of new options for effective epilepsy treatment and control are extremely important.Thanks to the recent advances in neuroscience and information technology, new treatments for epilepsy have developed rapidly. One of the most promising options is responsive neurostimulation. Unlike traditional open-loop neurostimulations, responsive system works in a "detection-suppression" manner, which first detects seizure onset from electroencephalogram via signal processing technology, and then delivers neurostimulation if and only if seizure occurs to abort epileptic seizures. Responsive system could highly reduce the amount of neurostimulation and significantly attenuate the risk of side effects and tissue damage; therefore it is promising to bring a new breakthrough in epilepsy treatment. Currently, responsive neurostimulation for seizure control has been a hot research topic so that research institutes and companies worldwide have put much effort in it, and products have been developed for clinical applications.This thesis studies key technologies of closed-loop "detection-suppression" responsive epilepsy therapy system. Focused on the key problems and difficulties in existing technologies, this thesis studies effective feature extraction algorithm on seizure signals, denoising method to handle the heavy noise in electroencephalogram data, and sequential modeling of seizure state transfer dynamics, to establish reliable and robust seizure detection method. Based on this, online "detection-responsive neurostimulation" software and hardware system is built and tested on animal epileptic models. The main contributions of this thesis are listed as follows:1) Focused on the problem that, the seizure onset patterns between individuals are highly diverse, this thesis proposes a new feature extraction method to automatically learn effective features from the signal itself. Based on traditional deep neural networks, this thesis develops a maximum correntropy criterion (MCC)-based neural network that could measure distance between samples according to sample distribution. Therefore, it could improve deep neural network’s feature learning ability under noise. Besides, optimal features could be learned automatically from seizure signals of each individual. Experimental results show that, our method could improve seizure detection accuracy by 10%.2) Focused on the problem that, artifacts caused by electromyography, electrooculography, and electrode noise lead to high false alarm rate, this thesis proposes a state-space model with Cauchy observation noise to estimate the dynamics of seizure state transfer. Since the noise in electroencephalogram shows non-Gaussian property, heavy-tailed Cauchy distribution is employed to model the observation noise with outliers. Besides, considering the brain state transfer process, sequential model is built to use information in preceding time windows for more accurate state estimation. Experimental results show that, the state-space model with Cauchy observation noise is robust to outliers and significantly reduces false alarms by 91%.3) Based on the stable and robust seizure detection algorithm, this thesis studies responsive neurostimulation system for seizure control. This system integrates seizure detection methods, signal recording and neurostimulation hardware and online control software system. The system is tested on animal epileptic model and the seizure control performance of responsive neurostimulation is evaluated. Controlled experiment on rat epileptic model shows that, the system could reduce seizure onset duration by 32%. The results strongly demonstrate that, responsive neurostimulation could effectively suppress seizure onsets.Overall, this thesis focuses on the key problems in electroencephalogram-based seizure detection and "detection-suppression" responsive epilepsy control system, puts forward effective solutions, and finally accomplishes reliable seizure detection and responsive neurostimulation system for seizure control. The feature extraction method for diverse patterns and sequential modeling method under heavy noise proposed in the thesis are innovative; and some algorithms not only bring breakthroughs in seizure detection, but also have great potential on other complex data, which worth to be further studied and applied.