Multi-scale and Multidimensional EEG signal processing with application to seizure prediction



Multi-Scale and Multidimensional EEG Signal Processing With Application to Seizure Prediction

Researching advanced bio-signal processing techniques for diagnosis, namely those based on computational intelligence, is an important activity in neuroinformatics. Epilepsy is one of the most common neurological diseases. About 30% of patients with epilepsy cannot be treated either with medication or with surgery, and must live with the seizures that can happen anytime, anywhere, “like a bolt from the sky." Many epileptics live in constant worry that a seizure could strike at an inopportune time resulting in humiliation, social sigma, and/or injury. A device that could reliably predict a seizure could dramatically change the lives of these patients by alerting them to the impending seizure or triggering another device to abate or suppress the seizure. The electroencephalogram (EEG) is the most used bio-signal to measure the electrical brain state. The quantity of information embedded in the EEG signals is yet unknown to a large extent.


This project aims to give a contribution to knowledge extraction from the EEG in order to predict a coming epileptic seizure. It is intended to research multidimensional EEG signals to extract linear and nonlinear features in different frequency bands in order to build a synthetic image of the brain state in real time, to be processed by computational intelligence techniques in order to detect the pre-seizure times. These researchers have developed a patient-specific classification algorithm for seizure detection and prediction using multiple features of spectral power from iEEG (intracranial electroencephalogram) recordings and using SVMs (support vector machines). Their ultimate goal is to develop an implantable architecture that can reliably provide seizure detection/prediction and allow sufficient time to trigger anti-epileptic therapy. Classification using SVMs is computationally expensive work. This is due to the many parameters of SVMs that need to be tuned, employing a costly optimization process through v-fold cross-validations, which requires repeated computations v times for every SVM architecture. The new computational intelligence methods such as Adaboost and Relevance Vector Machines (RVM) are also under study. The plan is to expand the simulations to cover good part of European database of Epilepsy well known as EpilepsiaE. 

Group name: