Analysis of EEG Epileptic Signals with Rough Sets and Support Vector Machines

摘要

Epilepsy is a common chronic neurological disorder that impacts over 1% of the population. Animal models are used to better understand epilepsy, particularly the mechanisms and the basis for better antiepileptic therapies. For animal studies, the ability to identify accurately seizures in electroencephalographic (EEG) recordings is critical, and the use of computational tools is likely to play an important role. Electrical recording electrodes were implanted in rats before kainate-induced status epilepticus (one in each hippocampus and one on the surface of the cortex), and EEG data were collected with radio-telemetry. Several data mining methods, such as wavelets, FFTs, and neural networks, were used to develop algorithms for detecting seizures. Rough sets, which were used as an additional feature selection technique in addition to the Daubechies wavelets and the FFTs, were also used in the detection algorithm. Compared with the seizure-at-once method by using the RBF neural network classifier used earlier on the same data [12], the new method achieved higher recognition rates (i.e., 91%). Furthermore, when the entire dataset was used, as compared to only 50% used earlier, preprocessing using wavelets, Principal Component Analysis, and rough sets in concert with Support Vector Machines resulted in accuracy of 94% in identifying epileptic seizures.