Approximately 1% of the world’s population is diagnosed with epilepsy. Epilepsy is a nervous system disorder that is characterized by episodic loss of attention or sleepiness and in more sever cases by loss of consciousness. Epilepsy is caused by abnormal electrical activity in the brain, more specifically by a synchronized discharge of neurons in the gray matter of the brain. Epilepsy impacts live of millions of individuals every day both physically and psychologically. Injuries, physical anxiety, impairment of productivity, job loss, and social isolations are some of the distressing effects of epilepsy.
Prediction of epileptic seizures would improve the quality of life of millions of people who are suffering from unexpected seizures. Different features of the electroencephalogram (EEG) can be employed for prediction of seizure. Neurologists and technicians who read EEG signals need to spend a considerable amount of time to read and interpret the EEG signals for each patient. Another disadvantage of this visual interpretation of the EEG signals is that the diagnosis may vary among the physicians, not to mention the possibility of visual failure to spot the precursory signals of impending epileptic seizures. Therefore, an automated prediction of seizures will not only assist the physicians in their diagnosis, but will also improve the quality of patient care and would eventually result in the development of implantable intelligent devices that reliably trigger warnings for patients to respond in time before a sever epileptic seizure manifests. A myoclonic seizure is depicted in figure 11. The abnormal electrical activity is manifest in the central area of the EEG.
Figure 1. EEG of a seizure activity
We propose the use of artificial intelligent system both for prediction and detection of seizures. Our introductory approach is to employ an artificial neural network and will be involved with four major steps:
One of the most challenging problems in training neural networks is determining the appropriate features that contribute the most to the training of the network. There are two different schemes for reducing the number of features for training the network: feature selection and feature extraction. There are also multiple methods for feature extraction. We propose the use of principal component analysis as well as auto-encoder for feature extraction.
Determining the best network topology requires trial and error to an extent. Backpropagation is the most common type of neural network. However, employment of RBF neural networks have recently been reported in medical applications. Therefore, we would like to compare the performance of both types of networks on prediction and detection of epileptic seizures.
In order to evaluate the performance of the networks fairly, we will consider sensitivity, specificity and accuracy simultaneously. Sensitivity is the percentage of correct prediction of patients who did not have epileptic seizures. Specificity is percentage of correct prediction of patients who indeed had epileptic seizures, and finally accuracy is the total percentage of the correct predictions.
Finally, we will use the best performers to achieve an intelligent system that could predict the epileptic seizures and improve the automation of EEG diagnosis.
We would like to thank Albert-Ludwigs University for sharing their epilepsy data set with us. (http://www.fdm.uni-freiburg.de/groups/timeseries/epi/)