In this document four different techniques such as Renyi Entropy, Shannon Entropy, Approximate Entropy, and Kurtosis are applied to EEG signals frame by frame for input to an artificial neural network for automatic seizure detection. The performance of each one is documented in relation to the performance achieved using only the average (mean) and standard deviation of each frame.

The Renyi and Shannon entropies were obtained in the frequency domain where all others are processed in the time domain. Performance parameters used are sensitivity, specificity, and overall accuracy. The artificial neural network is a Nonlinear Autoregressive with Exogenous inputs (NARX) neural network and the sample data are two sets of one hundred EEG signals of twenty three second time intervals. As far as we know this is the first time that the NARX network has been utilized for automatic epileptic seizure detection. Also it is the first time that Renyi and Shannon entropies have been used as features taken from the signal in the frequency domain for input to the neural network.

This study will show calculation times for each technique on a common personal computer and the complexity of each method using the big-O notation. We will show that taking the standard deviation as a feature yields the best results in detection of epileptic seizures using the NARX neural network. Furthermore the complexity of Approximate Entropy has been re-examined and shown that it is the least feasible method for feature extraction.