Pattern recognition of epilepsy using parallel probabilistic neural network

Chen Gong, Xingchen Zhou, Yunyun Niu

Abstract: Accurate and rapid pattern recognition of epilepsy from intracranial electroencephalogram (iEEG) recordings is important for medical diagnostics. In this paper, three algorithms based on discrete wavelet transform (DWT) analysis and parallel probabilistic neural network, SA-PNN, SA-PPNN, and LSA-PPNN, are presented to identify iEEG recordings and detect epileptic seizures. Simulated annealing (SA) and local simulated annealing (LSA) are utilized to optimize network parameters of probabilistic neural network classifier, respectively. The combinations of different features are utilized as the input vectors of classifiers to complete classification tasks. Experiments are conducted to deal with five different classification tasks. Compared with non-parallel probabilistic neural network algorithm (SA-PNN), the running time of parallel probabilistic neural network algorithm (SA-PPNN) is shortened by 2.18 times. Compared with SA-PPNN, the average operating time of LSA-PPNN is reduced by 9.97 times. The reason is that LSA-PPNN trains and optimizes parameters with local data firstly and then brings the parameters into the global training data sets to train the network for a test. As the amount of data increases, the superiority over LSA-PPNN is getting more distinct. Our methods are also compared with other existing relative research. Experimental results prove that our methods are much more competitive. In particular, for the classification task C-D, the classification accuracy of our method reaches 83.3%, which is much higher than previous results.