Supervised Learning In Spiking Neural Networks For Epileptic EEG Recognition

With the development of brain science and the accumulation of neuroscience research results,people have a more comprehensive and in-depth understanding of the structure and operation mechanism of the biological nervous system.The spike time coding and information processing mechanism found in the biological neural system has aroused many researchers' enthusiasm for exploring spiking neural networks.Compared to the traditional artificial neural networks based on spike frequency coding scheme,the spiking neural networks,which take more biologically realistic spiking neuron models as basic units,usually have more computing power and are very suitable for the processing and analysis of brain neural signals.Researchers have constructed different spiking neural network models and proposed some learning mechanisms.These calculation models have achieved good results in pattern recognition,decision-making and prediction.Accurate recognition of epileptic EEG signals has very important practical significance for the diagnosis and treatment of epilepsy.When using traditional machine learning techniques to process and classify EEG data,the accuracy of pattern recognition depends on the extracted feature values.However,the selection of features is usually determined by the extraction method used and the personal experiences of the researcher.Sometimes such feature values cannot accurately characterize the intrinsic attributes of EEG data.How to overcome the difficulties in EEG signal recognition and further promote the development of brain science and neurological disease diagnosis technology are becoming a challenging research direction.(1)For the feed-forward spiking neural networks,the desired outputs of the hidden layer neurons can be reconstructed reversely according to the target spike sequences of the output layer neurons.Therefore,a novel supervised learning algorithm based on the desired spike sequence reconstruction is proposed.This algorithm obtains the hiddenlayer desired output by assuming that there is a reverse connection between the output layer neuron and the hidden layer neuron in the spiking neural network structure.The experimental results of spike sequence learning show that the proposed algorithm can achieve high accuracy and have good performance.Compared with the Multi-STIP learning algorithm,the proposed algorithm based on the desired spike sequence reconstruction always requires fewer iterations and has more stable learning effect.(2)Using the CHB-MIT scalp EEG database collected at the Boston Children's Hospital as the research object,this paper proposes an epileptic EEG classification model based on spiking neural network.Then it will be used to detect epileptic seizures from EEG samples in this dataset.The model mainly includes the encoding scheme of epileptic EEG data,supervised training of spiking neural network,and the processing of neural network outputs.The first try is to perform the three-class recognition tasks on a small dataset that has a relatively balanced class distribution.All samples are divided into three categories: non-seizure EEG signals,EEG signals when the seizure occurs and EEG signals during epileptic seizures.The proposed model achieves a classification accuracy of 80.8% and 73.6% on the training set and the test set,respectively.Finally,the patients' actually EEG recordings are used for epileptic seizure recognition.After processing,the accuracy of the model can reach 95.38%.