Research On Detection And Classification Method For Overlapped Spikes

The nervous system deliverys information by the diffusion of spikes among the neurons,so the spike is an important proof to the study of nervous system’s collaborationmechanism.The data collected by multi-electrode array(MEA) may contain a large number ofnoises and the spikes in one array of data may come from different neurons. So the spikesshould be extracted from the signal trains and be classified to their corresponding neurons.This process of the collected data is called spike sorting. When two or more neurons firespikes simultaneously, superposition waveforms will be recorded unavoidable. Overlappedspikes increase the difficult to spike sorting. The decomposition of overlapped spikes hasbecome one emphasis and difficult problem of neurobiology analysis.The thesis focuses on key problems and methods of spike sorting, especially foroverlapped spikes. Overlapped spike detection method, feature representation method andspike classification strategy are proposed in the thesis.To solve the problem of ignoring overlapped spikes in a window or between adjacentwindows by window detection method, a new spike detection method is proposed in the thesis.The method adds threshold detection and nonlinear energy operator based on windowdetection method, and solves the repeated detection problem by estimating slopes.Experiments show that the method is good for any occasion whatever the low signal-to-noiseratio or baseline wander. Especially for the overlapped spikes detection, it has much lowerfalse-negative-rate than other traditional detection methods.Support vector machine (SVM) is introduced to classify the spikes by using of itsadvantage in solving nonlinear and small sample problems and the influcing factors arediscussed, such as the kernel function and noise level. Experiments show that SVM iseffective and stable to overlapped spikes, but the accuracy will be reduced when thewaveforms are highly similar or the distribution of different samples cross each other heavily.In order to solve the problem that SVM classify accuracy is reduced with highly similarspikes, a new feature representation method is proposed in the thesis. The method divide aspike into waveforms by windows, and make the slope of every window be the new attributevalue of optimized sample. Experiments show that the method could lower the similarity ofdifferent type spikes, and thereby benefit the classification accuracy.SVM will be over trained if the distributions of spikes with different types cross eachother heavily. To solve the problem, a new strategy is proposed. First the training set would beoptimized by k-means clustering algorithm, and classify the test samples by Euclideandistance and SVM respectively. Experiments show that the accuracy will be increased and thetime consuming will be reduced compared to using SVM only.