Neuron Identification And Classification Based On The Extracellular Action Potential Signals

The brain function depends on networks formed by neurons which connect to each other. State-of-art microelectrode array recording techniques can collect extracelluar action potential (i.e., spike) fired by a large number of neuron cells. Spike sorting technique is the first step to analyze spike time series. Using this technique, different neurons and neuron types (pyramidal neuron or interneurons) can be identified from spikes. However, at present, most spike sorting algorithms have two major problems. First, they can't deal with Burst. Burst is one of the common modes of neuronal action potential firing in brain. Evidences have showed that bursts play important roles in increasing reliability of neuronal signal transmission, in generating synaptic plasticity and so on. In extracellular recordings, a burst appears as a train of high-frequency firing spikes with gradual changes both in amplitude and in waveforms. The non-steady change feature of burst spikes is a challenge to the analysis of neuronal firing sequences. Second, those algorithms can only classify spikes, they can't provide information about neuron type, such as pyramidal neurons or interneurons.In order to solve the first problem, this thesis presented a spike detecting and sorting algorithm for tetrode recording signals. In this method, based on the spikes picked by a threshold detecting method, potential burst segments were selected firstly according to inter-spike intervals. Then, independent component analysis (ICA) was used to separate different types of spikes in each potential burst segment. Finally, principal component analysis was used to fulfill spike clustering. The test results obtained from both experimental recording data and synthetic data showed that the algorithm was not only able to sort the burst spikes and single spikes correctly, but also able to separate overlap spikes efficiently. Further more, the algorithm can avoid the source signal number limit of ICA technique and decrease the ICA computing time. Therefore, the algorithm provides a new method for accurate burst spike detecting and sorting.In order to solve the second problem, this thesis first established train set by manual classification, then developed an algorithm based on decision tree to evaluate feature importance. The results showed that, the time interval between the positive peak before negative peak and the positive peak after negative peak of spikes together with slope of the rising edge after negative peak was the most important features to classify these two neurons. Then linear discriminant analysis(LDA) was adopted to verify this conclusion, and to compare the most important features argued by other researchers. LDA results showed that, the most important feature proposed by this thesis had the highest classification accuracy.In order to investigate the difference between spikes fired by different neuron types, this thesis established simulation models of neurons in the hippocampal CA1 region and investigated the effects of dendrite currents, cell morphology on the formation of spike waveforms. The results showed that dendrite currents have significant effects on the spikes at the locations far from cell body, but not for those near cell body. The shape differences of various pyramidal neurons resulted in large changes in the spike amplitudes. However, the shapes of these different spikes were very similar. These results serves as a criteria for pyramidal neurons and interneurons identification.In summary, this thesis develops new algorithms for burst spike sorting, determines the best feature to distinguish pyramidal neurons and interneurons, and also investigates the difference between spikes fired by different neuron types, These results provide important information for signal analysis of neuron cell assembly and spike sorting algorithm research.