Abstract
When a neuron in the brain is activated, its membrane potential sharply rises up and immediately drops down. This phenomenon, which is called as a neural spike, is basis of many analyses about the brain in research fields such as neuroscience and neural engineering. Neural spikes are usually recorded by extracellularly voltage recordings and spikes from several neurons are recorded from an electrode. Therefore it has been required to develop the method to detect spikes and to determine which neuron generated each spike. Many previous methods have been proposed, however, they are not robust when more than two spikes are generated closely and their waveforms are overlapped. In this paper, we proposed a method to detect and sort spikes robustly under many overlaps. Extracellularly recorded voltage signals were modeled with hidden Markov model that can generate overlapped spikes. Hidden variables (corresponding to the existence of neural spikes at each time) and model parameters (the shape of spikes from each neuron and the standard deviation of noise) were estimated by alpha-beta algorithm and expectation-maximization algorithm. The number of templates was determined by maximizing BIC. The method was assessed using a simulated signal which contained many overlaps of spikes and additive white Gaussian noise. As the result, it was showed that our method could appropriately detect and sort spikes under many overlaps.