アイテム詳細

要旨

For the purpose of studying the mechanisms of information processing in the brain, understanding cooperativeness
of neuronal ensembles is essential. Therefore, it is necessary to analyze the simultaneous activity of
several neighboring neurons as obtained from extracellular recordings. Such signals usually contain action
potentials from multiple cells which need to be separated by spike sorting in a reliable fashion so that spike
trains from each individual cell can be used for further analysis. Efficient spike-sorting methods should ideally
allow for real time detection and classification of spikes, for good sorting performance in the presence
of overlaps and low signal-to-noise ratio and for minimal human interaction or supervision.
Linear filtering with optimal multichannel filters seems to be a promising approach [1,2]. For every recorded
neuron - and its specific waveform - an optimal multichannel filter is constructed. This filter should optimally
have a high output energy for the correct waveform and zero output energy for noise and waveforms of other
neurons. The task of spike detection is reduced to a simple detection of high peaks in the filter output. Since
every filter detects only one putative neuron, spike detection and spike clustering are done in the same step.
Overlapping spikes can be disentangled, because both the filter operation and the superposition of spikes is
linear.
In contrast to [1,2] we use an iterative algorithm to learn a set of multichannel templates autonomously from
a 10 to 30 second piece of recorded data. From the templates a set of optimal discriminative multichannel
filters is calculated. Thereafter the algorithm can run in realtime. We evaluate the method with simulated
and experimental data. The simulator convolves multichannel waveform templates from real recordings
with Poisson spike trains and adds Gaussian noise. The algorithm is also tested on simultaneous intra- and
extracellular recordings in slices of rat visual cortex as well as on data from macaque prefrontal cortex.
We compare the results to existing spike sorting methods including thresholding combined with principal
component analysis. We conclude that our algorithm is indeed able to successfully resolve overlapping
spikes and outperforms the other methods under realistic signal to noise ratios.