Death to the spike sorters!

First, good paper here: Spike train decoding without spike sorting by Ventura, V. (also available free on her site here)

Essentially, she looks at using a simple voltage thresholding of the continuous data stream to infer neuron identities using a very complicated process of turning curve fitting and statistics. One of the more interesting things, I thought, was the way in which noise drops out of the equation, since "noise" is all allocated to a "noise neuron". I am just looking at the paper for the first time, but I swear I remember her stopping by my SfN poster last year (we did a similar spike-sorting-less type of approach as on of our signal sources for comparison and it did perform quite well), and exclaiming something like, "I knew I wasn't the only one that thought this should work!"

If you're wondering what I'm talking about, the idea is this: You get a signal from an electrode channel that is the combination of many different types of "field potentials" and "unit activity". Unit activity is easy enough - those spike looking things that people use to discretely decode something or other. The tick marks in a raster, the counts in a histogram. "Field potentials" are more complicated, and represent the total electrical activity that the particular electrode is being subjected to and able to convey, based on its materials and geometry (thing like impedance, etc.). This represents things like the slow fluctuations in membrane potential of nearly cells as ions move through channels, overlaid activity from cells within the volume of brain matter, depending on location and electrode sensitivity it can include EMG and EOG signals, etc. Essentially everything that has an electrical charge big enough and near enough to be detected.

When you use unit activity to decode some type of behavior, you have to decide how many neurons you are recording from, or so the thinking has been. Excitatory signals spike, as do inhibitory, so people generally look at each different waveform as a separate cell's activity. In order to separate the waveforms, you "spike sort" (or "cluster cut" if you're an old DataWave user). This involves various ways of saying "this waveform is different from that waveform". It is much more difficult than you think, and is a major pain in the ass for a number of reasons. It also adds a ton of processing overhead, making it particularly annoying for BCIs, since you would like to do all the processing on some small chip that is implanted (sending just spike events is much easier than streaming continuous data for many reasons, power and heat being two).

For my last SfN poster, we looked at various ways of filtering the continuous data to look at various bands thought to represent different classes of field potentials. In our case, it was Multi Unit Activity (MUA), which is a high frequency band, thought to encode the output of a small cortical area. We compared the ability to classify which target a person with an implanted electrode array (Utah/BrainGate 100 lead grid) was attempting to move a cursor to, which they had neural control over using unit activity. We compared the performance of the MUA signal (using two different techniques) to the traditional spike sorting, unit based way using both "poor" and "good" sorting techniques.

That's the long way of getting to this. We then also compared unit and MUA activity to what Stark and Abeles called MSU - Multiple Single Unit activity. Actually, we used two definitions of MSU. Confused yet? Let' s get past this part and it will make more sense. MSU was defined in one case as "still use the regular unit activity technique, but every waveform on an electrode belongs to the same unit. No spike sorting. Then we also tried just thresholding the continuous data, and decoding using that. So that signal is the "if the signal is over X mV, that means something is happening and therefore there is an 'event' there" technique.

Ventura's paper is a huge formalization of using the second definition of MSU, and takes it many steps further. The biggest difference being the way that she creates and separates the individual 'events', creating, in essence, surrogate 'cells'. I say 'cells' because there is no indication that they are biological, individual cells, but an activity producing lump of brain goo that fits a complicated statistical model. Actually, I don't know if it is complicated. I read through 1/2 the paper and started to glaze over (skimmed the whole thign a few times), and this is the type of paper you need some time to digest. So yes, check it out. If anything drastically different than what I said come out of it, I'll post again.