[mattkrause]

If I may add some details....

The most popular implant is probably Blackrock Microsystems' "Utah Array", which has 96 electrodes arranged in a 10x10 grid (minus the corners). It looks like this: http://aerobe.com/wp-content/uploads/2016/11/utah-3.jpg For scale, the entire electrode grid is about 4mm on a side and the electrodes are between 0.5 and 1.5mm long (depending on the model).

There are a few other models (and similar stuff from other companies), but I'd be surprised if anything with thousands of contacts is in regular in vivo use. There are some in vitro (i.e., cells or tissue slices in a dish) systems with more contacts, but the signal quality isn't nearly as good.

We can read out the activity of single neurons--people have been doing it for single electrodes since the 1960s. It's slightly easier with a single (movable) electrode since you can creep up on the cell until its action potentials are fairly large and well-isolated from the background noise (here, large means about ±150 µV). You can't move the array or its individual electrodes, so you're stuck hoping that the individual shanks end up in good positions. Then, data is recorded at a fairly high sampling rate (say, 30 kHz) and the "spikes" are clustered based on their shapes to get individual neurons' responses.

The ADCs aren't directly at the contacts, but you want the amplifiers and ADCs as close to electrode as possible to avoid all sorts of weird EMI from the mains, other equipment, etc. Getting the grounding and shielding right is a bit of a black art and eats up tons of researcher time. (You'd think "throw it all in a Faraday cage" would work, but...it doesn't).

What else do you want to know? :-)

[hprotagonist]

I've done ephys in mice and gerbils. Spike sorting is nontrivial, and the effects on local tissue from jamming long shank electrodes into cortex are nothing i'd like done to me.

Unfortunately, less invasive recording techniques will never give you the ability to record from single units.

edit: pulled your google scholar and boy am I preaching to the choir...

[mattkrause]

And none of my array stuff is published yet (grrr!)

I did single-electrode experiments for my PhD and those definitely mess up the brain after a while. The Utah array stuff strikes me as "less bad" in there's only one* big insult to the brain, but it is a pretty bad one: the arrays are inserted with a pneumatic "gun".

I think you're right that non-invasive techniques will never give us single unit data, though I hope we can get some longer-lasting implantable electrodes soon.

[mdahlstrand]

I'm interested in the subject and appreciated your post, thanks. A couple of questions:

How does the brain adapt to having the electrodes in there, how long before the probes are accepted as being "part of" the brain?

Do you envision we need lots more sensors than in your example above, or is said number enough for precision input (say, text/words, or navigation in a 3 dimensional position & rotation plus a temporal dimension interface)? I guess the brain would work around the rough edges (or lack of sensor resolution) just like it already does with keyboards, mice, bodies, and language.

[mattkrause]

It depends!

For most electrodes, the brain doesn't really incorporate the implant. When a single electrode is inserted, you can start recording as soon as the contacts are inside the brain (in practice, you wait a few minutes since the brain is slightly elastic and stuff moves around). In humans, this is how deep brain stimulation is done--the surgeons use the neural activity to figure out when the electrode is in the right place. For larger implants like the Utah arrays, the insertion is a bit more traumatic. Allegedly, you get a pretty good signal right away, then inflammation makes it degrade for a while, and after ~12 hrs, the signal returns. However, the animal/patient is usually recovering during this time, so it's moot.

These electrodes are usually silicone and metal, usually tungsten, platinum/platinum iridium, or iridium oxide, so the brain doesn't really "accept" them. In fact, it tries to encapsulate and reject them, which limits the lifespan of the electrodes. In my experience, a two week old array might have nice, well-isolated neurons on more than half of the channels; after two years, you'd be lucky to get single units on more than a handful of the 96 channels.

However, there's a lot of interest in developing coatings that inhibit this immune response or actually encourage neurons to grow into the array. There's a lot of promising research on this, but nothing (as far as I know) that's commercially available.

As for the number of sensors...it also depends. You can do a lot with a 96 channel array implanted in the right spot, including spelling (https://elifesciences.org/content/6/e18554#bib27) and control of a robotic arm (http://www.jhuapl.edu/prosthetics/scientists/neural.asp) though neither of these is anywhere near "native" performance yet. More electrodes might help, but there's also probably some low-hanging fruit in figuring out the right control paradigms, decoding algorithms, and even where the electrodes are placed.

For research though, more and better arrays would be great. Many brain areas have a spatial structure. In visual areas, for example, cells representing neighboring spots in the visual world[0] are also near each other. Motor and sensory cortex also have a foot-to-face progression. Bigger arrays might let us sample from a more diverse population of neurons at the same time, which could be scientifically interesting and useful for BCI. Denser arrays might also allow for better recordings from single neurons. If you have a sufficiently dense array, you can record the activity of one neuron from multiple sites--this lets you isolate its responses better (this trick is commonly used with bundles of four wires, called tetrodes).

I would also love to get my hands on arrays made from multiple materials. Platinum is great for recording the activity of single units, but lousy for stimulation; its low charge injection capacity means that high stimulation currents damage the electrodes and/or nearby cells. Iridium oxide has a much higher charge injection capacity, but lower impedance and thus, fewer well-isolated cells. A "checkerboard" pattern of Pt and IrOx electrodes would be awesome, but is apparently difficult to make (no chance you run a fab, is there?)

The flips side of all of this is that amplifiers and ADCs are expensive (though much cheaper than they used to be), and adding channels rapidly increases the data files' size too. My experiments generate about 1 GB/minute, and we record 5-8 hours a day, 6-7 days a week.

What else? :-)

[0] The organization is really "retinotopic", meaning that cells receiving inputs from the adjacent parts of the retina are near each other in the brain.

[neurothrow]

Assuming that the 1 GB/min figure is because you're saving broadband (back of the envelope math assuming you have two Utah arrays with 32 bits/sample at 30 kHz), you can get enormous space savings with compression. Field potentials are highly correlated across channels, as you no doubt know. Depending on your amplifier you may be able to pack everything into 16 bits/sample too.

[mdahlstrand]

Thanks for your answer! I'm afraid I'm not a fabricator of platinum electrodes... I hope you find someone who is!

I'm thinking about the immune response from the brain, and whether implementing this array of sensors can be done elsewhere than the actual brain, connecting to nerves instead of neurons, essentially creating a virtual limb. I guess it kind of misses the point of this whole thing, and can't reach as far as brain-implementations, but with the advantage of being more feasible as a solution. I think what I'm getting at is whether we need lots and lots of sensors with very high resolution data, or if we instead can ensure the computer interface is consistent enough that a "muscle memory" can be formed for controlling the virtual limb. I guess I don't have a question really haha, thanks for your time and replies!