Sergey D Stavisky et al 2015 J. Neural Eng. 12 036009
doi:10.1088/1741-2560/12/3/036009

A high performing brain–machine interface driven by low-frequency local field potentials alone and together with spikes 
Sergey D Stavisky, Jonathan C Kao, Paul Nuyujukian, Stephen I Ryu and Krishna V Shenoy

[Title slide for a few seconds] 
[Sergey speaks into video, rig as background]
Welcome to the video abstract for our JNE paper, succinctly titled "A high performing brain-machine interface driven by low-frequency local field potentials alone and together with spikes". My name is Sergey Stavisky.

In this work, we set out to improve the long-term reliability of intracortical brain-machine interfaces, or BMIs. 
[video cuts to Keynote slides]
Most such BMIs work by decoding the extracellular spikes recorded using multielectrode arrays. When arrays record excellent spike signals, as they do in this example from our monkey J, this enables very high performance. Unfortunately, in other cases  - such as our monkey R - the arrays have fewer spike-detecting electrodes. A major challenge for BMIs is that eventually many arrays lose the ability to record spikes. 

Fortunately, an electrode that does not record spikes can often still detect a lower frequency neural signal called the local field potential, or LFP. We began this project with an offline decode comparison of different LFP features and found that the local motor potential , or LMP, was the best candidate feature. 

LMP is just the smoothed LFP voltage, and it is well-tuned for arm reaching velocity. These colored traces show the trial-averaged LMP during reaches in eight different directions. If we then half-wave rectify  these waveforms and compare their amplitude, we can see that this example electrode?s signal has a preferred direction of up and to the right.  Across all electrodes on our two arrays, LMP preferred directions span the two dimensions needed to control a cursor. 

We then implemented a closed-loop LMP-driven BMI. This video shows a monkey using it to control the velocity of the white cursor and acquire the green targets. Above, you can see each electrode's LMP feature. The control quality is far from perfect, but it does allow the monkey to effectively acquire the targets, and represents a roughly 50% improvement over the best previously reported LFP BMI.

Here we compare average performance when each monkey controlled the cursor with either his hand, decoded spikes, or decoded LMP. Importantly, the LMP-driven performance was still worse than spikes-driven performance.

We next wanted to see how well the BMI would perform if we decoded both LMP and spikes  together in a so-called ?hybrid decoder?. We found that in the monkey with mediocre spikes, this improved performance, but when spikes were much better than LMP , hybrid decoding was actually detrimental. 

Monkey R had useful LMP on many electrodes, whereas his spikes decoder relied on a small number of key electrodes. What if these best electrodes were lost, as might be the case if the arrays decay?   We tested whether hybrid decoding could rescue performance in this scenario.  As we disabled more and more electrodes, we found that hybrid decoding became increasingly useful for sustaining performance.  

Here's an example when the best 60 electrodes have been removed and both decoders are compared side by side. Spikes-only decoding, on the left, is virtually useless. In contrast, when LMP from the remaining electrodes is also put to use, the monkey is again able to acquire targets. 
One can easily imagine that if this BMI belonged to a patient with paralysis, this would be the difference between a neural prosthesis that still works and one that doesn?t.

[camera goes back to Sergey speaking]
We conclude that low-frequency LFP can serve as an effective alternative BMI control signal when spikes are unavailable, or as a complimentary signal when spikes signal quality is poor. We hope that this method can extend the lifespan of clinical BMI systems, and that this study motivates further closed-loop testing of different LFP decoding methods.