There’s a recent paper by the Maunsell lab in PLoS Biology that examines the origins of local field potentials in the low-gamma (30-80Hz) and high-gamma ranges (> 80 Hz) ranges. The paper uses a clever experimental paradigm that allows one to dissociate LFPs and spike activity through the phenomenon of surround suppression. Basically, in surround suppression, as the size of a visual stimulus is increased, neurons in visual cortex decrease their firing rate while low-gamma activity continues to increase. High-gamma, however, does not follow low-gamma activity, and indeed decreases as surround suppression kicks in. The implication is that high-gamma contains much of the same information as spikes. This of course begs the question: is high-gamma all garbage? Does it contain any information not present in multi-unit activity? Otherwise it’s pointless to continue to analyze high-gamma data.
This is an issue close to my heart, and we have a recent paper out in J Neurophys that examines this issue from a signal processing perspective. There’s good reason to believe that above a certain frequency much of the LFP is artifactual and actually caused by contamination by spikes, which contain a wee bit of low frequency content that can be amplified in some analysis scenarios. This is easiest to understand when you think of the spike-triggered LFP. Suppose you have access to a wideband signal. Grad student Alice decides to examine spike-LFP relationships by first low-passing the wideband signal to obtain the LFP, then taking the spike triggered average of the LFP; this is the accepted practice of the LFP-STA. Grad student Bob, on the other hand, decides to take the spike-triggered average of the wideband signal, and then low-passing that to obtain the low-passed spike-triggered wideband signal.
What Bob is doing is clearly wrong. If you spike-trigger the wideband signal, the result will be dominated by the mean spike; low-passing the result will spread the spike temporally, but it definitely won’t eliminate the artifact. But what Alice and Bob are doing is in fact mathematically equivalent. That’s because both spike-triggered averaging and low-pass filtering can be described as convolutions, and convolutions commute (if you’re not convinced about the STA being a convolution, read Dayan & Abbott chapter 1). So Bob and Alice are basically doing the same thing and THEY ARE BOTH WRONG. In both cases what they’ll get is something similar to Figure 7 in the Maunsell paper (shown below), an LFP-STA dominated by a large negative peak around the time of spikes, which is pure artifact, as the authors correctly point out. I’ve seen many a poster at SFN with similar figures presented as legitimate results; to be clear, THAT’S WRONG.
Both our paper and Maunsell’s present evidence that above 50 Hz the LFP is highly contaminated by spikes. It could be the case, in fact, that above some frequency, the LFP is pure garbage, that is, it contains the same information as the multi-unit activity (MUA) or the single-unit activity (SUA). Based on the two paper’s findings, I would venture and say that this magic frequency, if it exists, is much lower than the usual LFP cutoff (150-300 Hz), perhaps around 80Hz as the Maunsell paper seems to imply. Indeed, I have yet to see a paper make a strong claim that there is a dissociation between high-frequency LFPs and spikes in any circumstance (though I haven’t read every paper on the subject).
It’s going to be really hard to prove in an absolute sense that LFP above a certain frequency are purely artefactual. I think this is probably easier to attack from a signal processing perspective. The spike removal algo that we present in the paper is designed to remove foreground (sortable) spikes, not background (unsortable) spikes that can nevertheless be detected. My hunch is that the background spikes make up most of the LFP frequency content above 80 Hz. The question is, how can you remove spikes that you can’t even sort? I’m working on this problem these days and I think I have a pretty good solution. No definitive word on the +80Hz issue though. In the meantime, I would be very careful about writing some claim about LFPs above 50 Hz without some very thorough signal processing.
Ray S, & Maunsell JH (2011). Different origins of gamma rhythm and high-gamma activity in macaque visual cortex. PLoS biology, 9 (4) PMID: 21532743
Zanos TP, Mineault PJ, & Pack CC (2011). Removal of spurious correlations between spikes and local field potentials. Journal of neurophysiology, 105 (1), 474-86 PMID: 21068271
5 responses to “LFPs above 50Hz: is it all garbage?”
I’m studying your paper in order to apply the algorithm that you and your collaborators developed to some physiological recordings that has been provided to us (as I’m the one dealing with “spike-contaminated” LFP signal and the one who realised about this contamination). I’m a Post-Doc at IDIBAPS (Barcelona) but I’ve switched fields from Theoretical Particle Physics to Neuroscience so I lack the appropriate knowledge of the signal processing Mathematical stuff that lies behind the formulae and reasoning you use there. Could you recommend me some bibliography to learn the proper points to fully understand your paper and its supplemental materiel (which otherwise is very informative)?
Thanks so much.
[…] in LFPs. If these high-frequency components are similar to spikes (and they are according to another Maunsell paper), then a hand-waving argument says that the characteristic spatial correlation of high-frequency […]
Exactly! Although I’m pretty dubious about detecting high gamma outside of the skull, there are plenty of papers that make it seem possible. But I’m actually talking about subdural recordings in humans. Far fewer artifacts :)
Yes, actually that’s an interesting corollary, if high gamma contains mostly MUA then you can measure MUA (albeit with low resolution) outside the skull with EEG. At least that’s how I interpret this paper by the Logothetis lab: http://www.ncbi.nlm.nih.gov/pubmed/19874794
Excellent post and write-up. So initially I was drawn here with a big “oh crap!” sinking feeling when I saw your title. That said, for those of us working with human cortical data without access to spike data, “high” gamma is a critical spike surrogate for us!