Offset Causing Ripples/Harmonics in Spectrogram that go away with HPF
Hi all,
I have some strange spectrogram outputs from a Cyton recording that I've had trouble wrapping my head around. I would expect a spectrogram of the raw data to show activity fairly cleanly. However, pretty much all I see is very high powered bands approximately every 10Hz (fig 1). When I remove the DC offset with a HPF at 0.005 Hz they are pretty much gone (fig 2 (top plot is absolute value of the filtered EEG)) and I can see EEG data much more clearly. At an even lower cutoff of 0.001Hz the bands seem to show up again - most strongly when the offset is the highest (fig 3 (again, top plot is absolute value of the filtered EEG)). I've been googling around seeing if this is an expected behavior of spectrograms when there is a significant DC offset, but haven't found anything. Is this something expected from the Cyton or generally with signals that have DC offset? I'm plotting one channel here, but it was present on all 4 recording channels (the remaining 4 were turned off).
These plots are showing multitapered spectrograms generated using the Chronux toolbox (function mtspecgramc) but I confirmed it also happens with the MATLAB built-in spectrogram fucntion. Filters were made using the eeglab built-in filters.
Any insight is appreciated!
Comments
Harmonics from a DC offset can occur, usually as a multiple of the main or predominant frequency. The fourier analysis math to support this is quite complicated, mostly beyond my ability to explain, and has to do with the effects of minor fluctuations of the DC offset. Disappearance of the harmonics with filtering the offset supports this.
Brian, hi.
I would even say that your high pass filter at .005 Hz (a waveform that takes 200 seconds to cycle, 3.3 minutes) -- is much lower that is typically used. Even lower is .001 Hz, which corresponds to 1000 seconds, 16 minute cycle time.
More normally you would high pass at .1 Hz or .5 Hz. More discussion on this thread:
https://openbci.com/forum/index.php?p=/discussion/201/large-millivolt-data-values-fbeeg-full-band-eeg
By using a bandpass from say .5 Hz to 45 Hz you take out both the DC offset and the mains frequency. Many also use a notch at the mains frequency.
Regards, William
@Billh Thanks for the insight. It does seem like a very steady fluctuation since the bands are very consistent, no? Is that something about the hardware possibly?
@wjcroft I realize these filter values are a little out of the ordinary but I wanted to illustrate that these bands are related to larger offsets and very slow frequencies. I got into a bit of a rabbit hole of curiosity when just taking a look at the spectrogram of the raw signal and was surprised to see it completely washed out by these large bands, so I was also hoping to show my thought process a little. I figured showing the minimum cutoff value to remove them might help diagnose where they were coming from. If I'm doing any further analysis of these data I'll most likely be using a more typical bandpass and a notch filter per your suggestions.
Here's the issue of ISNR Neuroconnections, on some clinical applications of ILF (InfraLow Frequency) / ISF (Infra-Slow Fluctuations) bands. Quite a lot has been done,
https://docs.wixstatic.com/ugd/cba323_fdd03c418d2348f59879f475b82439ee.pdf
[Fall 2013 issue of ISNR NeuroConnections; "What's happening below .5 hz?"]
https://www.isnr.org/neuroconnections
[all issues]
Surely this isn't a physiological signal, though. It is remarkably consistent over the entirety of that recording - it's cropped to an hour above, but is consistent on all our Cyton recordings. A recording from a different subject shows the same pattern.
ILF / ISF is physiological and thought to be associated with the DC level glial / astrocyte potentials. See the previous link pdf.
You may be getting distorted ILF signals. The clinical usage of this demands silver chloride electrodes (with gel or paste) for their stability, and generally they are also sintered Ag-AgCl. It's also possible that your particular EEG environment might have EMF field distortions.
I've attached a paper that contrasts the various electrode materials, for use in SCP (Slow Cortical Potentials), another name for ILF / ISF.
I didn't mean to imply that there are no physiological signals at these frequencies, but that what these spectrograms are showing are not physiological as evidenced by how consistent they are in frequency and power. I looked at your links and Pubmed and couldn't find any publications showing a) these signals are extremely consistent as to show non-fluctuating bands - even between subjects and environments or b) they are so powerful as to drown out EEG signal in the typical frequency range when the raw signal is plotted. The environment has not been constant for these recordings (assuming you mean physical location). These spectrograms are from a recording done overnight, where electrodes were applied in lab worn overnight at home. I'm not convinced it isn't an artifact from the hardware since that is the most consistent variable. If these are indeed physiological, I would be most convinced if a recording from a different DC-coupled amplifier showed these same bands when the raw signal is analyzed. I haven't been able to track such data down today to test this idea with, however.
Cyton compared favorably to a g.tec medical grade amplifier. Below pdf link. I cannot see a reason for a hardware error causing distortions at other frequencies. Isn't it more likely that your spectrograms are just showing some type of mains harmonic / subharmonic noise?
https://arxiv.org/pdf/1606.02438.pdf
What kind of mains notch filter have you applied in your recordings?
That paper says that all the analyzed signals have been filtered at least above 0.5 Hz (see section 2.2.1), so I wouldn't expect they would see this.
All I have done is high pass filter the data for those plots - no notch filter. You can see in the bottom plot that the 60Hz noise shows up as a separate peak (at 60 Hz) from this offset-related noise (which peaks around 50-55 Hz). There's also not a good correlation between the power of the 60Hz noise and appearance of these bands (look at the clean part of the signal in the last 5 minutes on the bottom plot). If these bands are related to, as Billh suggested, "the effects of minor fluctuations of the DC offset," then is it not possible that the hardware is causing these fluctuations?
Please try your tests again with a proper notch at your mains frequency. Mains noise shows up not just at the exact mains frequency, but at various harmonics. I believe these harmonics are more acutely distorting the ILF range, which is why they disappear when you filter out ILF.
Virtually ALL EEG systems use a notch at the mains frequency to remove these distortions.
This is not the exact same recording, just one I happened to be working with, but the point is still illustrated here: notch filtering does not remove these bands while high-pass filtering does. Filters generated with the eeglab built-in; actual values for the notch/stop band freqs were 59-61Hz.
This looks like some type of harmonics noise to me. I'm not sure why you are trying to pin it on a hardware error. The majority of EEG users would be using say .5 to 45 Hz bandpass, plus possible a notch at mains. If you are doing ILF, you will be LOW pass filtering at .1 Hz and below. In both cases these 'bands' are not an issue.
I understand you are trying to illustrate your process in figuring it out. But now that you have the solution, why not just go with that? Harmonics are somewhat mysterious things. Subharmonics especially.
https://www.google.com/search?q=subharmonics
You could see what happens if you take the filtered data and add back a synthetic DC offset of about the same magnitude. If the bands come back that would mean it is the offset.