Fact or artifact
A new turn in the saga of my data processing. There has been some concern that the artifact from the stimulation is causing the thing I am seeing in my data. There are arguments to be made for both sides, but let’s go over what that could mean for me.
A hypothetical, you’re a researcher and you record electrical activity from the brain. You’re interested in the effect of peripheral nerve stimulation is on the electrical activity of the brain. The catch, using electrical stimulation causes your data to have what is called an artifact. What is happening is the stimulation causes the charge on your skin to increase/decrease depending on how you want to think about it.
Oversimplification and not my experiment, but the point is, no matter how far away you are from your stimulus, it WILL show up in the result. I could record from your brain and stimulate a nerve in the foot and you would see a spike in the data that corresponds to the stimulus onset and only represents the stimulus (as in not any sort of biological signal). It’s a fact of electrical stimulation and normally we live with it. In my case, it’s not that easy.
I’m interested in the effect of the stimulus directly after onset, but when we run a frequency analysis, we need to use a window to find the power spectral density of the signal, the window can be quite large too, we’re talking a one second window to resolve a 2 Hz signal when the response would be ~25ms after onset.
The question then becomes do we ignore the artifact? Try to remove the artifact? Leave the artifact because it won’t effect or data? Or something different all together. The answer isn’t so simple.
When we record EEG we pick up all sorts of electrical signals, one of which is line noise. Literally the electricity in the walls creates a 60 Hz (in the US) bump in our power spectral density analysis. Normally we ignore it, try to do fancy ways of filtering it out, etc. My problem is I have a stimulus that is 600 Hz. Which is far beyond the frequencies I’m interested in (orders of magnitude larger, I’m looking at like 0-20 Hz maybe up to 50 Hz).
In theory there shouldn’t be a problem. If I could resolve frequencies that high (I cannot because my sample rate is 1000 Hz and the nyquist frequency (which I realize now I never wrote about…) is twice our frequency so I would need to sample at 1200 Hz (minimum, we would need to sample higher in reality) to see a 600 Hz signal properly.
Now that I know tomorrow’s topic…
Here is the rub. I cannot see the power spectral density of the artifact since I didn’t sample that fast. This means that I cannot be certain it lives at 600 Hz. In fact, since we are below the nyquist frequency we can say with certainty that it won’t be there and that it will instead appear somewhere else on the frequency spectrum unfortunately.
Now I have some interesting results. They are not groundbreaking persay, but they serve a very important role in validating my new shiny way of doing things. The output is EXACTLY what I expected to find if my method worked. Moreover, it matches other animal model findings as well. This is great news for me… assuming it’s real. The issue is that apparently in certain cases, stimulus artifacts have been known to cause some interesting changes to the power spectral density of the signal.
This little problem throws a wrench in my plans because I cannot say with certainty that the artifact is not causing what I’m seeing. This means I need to do a few things. Or rather I need to start trying to figure out a solution to this problem, which there are a few.
One is to figure out how to filter out the artifact. That probably won’t happen without a lot of input from both of my PI’s. I don’t like this route for that reason, both of them come from different backgrounds so there is no guarantee that they will agree on how to do it fully. While one may be satisfied the other may not be completely and that does not work for me.
The second option is to attempt to show the power spectral density of the artifact to show it does not have an effect on my data. This is probably the best option. When we record EEG we have several different sensors we place to record noise signals that we can use later to filter with. We place four around the eyes to filter out eye movement/ eye blink artifacts for example. I could theoretically use one of those channels, locate the artifact, and run a power spectral density analysis on it to show it doesn’t cause an increase in the frequencies I’m interested in.
Lastly, I could just remove the section of data with the artifact and look at a section before the artifact and directly after the artifact. This would probably be the simplest way to solve my problem and wouldn’t be a bad way to do my analysis. In fact, it’s so easy I may just do it anyway to further complement whatever way I decide to go (which is almost certainly the second option).
So yeah, a lot of work ahead for me! I’m not thrilled with this, but as I tell my Co-PI I appreciate the fact that he’s skeptical, it makes me work for it. I would rather have my dreams crushed and this technique not work at all than to put something out into the world that isn’t real. For one thing, I’ll have other researchers come down on me and that would be far more embarrassing and humiliating than having my Co-PI challenge my findings.
Sometimes writing all this out helps me understand my options better. Today’s post was just as much for me to work out the next step as it was for all of you to see my progress. Good talk everyone. Now I know what I want to do and I’m going to set off to do it.
See you tomorrow!