Transcript for:
Baseline Normalization of Time Frequency Power

all right so moving along with the theme of baseline normalization of time frequency power in this lecture I want to discuss how to pick a baseline time window what it means to have a baseline time window and what are the implications of choosing not only the baseline time window but also how you baseline normalize across different conditions different experimental conditions the important thing to keep in mind is that baseline normalization does not change the time course of any of the data the only thing that baseline normalization does is shift the y-axis so it's going to change what values are zero what values are negative and what values are positive otherwise baseline normalization is not fundamentally changing the actual time series of activity and so what I have in this little toy example here is an illustration of of the importance of interpreting your results relative to the baseline and knowing which baseline period you've selected so here you can see the same cell time course of time frequency power you know these little five time steps choosing two different baselines and you can see that these two different baselines do not change the time course of the activity all they do is shift everything up or down on the y-axis and so here the time bin so ten bin one two three four five and here in this top illustration I've selected the first time bin to be the baseline and here the second time bin to be the baseline and now where this makes a difference is that for example in time bin three do we call time bin three zero change in activity or is time bin three actually a negative pass or suppression of power reduce power and the answer of course is that it's it's neither and both at the same time it really depends on your on your baseline and so this is an important point to illustrate that you do a baseline normalization your interpretation of the results has to change because you can no longer speak of absolute levels of power you can only speak about and interpret and also statistically analyze relative changes in power so power relative to whatever you specified as the baseline and so so you can have the same amount of power and interpret that either as zero change in power or a negative change in power suppression of power depending on what you pick as your baseline so this is really a non-trivial issue here I have some examples of the same data same electrode same analyses the only difference between these amongst these three plots is which time period I used for the baseline normalization and we will reproduce this figure in MATLAB in a few minutes so here you can see I used a baseline period of minus 500 milliseconds to minus 200 so that's from here to here this one is minus 300 to zero and this one is minus 100 to plus 200 I'll start with with penalty first obviously this is a ridiculous and obviously terrible choice for baseline activity because it includes post stimulus onset activity but you know it's an extreme case that really illustrates the point that if you if you compare panel D to panel B a lot of this post stimulus activity is gone and what you what it what appears to be happening instead is that you of this there's really strong suppression of of you know whatever this is up birthday theta to alpha activity in the pre stimulus and post stimulus period but really this is not or it's not really interprete Balazs a suppression it's really just that the increase the post stimulus increase in power sort of leaked back into the pre stimulus period and and made this a and could yeah change this into it into a relative suppression so obviously this is a terrible choice but you know it's interesting to see these extreme cases to really sort of drive this point home so there's a bit of a more subtle difference between panels B and C and here you see they generally show similar results in panel B you see a bit stronger theta a response is kind of sustained theta response that starts earlier that's a bit more prominent here in panel B compared to here in panel C and what's happening you can see that a little bit of this power is leaking back before time zero so this is an interesting situation it is possible in your experiments if you have a constant inner trial interval or some kind of a pre stimulus warning cleek you that your subjects know exactly when the next stimulus is going to come on the screen so they can kind of anticipate it they start processing stimulus already before it comes on screen that's a situation where you could get a brain activity before time zero but I think what's more likely in this case what often happens that activity that happens very quickly after time zero because of the inherent temporal smoothing properties of time frequency decomposition some of the very early activity is going to leak into the pre stimulus period so this is temporal leakage and probably what's happening here is that you know the in in the actual brain in a real experiment there was an oscillatory response that began fairly early somewhere around here and because of the temporal smoothing properties from the wavelet convolution that kind of yeah I leaked a little bit into the pre zero time bin and so then when the baseline goes all the way up to exactly times zero we're actually including a little bit of this post stimulus activity into the baseline time window and that pushes up the baseline power and then the task related power it seems to be relatively less and in fact you can also see a little bit of a suppression here an apparent suppression which you don't see here and that's really just because a little bit of this kind of shoulder leakage is is bleeding into this earlier activity so for exactly this reason because of the the decrease in temporal precision and the temporal smoothing resulting from time frequency analysis particularly at lower frequencies I think it's generally a good idea to do have your baseline period not go all the way up to x 0 but instead to end a little bit before x 0 so something like minus 500 to minus 200 euro of minus 400 to minus 200 something like this of course the exact choice of the baseline period depends a lot on the specific design of your experiment and so you know I I couldn't say that you should always use minus 500 to minus 200 because that's not going to work for everyone and for every experiment the point here is that you really need to think very carefully about what to use as your pre stimulus baseline period so let's switch over to MATLAB I'm not going to go through all of this code line by line because most of it is is is is stuff that you've seen before so here I'm going to choose a bunch of different baseline time windows and basically we are going to be reproducing this plot that with that I was just discussing in PowerPoint so yeah here we specify the time windows and you know of course I encourage you to change these numbers around we're going to do four separate analyses using these four different baseline time windows and you can change these numbers to see what the effects are going to be here we set up our wavelet analysis properties we're going to use channel oh one you can of course change any number of these of these parameters here's where we actually do the wavelet convolution of course all this stuff looks familiar you can see that we are not actually changing any of we're not performing any normalization here in this plot or in this code instead all the normalization happens here and here I'm going to compute decibel normalization and also time frequency power change so you can see both of these results so let's see I believe this one is just plotting percent change and yes these are not exactly the same data as I showed in a PowerPoint slide but you get the idea so here again is this awful choice of baseline which is actually really the post jameelah's period of course that obliterates the actual stimulus related activity and it puts all of this activity or the inverse of all of this activity into the baseline and post stimulus period let me see-oh this also it does compute decibel so here in Figure one you see the decibel normals normalized results here in Figure two you see the percent change normalized results and there are a few subtle differences between these most of the apparent differences between a percent change and and decibel are actually related to a color scaling so if you would change the color scaling here you could actually get these two plots to match even more than they already do but the important point is that the percent change in decibel change are not really hugely different its enormous ly unlikely that you would come to completely different conclusions about what's happening in the brand based on using decibel or percent change and for that reason unless you have really extreme values it's it doesn't really matter you can choose decibel or percent change whichever you feel more comfortable with so that is basically all I wanted to say about the baseline time window there is another issue related to baseline normalization that happens when you have multiple experimental conditions and here the question is should you compute the baseline power based on the condition average or based on each condition separately and so here imagine you have three time courses of some frequency band specific power for different conditions what we could call this condition a or I guess condition red is it easier to remember so here's the time frequency power for condition red and this is raw values this is not doing any any baseline normalization yet and here's for condition yellow and condition blue so you see that these conditions do differ in terms of their the amount of power that's elicited by the different conditions but it seems like the the phasic response to the stimulus onset is pretty much the same for the different conditions and really what differs is that there is a a kind of atonic shift or a baseline shift between these different conditions now depending on your experiment this might be related to noise or maybe this is related to some meaningful thing let's say you know if subjects were indifferent I know motivational situation at contexts in these different conditions or maybe you have a you're looking at at fatigue or something and this was early in the experiment and middle in the experiment and late in the experiment and then you see these tonic differences so now if you were so now the question is when you apply a baseline normalization so should you average these conditions together to get the baseline the estimate of the baseline activity or should you baseline each condition relative to its own baseline and so there's there's actually no right answer here if you would do a condition specific baseline then the results you would get would look something like this and you would see so I separated them here just so you don't think that I got lazy and only plotted one line but so you know here you see with a condition specific baseline you're removing this tonic offset across these three conditions and then the conclusion you would arrive at is that all these conditions are these three conditions have the same elicit the same amount of time frequency power after time zero in contrast them if you were to use a condition average baseline then you would see so this line is is reflecting zero decibel then you would you would preserve this condition difference in activity but then this would you know you would have to be very careful not to interpret these differences as as being related to phasic differential responses to stimulus but instead these differences reflecting tonic differences between these conditions so you know which one is right if you know these are both correct solutions these are both appropriate and valid baseline normalization procedures but they would give you different interpretations of the results so let me see this is a a fairly unusual situation where you would have such strong tonic differences across the three different conditions so in general I prefer doing condition averaged baseline so you put all the conditions together and then you compute the baseline and then you apply that baseline to each individual condition I like this approach because it increases the signal to noise and you you want to have as pure and high signal-to-noise estimate of the baseline power as you possibly can and so to do increase the signal-to-noise of the baseline you want to average to gather more data basically so you put in longer time windows and you put all of your conditions together that is why I generally tend to prefer a condition average baseline but of course that said you know there are situated there's always some situation there's always some experimental design where the condition average baseline doesn't really make sense and the condition specific baseline does make more sense again there's no right or wrong answer here but it's important for you to think about this issue this really has non-trivial implications for the kinds of interpretations of your results that you will make and yeah so it's something that you should think a lot of think think about very carefully but yeah I guess they okay I have nothing else to say that I haven't already said