Experimental Design and Analysis in fMRI for Cognitive Psychology

Hello and welcome to the course basics of experimental design for cognitive psychology. I am Dr. Arwarma from

department of cognitive science at ID Kpur. This is the eighth week of the course where we are discussing

methodologies of experimental research. In this lecture, I will continue talking about neuroiming methods.

Now we have seen uh a broad we given you a broad overview of fMRI and PE in the previous lecture. In in today's lecture

I basically going to touch upon very superficially some of the design questions in fMRI experiments. So as we

know fMRI measures hemodynamic changes. It does not measure neuronal activity directly and the signal the uh ratio of

oxygenated to deoxxygenated blood basically varies uh by physiology. It is a physiological signal. It basically

also is affected by heart respiration patterns and so on. The scanner and the motion basically also introduce

additional noise. So I'll show you the picture of a scanner shortly and any kind of motion that the uh body does

also introduces additional noise which needs to be taken care of when before you are interpreting data from fMRI. It

requires statistical modeling generally general linear model to infer the activation the cause source of

activation and how the activation must be interpreted. Sometimes if you're not very careful with the design of an fMRI

experiment it can lead to false or even misleading calculations. This is typical example of the fMRI

scanner. You can see this is the uh you know circular magnet where the person is laying and the magnet basically is

because of the u proper magnetic properties of the blood. It is able to uh track where the blood flow is more or

less in the uh you know in the in the brain which areas are receiving more oxygenated blood which has feromagnetic

properties and which areas have less receive less oxygenated blood. So that is basically what we are trying to do

here. Now neural activity uh wherever the brain is involved in uh any cognitive function there is more neural

activity and this neural activity increases the metabolic demands on the brain. So the brain starts consuming

more oxygen. The cereable blood flow increases uh and you know it reaches uh that area of the brain to provide for

more oxygen consumption which basically reduced to deoxxygenated or deoxy hemoglobin in a particular area which is

basically referred to as the blood oxygen oxygen level dependent response or the bold response which forms the

basis of most of the fMRI studies that we you will talk about. Now as we know this is an indirect measure. Bold

reflects mostly vascular uh activation basically how the blood is flowing through these uh you know arteries and

it is not merely a neural signal. It is it has a larger spatial uh spread because the veins are you know some of

the veins there is a lot of blood flowing up and some of the veins are draining because the blood uh flow is

moving from one direction of the brain to the other. Also sometimes these signals may originate away from the

actual neurons where the activity is happening and this is something that we should keep in mind for better

interpretation of fMRI data. Now there is certainly a temporal delay in the development or in the buildup of

this bold response and this delay is generally to the order of uh uh you know neural events typically happen uh in

milliseconds but the bold onset the blood oxygen level dependent response. It basically onsets generally around 2

to 4 seconds after the neural activity has happened. It peaks in around 4 to 6 seconds and it returns to its base in

around 12 to 16 seconds. So you can see the spatial activity in the brain or the uh this uh you know vascular activity in

the brain the blood oxygenated oxygen level dependent response actually builds up much later than the neuronal activity

has happened. That is the reason uh that there is very limited temporal precision. There is a very uh you know

uh constricted temporal resolution of this method but it has a very good spatial resolution which basically tells

you which areas of the brain were involved in performing a specific cognitive function. The uh basis of how

we analyze fMRI data is based on the hemodynamic response function. The hemodynamic response function contingent

to the cognitive operations I'll show you very shortly is it is slightly delayed and it uh rises gradually. It

includes peak where it reaches its maximum and the undershoots when it sort of comes down to rest and it varies

across different brain regions. So this is also something important that we don't know uh that the HRF or the

hemodynamic response function in different brain areas will be different and it makes the interpretation of this

data slightly difficult. Canonical HRF is what is typically used in modeling of this data and statistical analysis.

So uh sometimes there are uh you know overlapping uh HRF functions which create ambiguity of which reasons uh

regions of the brain they are emanating from. Uh sometimes the when the stimula are placed very closely together we we

have seen the time course of the development of the bold response. If the two stimula are placed very close in

time with each other let's say 2 seconds or 3 seconds then they will have overlapping HRS and it'll make it

difficult and will create ambiguity. design therefore must you know the a good experimental design for fMRI must

uh minimize colinearity that the uh you know uh stimula of interest are not presented just one after the other with

small time in between they should be actually well spaced out so timing optim optimization in these cases is extremely

critical now one of the ways in which we can sort of work out the hrf function is through

convolution which basically combines and overlaps the neural events with hrf it produces predicted bold signals uh with

uh you know the in the areas where the brain uh uh is active and it is the one that is used to create regressors that

will basically help us model the response or model the hrf function. This is one of the core steps in the

generalized linear model that is used to understand or analyze the uh fMRI data. This is basically the explanation of the

hrf. You can see the HRF describes how neural activity is translated into a measurable bold signal. After a neural

event, say some event has happened, a phase has been seen, the blood flow increases leading to the delayed uh rise

in signal, the response peaks at around four to six uh uh you know seconds and then it returns to the baseline.

Sometimes an undershoot often follows due to delayed vascular recovery so that the response is still there. So this is

something that uh you know as uh uh statisticians as data analysts fMRI researchers are actually looking for.

Convolution is uh here say for example neural events when they are plotted together when they are combined with the

HRF through con convolution this pro this produces the predicted bold signal using the generalized linear model. So

you can see these blue uh peaks are where the stimulus is presented. The hr function is this one and the bold

convolution basically it it tells you how the signal is coming in response to the neural activity.

Uh that is why uh experimental designs in fMRI can be sometimes uh blocked or event uh related and we will talk about

them. Block designs however are uh you know better as they produce sustained signals with relatively high power. You

can see these blue uh uh you know sign curves are the block designs which basically create sustained activity in

the bold response and the these orange ones are the event related designs which discrete trials are presented with the

uh critical stimuli. So two types of designs are used. Block designs which basically create sustained activation.

Event designs which are event related fMR designs which basically involve presentation of stimuli one by one one

by one in discrete trials. There are also mixed designs which basically include a both sustained and transient

sort of buildup of activation. And the choice of which design and the choice of which design that you want to follow

basically depends on the research question that is being uh attacked in a given fMRI experiment.

Now block design basically has a typical onoff condition. The stimulus is present then there is silent for uh some period

then the stimulus is presented again and there is silence for some period then it is presented again. It typically

produces a strong cumulative signal in response to that particular stimulus and it makes it easier for the researchers

to detect the activation the bold activation in response to that particular stimula.

Uh block designs also have pros and cons. For example, the very good advantage is that it has high

statistical power. It basically allows for robust detection of activation. But in in some cases you can see that it is

very predictable and it might uh have an influence on the way the brain is cognitively processing the stimulus and

it is something that cannot uh you know isolate individual trials when specifically in time a particular

stimulus was presented. So that is why to uh you know counter that an event related design was uh

thought of which basically uh allows for discrete stimulus presentation. similarly are presented in separate

trials uh with sometimes random or jittered uh timing. Each trial is modeled separately. So each trial when

the bold activity develops and goes to the baseline is modeled separately and it therefore allows trial level

analysis. So when you want to compare different kinds of stimula within a particular block, the event related

design is the one that is more suitable. It also has its pros and cons. For example, the event related designs are

more flexible and they are more realistic. They can examine variability of how the brain reacts across trials,

but they're basically slight uh you know uh they are lower sustained activation than block design and it requires many

more trials to actually be able to infer how the brain activity is happening in response to specific stimula.

Now the mix design basically has best of both worlds. It compi it combines both block level and event related effects.

It captures both sustained attention and transient responses across trials. It is slightly more complex. Therefore, more

complex modeling is required to analyze data from a mixed fMRI design. Timing is of extreme importance in fMRI design.

So, the inter stimulus interval actually determines whether the responses will be overlapping or not. Typically it is

preferred that the stimula are evenly spaced and they are very well separated from each other so that the hrf for

different functions are non-over overlapping and not creating ambiguity. Short ISIS if they are there they will

create overlap in the hrfs. Long ISIS will basically sometimes lead to inefficient scanning. So if the uh two

stimula are placed farther than 10 millconds or 8 millconds from each other then also it will lead to inefficient

scanning. So a balance is absolutely necessary that you present the stimulus for the exact amount of time so that it

leads to the development of the bold response but it is not so far spaced apart that you cannot actually interpret

why and how these effects are building up. Jittering is is therefore a very important uh you know uh mechanism that

you can you know jitter or randomize the timing of the stimulus presentation. It basically improves the estimation of the

HRF because uh you know one is happening and then after uh you know an un specified time of uh uh amount of time

the other trial is happening another unspecified time of uh amount of time another trial is happening. It basically

makes interpretation of the HRF much more easier. It reduces the effects of pred predictability across participants

and across uh individuals and therefore it helps in how the individual is processing a given stimulus.

Both of these are more efficient and basically the efficiency of an fmri design uh depends on the design matrix.

The goal is to basically achieve maximal statistical power and there is obviously a trade-off between detection of

activity and estimation in terms of the neural activity. Now the overall analysis and this part I will do very

quickly because uh I'm not going to cover the generalized linear model uh in in any detail. They basically model

vauil time series. So the uh volutric section of the brain the vauil and how the activity is developing over time in

that it uses multiple regressions and bold uh and basically tries to predict the blood oxygen level dependent

response from the design uh features. Basically the idea is that the observed signal whatever you are getting part of

it will be predicted part of will be error due to different reasons. The model is created such that it can

estimate uh and minimize error. A very good fit is basically one that leads to accurate in in accurate inferences and

uh takes away the noise from the data. Now the design matrix basically is how the experiment is actually designed. The

columns are the regressors which are uh going to be used in interpreting the data. The rows are the time points where

the data is uh taken from and it basically encodes how the different experimental conditions are involved in

the experiment. The regressors basically are task regressors which are convolved with the HRF. Motion regressors are used

to remove artifacts. Drift terms model basically slow and understand the changes. Uh there are also beta weights

which basically uh look at the contribution of each regressor that you're using to estimate how the bold

function is actually going on and this is the one that is absolutely useful for hypothesis testing. Uh basic analysis in

fMRI is basically contrasting the two conditions the task condition and the baseline condition and this is basically

what is most important. So this is what is referred to as the subtraction condition. You look at the activity of

the brain uh at rest at the baseline level and you compare it with the activity of the brain at uh the time

when the task is basically getting performed and the task is going. You subtract that you get the activity that

is contingent to the specific task that you have asked the brain to perform. So it basically is is a extremely important

aspect of the analysis. Then you get your statistical maps. Each vauil is analyzed independently. It produces

statistical map of activity uh throughout the brain and it can be visualized as the brain activation

during the particular task or the baseline. There are multiple comparisons of these kinds possible. Thousands of

vauels are there to test it is and because there are thousands of these comparisons possible. There is a high

chance of type one errors or false positives. Something that we were talking about in uh ERP data analysis as

well. So it requires a lot of cleaning. It requires a lot of correction. There are different kinds of correction

methods. Uh you know you can have different kinds of uh cluster correction using spatial extent and so on. Uh I'm

not going into a lot of detail of this because we're getting into the analysis part. But the idea is that this data

really uh you know it's it's a lot of data. It's rich data and it needs to be cleaned very carefully in order to reach

the correct inferences. There is a pre-processing uh protocol there. So raw data contains obviously

noise pre-processing removing the initial noise will improve quality. It is definitely essential before you uh

you know push the data into analysis and the analysis pipeline you'll have to have slice timings which areas which

time periods of the brain you are basically acquiring uh you have to uh you know use motion correction uh

realignment. So uh any sources of motion have to be uh you know taken out. Normalization of the standard brain uh

basically smoothing also improves uh this function. Uh you have to uh remove motion related artifacts correct via

realignment and these motion parameters are added to the GLM so that it takes care of them while interpreting the

overall uh thing. Typically uh one of the very important steps is normalization. It maps the

entire brain to a standard brain a standard space. It allows for comparison between participants and allows for

comparison between groups. And it also enables averaging across subjects. So that you know that okay vauil a bc is

the same as voxil d uh ef in subject one versus subject two and you can actually compare the activities in these areas.

Smoothing uh is required to sort of take away the uh extra noise, improves the signal to noise ratio but it reduces you

know spatial precision in terms of which specific areas the uh activity was arising from. Obviously there are

several sources of noise. So heartbeat, respiration, scanner drifts, subjects motion artifacts all of these are there

which as I have said you have to be taken care of. Now interpretation of fMRI is extremely important. it bas one

needs to understand that fMRI basically shows correlation data and not causation data. This is something I was mentioning

in the previous lecture also and activation does not involve the necessity of that area. If a particular

area of the brain is active during a particular uh cognitive function, it does not imply that that area is

absolutely necessary for that cognitive function to be performed. Therefore, a very careful interpretation of fMRI data

is uh is required. The fundamental uh rule that one has to remember is fMRI is correlational data. It is not causal

data as to it does not really tell definitely that this area is definitely involved in this particular cognitive

function. Obviously there are some uh you know limitations of the uh method. For

example, as I said the bold signal takes times to develop takes time to develop and therefore it has low temporal

resolution. It is an indirect metabolic measure. So you it it is not direct neuronal activity that we are measuring

and it is obviously s sensitive to vascular defects defects in patterns of blood flow and so on. So these are some

of the things that I sort of wanted to talk about. There are obviously best practices that you have to optimize your

design before scanning. You have to control for confounds. Everything that we have talked about in experimental

methods so far is something that you have to take care of in any of these experimental methods that you're using.

whether it is eye tracking, whether it is ERP or whether it is these neuroiming methods, avo proper corrections, control

confounds and obviously most important of all avoid over interpretation. So these are some of the things that I

wanted to talk about and uh I hope this course has been uh useful for all of you and you have learned some of the

critical aspects of experimental design and a little bit of a uh overview of different experimental methods that are

used. Thank you. I hope this course will be useful for everybody.