There is no global way to improve an experimental design. It depends on what the researcher wants to measure. There are two statistical properties and two psychological properties that can be optimised. It is the researcher's choice to decide the relative importance of each of those properties. The two statistical properties define the main possible outcomes of an experimental design. However, there are two potential unwanted psychological effects that can be avoided by optimising the psychological properties.

Statistical properties

• Estimation efficiency: A possible goal of the research is to estimate the exact shape of the signal (the haemodynamic response function - HRF ) after a stimulus has been presented. This can be interesting for research into how the HRF varies accross subjects/regions/... This outcome is also desirable when estimating connectivity between region/time series. The optimisation algorithm will improve the likeliness to have a good estimate of the brain response at following a stimulus with a good temporal resolution.
• Detection power: Another potential goal of an fMRI study is to detect brain activation related to a certain task. The keyword is contrast. For example: the researcher wants to find which part of the brain is responsible for the contrast between seeing faces and houses, the contrast between angry and happy faces, or the contrast between auditory stimulation and nothing (baseline). In this case, the optimisation algorithm will improve the statistical separation between the modeled conditions.
• Estimation vs. detection: The difference between the two testing settings is in the design matrix. Estimation uses a finite impulse response (FIR) model, while detection uses HRF-convolved regressors. In general, detection power is more common.

Psychological properties

• Predictability: When a subject in the scanner can predict which stimulus will follow, the effect of that stimulus might be biased. As such, it is important to avoid certain contingencies.
• Stimulus frequencies: Sometimes, you can get higher detection power or estimation efficiency when the frequencies of the stimuli are slightly changed. For example: in a stop-signal experiment, you want 30% of the stimuli to be stop-trials and 70% to be go-trials. However, the genetic algorithm has a better detection power when the frequencies are changed from 30-70 to 25-75%. In certain tasks (like a basic stop-signal task), changing these frequencies can have a large psychological impact and you want to avoid changing the frequencies.

In the first step, you are asked some parameters about your design/scanner:

Design parameters

General

Trial structure

We ask to specify the stimulus duration. However, very often your trial consists of fixation cross etc, where you don't expect activation. Therefore, you should specify the duration of the trial before the stimulus (or target) and the duration of the trial after the stimulus.

Contrasts

If you are comparing conditions (or a condition with baseline), you need to prepare contrasts for your analysis. The optimisation of your design also depends on these contrasts. If you have 3 conditions, and you want to compare all of those, your contrast matrix would be:

The first line compares condition 1 with condition 2. The second line compares condition 2 with condition 3 and the last line compares condition 1 with condition 3.

By checking to inclide all pairwise contrasts, all these one-by-one contrasts will be added.

If you want to specify custom contrasts, you can do so too. In an example with three conditions - seeing angry faces, seeing sad faces and seeing neutral faces - you'd want to compare the average of emotional faces with neutral faces. Your contrast would look like this:

Inter-trial interval (ITI)

You want to specify the time between stimuli, the inter-trial interval (ITI) or also known as the inter-stimulus interval (ISI). You can specify a fixed ITI, but also a variable ITI. We strongly suggest a variable ITI for a better estimation of the activation. For a fixed ITI, you only have to fill out the average ITI; for a variable ITI, you can either sample from a uniform or a truncated exponential distribution. You can also specify a minimum and a maximum and. For a truncated exponential distribution, you can additionally add the mean of the ITI's (which should be smaller than the mean between the minimum and the maximum).

Rest blocks

Very often, a researcher would include some rest blocks during the experiment to let the subject rest and refocus. You can choose how long these rest blocks are and how many trials there are between rest blocks.

Note that if you're modeling a very long experiment, it might be advisable to optimise 2 different runs rather than 1 long run with a long rest block.

Design optimisation parameters

On the previous page in the tutorial, we explained the different optimisation criteria. Here, you are asked to give weights to each of those criteria. Each criterium is scaled between 0 and 1, with higher being better. The total optimisation will be the weighted average of those criteria.

Psychological confounding factors

Trial contingencies

Trial blockedness

Imagine an experiment with 3 stimulus types. The probability to have randomly a block of 6 stimuli of the same type is 1/3^6. However, with a design of 300 trials, the probality of a random design to have a block of 6 stimuli is close to 0.4. As such, even a design that is completely random has a large chance of containing a long block of the same stimuli. You might want to avoid such long blocks, it might interfere with the expectance of the subject by setting a maximum number of repeated stimuli.

Watch out though. In an experiment with only 2 stimuli, there is on average a block of 6 repeated stimuli every 64 trials. With only two stimuli, it will not only be difficult to randomly find a design without such blocks, but it might also inversely impact the expectance of the stimuli. You might want to set the maximum number of repeated stimuli higher or leave this field empty.

Contrasts and Probabilities

Stimulus probabilities

Here, you are asked to set the probability of each stimulus type. You can fill this out in whatever unit, as long as they are relative to one another. If you want a design with 200 trials with 4 stimulus types with equal probabilities, you can fill out [0.25, 0.25, 0.25, 0.25], or [50,50,50,50], or even [1,1,1,1].

You can fill out integers or decimal values, but not fractions. For fractions that have infinite decimals (like 1/3), you can ignore the denominator. For example [1/3,2/3] can be written as [1,2].

Hard limit on the probabilities

During the design optimisation, it might (and it probably will) occur that the final probabilities are not exactly what you wanted for the sake of having an optimal design. To avoid this, you can set the probabilities as an optimality criterion, but it might still not be perfect. If you only want designs with the exact probabilities you wanted, you can check this box.

However, this will have a large impact on the optimisation. It will largely restrict the number of designs to search over, and the optimisation will take a lot longer and we might even end up without optimal design. If you check this box, we will automatically set the number of designs in each step of the algorithm a lot larger to avoid finishing without design, but there's nothing we can do about the duration.

Contrasts

We've explained in short how to construct a custom contrast above. For a clear overview and more information, see this video from Jeanette Mumford's blog.

In the review panel, there is a button that will take you to some settings of the design and the genetic algorithm. These settings are based on research, so you might leave those to the default settings. We will explain all these parameters in short here.

Design and optimisation parameters

Genetic algorithm parameters

To understand these parameters, ideally you read about the genetic algorithm (see methods). Here's the algorithm in very basic terminology.

1. First parent generation: A certain number of designs are constructed to start
2. Cross-over happens: The designs make babies by mixing part of one design with part of another design.
3. Mutation: Babies aren't perfect clones of their parent designs. A certain percentage of trials are randomly changed for other trials.
4. Immigrants: To avoid inbred designs, new baby designs are added to the already existing baby designs.
5. Natural selection: the best babies are selected and now become the next generation of parents.
6. Step 2 to 5 are repeated for a large number of generations

In the tab 'Genetic Algorithm Status', you can start the genetic algorithm by clicking 'start'.

The first step will be to find the maximum values of two of the optimality criteria (unless they are set at 0). You'll be able to see the percentage done.

Next, the genetic algorithm starts. You'll see a figure with the weighted optimality, and the individual optimality criteria. You can hover over the lines to see the exact values. Below, a figure will show the expected BOLD signal (in a perfect voxel without noise) for your design for the different conditions of the best design so far.

Once the optimisation is done, there will be download button visible that will allow you to download a zip-file with a file for each stimulus with the onsets of that particular stimulus.

During the optimisation, you can go to the different tabs to check your settings. However, once the algorithm has started, changing these parameters will not affect the optimisation.