Machine Learning Tutorial #7: Group Analysis


Overview

As in our tutorials on fMRI analysis, our goal in analyzing this dataset is to generalize the results to the population that the sample was drawn from. In other words, if we see a significant classification accuracy in these subjects, can we say that it would likely be seen in the population as well?

With MVPA, we have two different types of group analysis we can do: One with the ROI results, and one with the searchlight results. We will review how to do each of these in turn.

Group Analysis: ROI

One of the first-level analyses we did was a classification analysis within a region of interest, or ROI. In this case, we used the masks created in the original Haxby study, and generated a single classification accuracy for each condition for each subject. When we scripted our analyses in the last chapter, these results were stored in the directory ROI_Results; in order to do a group analysis, we will need to extract the classification accuracy for the conditions we are interested in, and then compare the accuracy across conditions.

The code below will extract the classification accuracy and store them in variables. For example, the cell (4,4) indexes the classification accuracy for “Faces”, since that was the fourth condition that we entered in our list of labelnames for the classifier:

%%% Define the variables

subjects = [1 2 3 4 5 6];
Group_Results_Face = [];
Group_Results_House = [];
Group_Results_Scissors = [];

%%% Extract confusion matrix accuracy for Faces, Houses, and Scissors

for subject=subjects

    subject = num2str(subject);

    ROI_Results = load([pwd '/SPM_Results_' subject '/ROI_Results/res_confusion_matrix.mat']);

    Group_Results_Face = [Group_Results_Face, ROI_Results.results.confusion_matrix.output{1}(4,4)];
    Group_Results_House = [Group_Results_House, ROI_Results.results.confusion_matrix.output{1}(5,5)];
    Group_Results_Scissors = [Group_Results_Scissors, ROI_Results.results.confusion_matrix.output{1}(6,6)];

end

We can then compare the classification accuracies for Faces against that of Houses, for example; or that of Faces versus Scissors:

%%% Example t-tests %%%
[F_H_h, F_H_p, F_H_ci, F_H_stats] = ttest(Group_Results_Face, Group_Results_House);
[F_S_h, F_S_p, F_S_ci, F_S_stats] = ttest(Group_Results_Face, Group_Results_Scissors);

In the first case (Faces compared to Houses), the average classification accuracy within this ROI is virtually identical - the resulting t-test is not significant. The comparison of Faces against Scissors, on the other hand, yields a t-statistic of 5.80, p<0.001, representing a significant difference between the two.

Group Analysis: Searchlight

The group analysis of the searchlight results requires more steps. Remember that our preprocessing pipeline did not include normalization, since we wanted to introduce as few interpolations as possible; this also enabled us to use subject-specific masks for our ROI group analysis.

Our searchlight maps, located within the SPM_Results directories, will instead need to be normalized before they can be used in a group analysis. To do this, open SPM, and then open the Batch editor. Select BasicIO -> File/Dir Operations -> File Operations -> Named File Selector. Set the Input Name to “Anatomicals”, and create 6 File Sets. Select each subject’s realigned anatomical image for the corresponding File Set, and then create a Segmentation Module by selecting SPM -> Spatial -> Segment. Highlight the Volumes field and click Dependency, selecting all of the images from the Named File Selector. In the Save Bias Corrected, change the selection from None to Save Bias Corrected. At the very bottom of the window, change Deformation Fields to Forward, and then click the Green Go button. This will only take a minute or two.

When it has finished, go back to the SPM GUI and select Normalise -> Write. Create 6 new Subject fields, and select the corresponding deformation field in the anat directory and the res_accuracy_minus_chance.nii image for each subject in their results folder. Click Done, and then click the Green Go button. The normalized classification map for each subject, which now has a w prepended to it, can be overlaid on an MNI template of your choosing to make sure that it was normalized properly. You may also choose to smooth these normalized results in order to make the assumptions of the t-test more valid; to do this, click on Smooth and select the normalized images in each subject’s folder.

From the Matlab terminal, create a new directory by typing mkdir 2ndLevel_GroupResults. Then click on Specify 2nd-Level and select this new directory in the Directory field. Double-click on the Scans field, and select all of the smoothed and normalized accuracy maps; then, click on the Estimate button and estimate the model. When this is finished, click on Results. Just as with the fMRI studies, you can create a single contrast weight to determine where the classification accuracy is above chance. In this example we will use a voxel-wise threshold of p=0.001 and a cluster threshold of 50; see the link below for a guide to cluster correction if you want to fine-tune this. Notice that the results are strongest in the ventral temporal and occipital areas, similar to the results reported in the original Haxby paper.

Note

For a whole-brain analysis of accuracy minus chance, the assumptions of a t-test may not hold completely; a paper by Allefeld and colleagues found that another method called prevalence testing may be more appropriate. In our case, this method may not be sensitive enough to work with only six subjects, but we will explore other approaches in future chapters.

Video

For a video demonstration of group level analysis with MVPA, click here.

Next Steps

Now that you have learned the basics of MVPA from preprocessing to group analysis, you are ready to use The Decoding Toolbox on a dataset of your choice. In the next chapters, we will review other features of The Decoding Toolbox, such as Representational Similarity Analysis (RSA).