Segmentation of EEG Files into Micro-States

 

Segmentation is run first to extract template maps , which can be back-fit to the individual subjects on a second stage.

 

Brief introduction to the segmentation
Segmentation of EEG files into micro-states
How to run the Segmentation
Files Dialog
Parameters Dialog
Technical points & hints
Results
Segmentation in the Source Space

Brief introduction to the segmentation

Segmentation into micro-states is the process of finding periods of stability in the map representation of the ERP's. The data (usually grand means) are then segmented / cut where changes occur , and template maps are computed for each of the resulting segments to act as single representatives of the segments. The segmentation therefore produces:

 

Given a set of averaged maps, it is possible to fit them back onto individual ERP's:

 

For the display of the segmentation, see all the Segments Display .

Segmentation of EEG files into micro-states

Methods used for the segmentation

It is actually made of three successive steps:

  1. A mathematical clustering,

  2. Some temporal post-processings,
  3. Computing some quality measures.

 

Let's introduce briefly these three points:

1. Mathematical clustering

This is a well-known mathematical process, which does not account for the time dimension of the data. That is, data (maps) are put in a "bag", then picked from there without caring for their actual time stamp. Three methods are available for this stage:

2. Temporal post-processings

These are optional processings, which have the nice properties, for most of them, to account for the time dimension. They are totally independent of the type of clustering method used in the first stage. The methods available are (in this order):

3. Computing clustering quality measures

Cartool computes 9 different clustering quality criteria (as of 2024), and then combine them into a meta-criterion. This meta-criterion has a statistically more robust behavior, i.e. it outperforms each individual single criteria alone for correctly guessing the optimal number of clusters.

See the full discussion about clustering criteria .

 

Next, we are going to delve into the details of these three stages.

Mathematical clustering

The K-Means clustering

We summarize here the K-Means algorithm as it is used for the EEG data in Cartool. For the sake of clarity, we use practical numbers, by searching for 5 clusters out of 4 conditions of 100 time frames. Time is not relevant in this algorithm, so the total of 400 maps are just "put into a bag".

 

  1. Initialization

    • Of the templates: randomly pick 5 non-identical maps out of the 400 data maps.

    • Of the labeling: test each of the 400 maps, and label (index) to which of the 5 templates they correlate the most.

  2. Convergence loop

    • Compute new templates: average all the maps that have the same label.

    • Compute new labeling: test each of the 400 maps, and label (index) to which of the 5 new templates they correlate the most.

    • Compute a measure of quality of the current labeling ( Global Explained Variance, GEV) .

  3. Repeat step 2, until the GEV doesn't improve anymore.

  4. If the end results has a higher GEV than the current best one, keep it as the current best.

  5. Repeat again steps 1 to 4, until enough repetitions have been done.

 

At the end, we have an estimate of the theoretically best segmentation.

As there is a random stage in this algorithm, we can only get close to the perfect solution. Also, running again the whole stuff will give slightly different results for the same reason. However, the process is reliable and is a standard procedure. See more technical points here.

The Atomize & Agglomerate Hierarchical Clustering (AAHC)

Important note: the AAHC is not used anymore in Cartool, only its modified version T-AAHC is available. However, the T-AAHC being based on the AAHC, the explanations below still hold.

 

The usual Hierarchical Agglomerative Clustering is a bottom-up approach. It starts with all maps being a cluster, then it merges the two closest clusters (of the current remaining ones) into a new one, repeatively. In this way, we have all the segmentations from the number of Time Frames down to 1.

Cartool implements a modified version of the Hierarchical clustering, specifically tailored for the EEG case, called the Atomize & Agglomerate Hierarchical Clustering (AAHC) .

It behaves the same as a regular hierarchical agglomerative clustering, starting with every map being a cluster. Then the main idea is to pick the "worst" segment of all currently availables, to atomize it (split it into indivividual maps), and to re-distribute these maps to the segments they fit most. This also results in all the segmentations from the number of Time Frames downto 1.

 

  1. Initialization

    • Of the templates: each map is a template, starting therefore with 400 clusters.

    • Of the labeling: each map labels itself.

  2. Removing one cluster

    • Pick the worst cluster: the one with the lowest GEV (Global Explained Variance) .

    • Atomize & distribute: take each map of this cluster, and assign it to the remaining cluster it correlates the best.

    • Compute new labeling: test each of the 400 maps, and label (index) to which of the remaining templates they correlate the most.

  3. Repeat step 2, removing one cluster at a time, until we reach 5 clusters.

 

The main advantages of this method over the regular Hierarchical Agglomerative Clustering are:

There are many advantages of this method over the K-Means:

 

A final word could be: try both segmentations, and see the differences by yourself, it's not good to trust blindly (even Cartool). Actually, if there is something in your data, it will show up with both methods anyway!

 

See f.ex. a AAHC versus a K-Means for 18 segments (though it could be hard to compare directly):


The Topographic Atomize & Agglomerate Hierarchical Clustering (T-AAHC)

The T-AAHC is actually a further refinement of the above AAHC. The main difference being that the T-AAHC does not account for the powers of the maps, but only for their topographies , as the K-Means was doing.

Though the power of maps can be significantly used in many cases, there are other cases where they are not. If a given brain area is activated with only a low power, still, this is an interesting event to consider. Or if you want to re-segment a bunch of template maps merged together, they all have the same GFP so power is of no help in this case.

The T-AAHC takes the best of both the K-Means, by working only topographies, and the AAHC, by being deterministic and fast, without their respective drawbacks.

 

There is actually only one technical difference between the T-AAHC and the AAHC. At one point, both methods have to pick "the worst cluster" before atomizing it and re-distributing its parts to the other clusters:

 

The T-AAHC results are very close to the K-Means ones, see here an example for 15 segments (first one is the T-AAHC, second one is the K-Means):


Temporal post-processings

Temporal smoothing

This was the historical way to remove small segments. We kept it here, but we do prefer the Small maps rejection procedure instead, because its actual behavior is difficult to predict (very parameter dependent).

(You can refresh your memory with all the formulas here )

In the labeling process, when assigning a map for each time point t, a distance d is computed, and the map with the lowest distance d chosen:

To express the idea that micro-states tend to stay stables, we introduce a smoothing factor in the distance d, by defining a new distance d' that has to minimized the same way:

with Nbkt being the number of maps k in the interval t-b to t+b (excluding time t itself), and >0 being a weighting factor to control the smoothness. equal to 0 gives d' = d, while greater values will allow more smoothness. b is a window size, before and after current time point t.

 

 See f.ex. before and after a smoothing (Window half size of 3, strength of 10). Note that not all small segments have been removed:

Sequentialization

The mathematical clustering, as already explained, does not account for the time dimension. It therefore happens very often that maps attributed to a given cluster be in fact spread into very distinct time windows . See for example here cluster 3, or cluster 4:

 

The clustering did a good job grouping maps together, as the best trade-off for 8 clusters. But based on physiological evidences, segments sharing the same label but being temporaly disconnected do not reflect the same neuronal activity . We are then untitled to sequentialize those parts, i.e. to split the clusters with many non-overlapping windows into some sub-clusters.

 

The second generation of the sequentialization process works this way:

 

See here the original clustering without sequentialization (above), then with sequentialization (below). Worth of interest is the splitting of old cluster 1 into new 1 and 9, old cluster 6 into new 7 and 11, old cluster 7 into new 13, 15 and 16. Also note that old cluster 8 remains unchanged (though renamed 14) due to the fact that all its parts overlap:


 

Merging correlated maps

This procedure blindy merges together all maps that are correlated above a specified threshold. To be used when you really want to force (and only when) the merging of some segments together.

Not to be confused with the re-merge step of the Sequentialization. This is another, and independent, processing.

 

See f.ex. before and after merging maps above a correlation of 0.92:


Small segments rejection

Repeatidely scan for the smallest segment, and if it is smaller than a given size, it will be deleted. This is done by splitting it into 2 parts, each part being the one that correlates the most to one the two neighbor segments. If there is only one neighbor, it is simply merged into it.

The size is given by the user, and shouldn't be greater than the smallest segment you logically expect from your experiment! (paradigm, physiology and sampling frequency dependent, at least)

 See f.ex. before and after a rejection of segments shorter than 3 time frames (the coloring changed between the two):

The infamous "Optimal number of clusters"
and the Meta-Criterion

Forewords

A point worth mentioning is that we said at the beginning that we were looking for 5 clusters. But actually, we have no a priori knowledge of how many clusters are right. So we have to run the segmentation sequentially, and independently, through a range of clusters, for example from 1 to 20 clusters. Each of these segmentations is perfectly valid in itself, but we wish to have a way to compare between them and select the "optimal one".

 

Many criteria exist, and have been developped based on different assumptions. However, it is also unfortunate that most of them are useless for our case. Indeed, a lot of criteria are not well suited to our type of data (normalized maps, equivalent to using a cosine distance). And if they do, they can still give unreliable results. All in all, on the 20 criteria that have been implemented in Cartool, only a subset of 9 criteria (since 2024) seem to be robust enough . But none of them does work accurately in all cases. So the approach taken in Cartool is to run all of them, and to merge their results into a compound called the Meta-Criterion .

 

These criteria and meta-criterion were discussed in this article:

Bréchet, Brunet, Birot, Gruetter, Michel, Jorge - "Capturing the spatiotemporal dynamics of self-generated, task-initiated thoughts with EEG and fMRI" - NeuroImage, July 2019

Note 1: the conclusions above are the results of 20+ years of working on this very specific subject. The selection of the best criteria and the formula of the Meta-Criterion were both the results of empirical observations, but, most importantly, of running simulations. These simulations where used to produce series of quasi-stable micro-states, of random durations, with a range of 3 to 15 seed maps, each correlated in a range of 10% to 90%, and Gaussian noise variance ranging from 50 dB to 0 dB (as much noise as data!). Only the criteria surviving these whole ranges of parameters were kept.

The 9 criteria

The best 9 criteria that proved to be the more reliable for our type of data are:

  1. Dunn, Robust version: An evaluation of the goodness of separation of all clusters, with a robust version formula.
  2. Gamma: An adaptation of Goodman and Kruskal, based on concordant vs. discordant clustered pairs.
  3. Gamma, second derivative of Robust version: Cartool robust version of the Gamma, then using the second derivative.
  4. Krzanowski-Lai Index: A ratio of the relative difference of the withinclusters dispersion.
  5. Krzanowski-Lai Index, Cartool version: Cartool implementation of the internal second derivative formula.
  6. Point-Biserial: A point-biserial correlation calculated between the distance matrix and a binary cluster index.
  7. Point-Biserial, second derivative of Robust version: Cartool robust version of the Point-Biserial, then using the second derivative.
  8. Silhouettes: Evaluation of the consistency of each cluster through its goodness of fit.
  9. Silhouettes second derivative: Second derivative of Silhouettes.

 

Naming convention for criteria listed in the .error.data files is the following:

 

For example, current criteria names are: DunnR, Gamma , GammaR'', KL, KLC, PtBiserial, PtBiserialR'', Silhouettes and Silhouettes''.

 

Here are the main references for all of these criteria:

Charrad, Ghazzali, Boiteau, Niknafs - "NbClust an R package for determining the relevant number of clusters in a data set" - Journal of Statistical Software , 2014

Krzanowski, Lai - "A Criterion for Determining the Number of Groups in a Data Set Using Sum-of-Squares Clustering" - International Biometric Society , 1988

Milligan, Cooper - "An examination of procedures for determining the number of clusters in a data set" - Psychometrika, 1985

Note 2: We used to use a Cross-Validation (CV) criterion for a long time. However, the extensive testing above clearly showed that it was not a reliable criterion. It was "undershooting" all the time, having a tendency to report fewer clusters than the real numbers. This could have induced previous analysis to under-estimate the actual number of clusters, especially for Resting States analysis. The CV was therefore removed from our VIP Club of Best Criteria - RIP.

 

Note 3: All criteria have been implemented to have the max value indicating the optimal clustering. Some criteria work the other way round, with the min value pointing at the optimal clustering. In these cases, results were simply inverted to also work with the max value.

The Meta-Criterion

The meta-criterion is simply defined as the median of all optimal numbers of clusters across all criteria :

 

 

Here is an example of the 9 criteria (top tracks), their geometrical mean, and the Meta-Criterion (bottom track). The horizontal axis being the number of clusters, the red squares show the maximum position of a given criterion:

 

Lowest number of clusters

The minimum number of clusters that Cartool can return is actually 4. We are going to detail the reasons why:

 

Note that there are no constraints on the maximum number of clusters, though (apart to be less or equal to the number of data points). In case of L-corner or second-derivative criteria, Cartool will silently expand the range of clusters by 1 or 2 to be able to compute these criteria. But they will not be part of the output files.

Afterwords

The optimal number of clusters is a difficult topic. We hope to provide a more reliable tool with the Meta-Criterion to guide the researcher. By using different criteria, each being selected for their appropriateness, and combining them together into a Meta-Criterion, we have a more reliable estimator of the optimal number of clusters.

 

For Resting States analysis, we recommand the automatic use of the Meta-Criterion when working in batches at the subject's level, i.e. finding the optimal number of clusters per subject. Then, on the group level, we suggest to use the Meta-Criterion more as a guidance, from which you have to interpret its suggestions. Indeed, it happens the result is really clear-cut, with a single max position of the Meta-Criterion. Easy case, no needs for second thoughts. However, other cases can show 2 different local max positions, suggesting that the data can legitimately be split with either of these 2 numbers of clusters. In this case, the type of data, the paradigm, or even experience should guide your final choice.

How to run the Segmentation

There are two ways to call the segmentation dialog:

Then a dialog in 2 parts appears:

Segmentation Files Dialog

segmentation files dialog

 

Files Presets

Pick the main type of processing you want to do:

  • ERP segmentation
  • Resting States segmentation, on subjects' level
  • Resting States segmentation, on group level

(1) Segmenting Groups of Files

Read some important informations about the meaning of Groups in the Segmentation context.

You can use the very convenient Drag & Drop feature here.

Number of groups:

A counter of how many groups have been given.

Add New Group of Files

Enter a new group of file(s).

Remove Last Group

Does what it says.

Clear All Groups

Clear out all the groups at once.

Read Lists from File

You can direclty retrieve the lists of groups previously (see below).

See also Drag & Drop.

Write Lists to File

You can save the lists of current groups into a file, in case you want to re-use them (much recommended!).

See the file formats available.

Sort Files within Lists

A strange behavior of Windows is to not respect the order of the files dropped in a window. To help cure this silly habit, you can sort all the file names of all the groups already entered.

(2) Epochs

Specify which time period to segment.

No Epochs

The whole input files will be used for the clustering.

List of Epochs:

Specify a list of epochs, each epoch being clustered independently from the others . In this very case, epochs overlapping or not is irrelevant.

From

From time frame (included)...

To

...to time frame (included).

Add Epoch

Actually inserting the specified epoch to the list. Don't forget to press on this button to enter the current values!

If the list remains empty, the segmention process will be run on the whole dataset.

Remove Epoch

Popping out the last epoch from the list back to the edit fields.

Resampled epochs:

This will resample your dataset, generating a set of epochs from random data points. The intended use is for Resting States.

The first time you select this option, Cartool will estimate the parameters below. The aim is to statistically cover 95% of your data, but you can change that, especially if the number of resampling is too high.

See this paragraph about resampling.

XX Epochs,

The number of random epochs, each epoch being clustered on its own.

You can change the number of epochs, and see the corresponding % of data coverage in real-time you'll get.

of XXXX [TF],

The epoch size. This amount of random time points are concatenated into each epoch.

It might be interesting to you (yes, you) to know that the temporal ordering is preserved. Data points, although randoms, appeared in the epoch's order.

You can change the epoch size, and see the corresponding % of data coverage in real-time you'll get.

Covering XX% Data

Combining the number of epochs by the epoch size, it is possible to compute the probability to resample any given time point. Therfore, we have an idea of the data coverage.

For optimal results, it is of course better to cover most of your data. However, if the data were to be totally recurrent, you might ease up this number.

You can change the % of data coverage, and see the number of epochs needed for that in real-time.

(3) Files Options

 

Output Base File Name:

Specify here a basis for all the file names that will be generated during the segmentation process.

Segment Files (.seg)

A visual rendition of the segmentation, which also includes the GEV, GFP, and correlation of templates with the data.

See the .seg specification.

Template Files (.ep)

The Templates (or centroids) of the clusters, a weighted avarage of each clusters.

See the .ep specification.

Data per Cluster

Original data splitted by clusters.

Currently in .sef format.

Synthetic EEG Files

Counterparts of each original file, where the data has been replaced by the corresponding template at each time frame, so it is easy to visualize the temporal evolution of multiple conditions.

The power of the synthetic maps also follow the power of the original data, too.

Currently in .sef format.

Resting States:

The following options are aimed at improving the Resting States batch processing.

Force Single Files Processing

Each file from each input group will be processed individually .

So basically, it doesn't care for conditions, and you can Drag&Drop all your files at once.

Common Best Clustering Directory

After each file has been processed, be it the whole subject or some resampled epoch, the main results are copied to a Best Clustering directory .

Files copied are the .error.data, the optimal templates file , and the corresponding verbose .vrb file. Directoy name ends is  <base file name>.Best Clustering

A useful trick is that you can run multiple Cartools in parallel, each with their own subset of subjects. You can then set the exact same output base file name in all of them, and select the Common Best Clustering Directory. These Cartool instances will therfore write concurrently to the same output directory. They will not overwrite each others results, though, as file names should be unique across subjects. This way you will divide the total processing time by the number of Cartool you run...

+ Delete Individual Directories

This option is only available when the previous one is selected.

If selected, once the optimal results have been copied to the Common Directory, the current file's results will be deleted. At a subject's level, we usually don't care for all the non-optimal clustering results. As these could amount to a lot of disk space, it is totally OK to dispose of them.

<< Previous  |  Next >>

Use these buttons to navigate through the previous and next dialogs (if any).

See which current dialog you are in, and to which other dialogs you connect, in the tab-like part at the top of the dialog under the title.

Process

Runs the segmentation.

This button remains disabled until all the parameter dialogs have received enough (and consistent) informations .

Cancel

Quit the dialog.

Help

Launch the Help to the right page (should be here...).

Segmentation Parameters Dialog

segmentation parameters dialog

 

(1) Clustering Parameters

Corresponds to Part 1 of the segmentation.

Computation Presets:

Pick your choice from the drop-down list, according to your data type and experiment.

This will set the input data type for you (can not be manually changed) and the segmentation parameters with the most common values (which you can manually tune-up).

(1) Data Preprocessing

All sorts of preprocessing to tenderize your data before the data munching.

Spatial Filter, using XYZ file:

Applying a Spatial Filter to the data, to remove outliers and smooth out the noise .

K-Means clustering doesn't like noise, lower SNR data in input also leads to lower SNR results...

However, it is not recommended to apply this filter more than once! If your EEG preprocessing pipe-line already included the Spatial Filter, you shouldn't use it here! Cartool tries to be smart here, if your files contain the characters "Spatial" in their names, it will automatically disable this option.

Using Whole Data

Using all of the data specifed from the epochs options. So either the whole file, some specific time interval, or some resampled epoch. But the file given to the clustering will be used as a whole.

Mainly used for ERPs analysis.

Using only GFP Peaks Data

Keeping only the data at local GFP Max positions. These are the data points with the higher SNR, and centered on the segments' most stable part. Data points with lower GFP are usually transitions due to dipoles oscillations, when not at their peak values.

This is used only for Resting States analysis.

Note that the GFP Max extraction part is done before the resampling . All GFP Maxes are first concatenated into a temp file, which is then resampled. Resampling disrupt the time line consistency, therefore a local GFP Max can not be correctly retrived. This also ensures that all the resampled epochs will have the same size.

Automatic

The GFP track is computed automatically from the input file, then the GFP Maxes positions extracted.

At Markers:

The GFP Maxes might have already been computed, or the user has some specific requirements about what data points to keep. With this option, you can specify whatever marker names for the time positions to be kept.

But Excluding Bad Epochs:

Select this option to skip some time periods from your data. Data preprocessing is never perfect, and some parts of the data should still be ignored.

Note that this step is done before the GFP Max and the resampling.

Automatic

Cartool has a bad epochs automatic detection (you might want to give it a try from the EEG Window, under Markers|Scanning Bad Markers menu).

This option is seen as a last chance before the clustering. It would definitely be better to run the detection before, then visually assess the bad epochs, then only run the clustering.

At Markers

Give the marker names of the bad epochs to be ignored.

Either somebody did put some markers manually to isolate the bad epochs, or have already run the automatic bad epochs detection.

(2) Clustering Parameters

 

Data Type:

This is the input data type, set by the Presets list above.

Only Positive

Data consist of positive only, scalar data. This could be spikes from neuron recordings, or the Results of Inverse Solution, f.ex.

This will logically turn off the Polarity & References options.

See this point on positive data and also this point.

Signed

Signed scalar values, like, you know, EEG.

Vectorial

Used on Inverse Solution  .ris data files, when the results are vectors for each solution points.

See this point about segmentation in the Inverse Solution space .

Data reference

 

No Reference

Data are used as they come from files, no changes occur.

Average Reference

Data are average reference-d.

Maps / Patterns Polarity:

 

Ignore

Polarity of maps does not matter, so ignore it. Inverted maps are considered the same (same underlying generators, but with reversed polarity).

Used for spontaneous EEG recordings.

Account

Polarity of maps matter, that is, inverted maps are indeed considered as different.

Used for ERPs.

Clustering Method:  

K-Means

See K-Means clustering. 

Number of Random Trials

Number of time that the process of randomly picking maps is initiated ( steps 1 to 4 ). Of course, higher values increase the chance to converge to the optimal solution, but at the cost of more processing time.

T-AAHC Hierarchical Clustering

See Topographic Atomize & Agglomerate Hierarchical Clustering .

Range of Clusters

 

From

Minimum number of clusters, keep it to 1.

To

Maximum number of clusters, usually around 20 for ERPs, and 12..15 for Resting States.

You might want to increase that number if, f.ex., your ERP epoch is longer than ~500 ms. Or if the optimal max appears quite close to the higher limit. Of course, the higher the maximum number of clusters, the longer the processing time.

Decreasing this number is a bit more delicate. By doing so you might miss the optimal clustering, without any clue about the miss! The range suggested by Cartool is based on our experience at the FBMLab, you have to have some good reasons to decrease that range.

Labeling at Low Correlations:
No Labeling if Below

If selected, you can specify a minimum correlation threshold for data points to be assigned to a given cluster.

Default correlation threshold is set to 50%, which is quite conservative.

See this outliers rejection paragraph.

Alternative Templates:
EEG To Inverse / Inverse To EEG

This option can help you computing the inverse solution of your EEG template maps , or the EEG maps from your inverse templates if you run the clustering on RIS files.

There are more explanations here.

(3) Temporal Postprocessing

Part 2 of the segmentation.

Sequentializing Segments

Scans the clusters, and split those who belong to the same cluster but don't overlap in time.

Merging Correlated Segments

Blindly merge together clusters that are correlated above a threshold.

This is quite a harsh post-processing, use with care.

If Correlated above:

Specify the correlation threshold as a percentage.

Segments Temporal Smoothing

Redo the labeling, but with a temporal smoothing factor .

Window Half Size:

The temporal smoothing operates with a sliding window, which you specify here the half size (in time frames). The actual window size is therefore ( 2 x W + 1 ).

Strength (Besag Factor):

Strength of the smoothing, actually the Besag factor from the article on the segmentation process.

Rejecting Small Segments

Deletes short segments, and fuse them with their 2 respective neighbors.

Shorter than or equal to:

Size below which a segment is removed.

<< Previous  |  Next >>

Use these buttons to navigate through the previous and next dialogs (if any).

See which current dialog you are in, and to which other dialogs you connect, in the tab-like part at the top of the dialog under the title.

Process

Runs the segmentation.

This button remains disabled until all the parameter dialogs have received enough (and consistent) informations .

Cancel

Quit the dialog.

Help

Launch the Help to the right page (should be here...).

Segmentation - Technical points & hints

Groups of Files in the Segmentation

Each group of files provided in the list will be processed independently of the other groups . Conversely, all the files within a group are concurrently segmented together (a different approach from the Fitting).

In this way, we have a sort of batch processing, where you can input groups with different number of files, and / or different dimensions (in time, in electrodes). However, the same segmentation parameters will apply for all these segmentations. The output directories / files will be numbered if more than one group is given.

You can Drag & Drop these files directly from the Explorer:

It is strongly recommended to use these Drag & Drop features which will tremendously ease your work:

File formats to save or retrieve the lists of groups

Maps

In all the segmentation and fitting algorithms, normalized maps are used as only the shape/topology of the maps is of interest, and not their strengths. The only exception is when generating an average map to represent a given segment, the original maps are taken to therefor make a GFP-weighted sum, giving more emphasize to maps with higher GFP (which are likely to be less noisy and of higher interest).

Resampling your data - why should I do that?

In the Files Dialog, you can set some resampling option, aimed at Resting States processing. Here are the main points why you should favor the resampling (apart from looking cool, of course):

 

TL;DR: Shorter files work better, repeating the K-Means multiple times is better, using different random parts of the data improves both the K-Means and the optimal number of clusters.

 

I hope I made a case to use this resampling technique. Resting States analysis is not an easy task, and the results appear to be much more unstable than ERPs. This is why we should put more chances on our sides to find the optimal clustering. You can still ignore it, though, f.ex. if you want to reproduce some previous results where it was not available.

Global Explained Variance computation

During the segmentatioin process, the GEV value is computed across all the provided files, as if they were aggregated into a single one . It gives a global quality measure of the overall segmentation process, plus the relative weight/contribution of each segment into the global segmentation.

However, during the fitting process, the GEV is computed on a set of groups that represent some natural within-subjects size , which in turns depend on your paradigm. It can even be computed on a per-file basis , if subjects were not the same across all groups.

This has roughly the same consequence as for GFP Normalization , in that the denominator part of the equation will be different according to the number of files it is computed across. Not that the absolute sum is of any importance, but rather that individually scaling each file could introduce some artifacts if some big component was to be found in one file and not into the others. Which could then make the other components appear relatively weaker in the other files, thus wrongly generating significant statistical differences.

Outliers rejection

By construct, and no matter what, the K-Means clustering will assign every data point to one cluster, the one it correlates the most. But sometimes, some data points will not correlate well with any of these clusters. To avoid degrading the existing clusters with outliers, data points whose best correlation are still below a given threshold will not be assigned to any cluster .

Data points too far from any clusters' cloud will be labeled as unclassified . It can also be seen as a hidden, additional "garbage" cluster used to gather all of the outliers. Cluster that will be subsequently ignored for the remaining computations (centroids, criteria...).

The default correlation threshold is set to 50%, which is quite a conservative value. Hence, no more than a few % of the data should be classified as unlabeled in the end. Check with the verbose file for the amount of unlabeled data, and if too high, then something might be wrong with the data. If you increase the threshold, you might end up with too much rejection. If you decrease it, you might incorporate outliers back to the clusters.

Alternative templates

(The following explanations are for the regular EEG clustering)

Cartool provides an option, the "Alternative templates" , to compute the corresponding localization for each EEG template map.

First of all, it expects that all of your input EEG files have some RIS files counterpart, f.ex. as generated by the RIS Computation Toolbox . Cartool expects an exact one-to-one match between each EEG and RIS file names (similar file names, identical directories). Cartool will complain if some files appear to be missing. However it will not whine at all in case no counterpart files were found at all , it will just skip this option silently.

The clustering will then run as usual at the EEG level, without any difference. The final templates are computed as the average of each clusters' cloud, as usual. Then, the labeling from the EEG clustering will be applied on the alternative dataset , allowing to compute the average of the corresponding inverse solutions. This way, each EEG template map will have its corresponding inverse template.

 

Here is a simplified diagram, showing 18 EEG maps clustered into 3 groups. Top part shows the results of the regular EEG clustering, with the 3 clusters containing each 6 maps, and below them their average maps. Bottom part is the alternative dataset, here the inverse space. It is using the same labeling ("map numbers") to compute the average brain activities corresponding to the average EEG maps :

 

 

Note 1: This option is actually working both way. If you run the clustering on the RIS data, then the alternative dataset will be the EEG maps. The main templates will be brain regions, the alternative templates the corresponding EEG maps.

Note 2: You can also compute the average inverses through the Fitting, by saving all data per clusters. Then running the RIS Computation Toolbox with the proper Preset.

Segmentation - Results

(See also the Segments display)

Segmentation in the Source Space

Instead of using EEG data, it is possible to directly segment the Results of Inverse Solution . Instead of the electrical values at each electrode, it will use the estimated dipoles at each solution points within the brain.

This part has to be totally rewritten due to current evaluations...