Before reading this tutorial

The tutorial uses an example dataset. You can download it here. When you have the example dataset proceed as follows...

Tutorial for using BDCA with EEGLAB

Step 1. Epoch your data

This is standard EEGLAB stuff. If you load the dataset in EEGLAB and scroll through it, you will see a bunch of events. We'll use those to epoch the dataset. You can do it from the EEGLAB menus, or simply run the following from the command line:

EEG = pop_epoch( EEG, {  '80'  '160'  }, [-0.1         0.5], 'newname', 'BDCA tutorial dataset epochs', 'epochinfo', 'yes');

So, now we have epochs available.

EEGLAB BUG: Note that if you set 0 instead of -0.1 above then EEGLAB will make a sometimes/somewhat useless event structure. Therefore, always include some samples preceeding the event, e.g. -0.1 instead of 0.

Step 2. Make BDCA know which labels to use for what

Ok, so we want to train a classifier which discriminates between epochs of label 160 and epochs of label 80. Now we use the BDCA toolbox:

EEG = pop_bdca_event2labels(EEG,160,80);

This tells the classifier that you want to discriminate epochs of type 160 against epochs of type 80. What happened is that a vector is now to be found in EEG.bdca.labels ! take a look at it if you want:

figure, plot(EEG.bdca.labels);

Note: if you wanted to discriminate, say 160 AND 170 versus 80 then you could do e.g.

EEG = pop_bdca_event2labels(EEG,{160 170},80);

... but we don't need that for this tutorial.

You can manually set the labels by changing EEG.bdca.labels. The function pop_bdca_mat2labels is meant to help you with that.

Step 3. Tuning sigma / scaling the data

We're gonna tune sigma using 5-fold cross validation. Our initial guess is sigma=0.1, and also let's set R=1 because it is much faster than R>1 and it is sufficient for tuning sigma:

R=1;
sigma=0.1;
EEG = pop_bdca_train(EEG,R,sigma,[],[],[],[],[],5);

when it is done, you can see the performance ROC area in the variable EEG.bdca.cvstats.Az. The cross validated log-likelihood is also there in EEG.bdca.cvstats.loglik. So you can tune your model looking at either one.

So, now you could imagine tuning this by hand, or making a for loop and choosing sigma based on the Az or loglik. Note that sigma should be sweeped on a logarithmic scale, say

happy = [];
for sigma=logspace(-2,0,30)
  EEG = pop_bdca_train(EEG,R,sigma,[],[],[],[],[],5);
  happy = [happy; sigma,EEG.bdca.cvstats.Az];
end

and then choosing the sigma which has the highest Az in the result happy. Or, like I said, use the log-likelihood.

Due to numerical issues, you might wanna scale your data so that the optimal sigma is roughly 1. Try for instance to scale your data by the reciprocal of the average channel standard deviation.

Step 4. Tuning temporal smoothness

Now the time has come to start tuning the temporal smoothness. I recommend setting nut=2.5 but you might find a different value to love. Anyway, tune lt...

lt=30;  % this means that we expect 30 milliseconds smoothness
nut=2.5;
EEG = pop_bdca_train(EEG,R,sigma,lt,nut,[],[],[],5);

and, again, look at EEG.bdca.cvstats.Az or loglik. As before with sigma, make a for-loop for finding the best lt.

Step 5. Tuning spatial smoothness

Tuning the spatial smoothness requires that you have channel locations for all channels and that no two channels have the same coordinates. (I'm just saying so that you don't get tempted and simply copy locations from other electrodes to channels that have no coordinates).

ls=0.1;  % this is relative to the EEGLAB head radius which I believe is 0.5
nus=100; % 100 means: no ripple. Decrease this number to allow more ripply topographys
EEG = pop_bdca_train(EEG,R,sigma,lt,nut,ls,nus,[],5);

So, tune ls and nus using cvstats as before.

Step 6. Increasing the number of components

Finally, try increasing R to see the effect of having more components. Note that it will start to go pretty slow.

Step 7. ... much more to come ...

Tutorial for using BDCA without EEGLAB

Tips on tuning the parameters with bilinlogistregmulti

First, set R to 1

>> R = 1

Set sigw0 to some high number, like 5

>> sigw0 = 5;

Then, tune sigma for space and time

>> gpa = [sigma];
>> gpb = [sigma];
>> spaceunits = [];
>> [w0,a,b] = bilinlogistregmultigp(X,y,R,sigw0,gpa,gpb,spaceunits);

repeat the above step until you find a good sigma (use cross validation!). Then, start tuning the temporal smoothness. I recommend setting nut to 2.5, and then try different values for lt.

>> nut = 2.5;
>> gpb = [sigma lt nut];
>> [w0,a,b] = bilinlogistregmultigp(X,y,R,sigw0,gpa,gpb,spaceunits);

repeat the above step until you find a good lt. If you have channel locations, then you can tune the spatial smoothness. First, you have to set spaceunits non-empty. Then start tuning nus and ls.

>> gpa = [sigma ls nus];

Finally, you can start to increase R in increments of 1.