Hello,
I finally managed to use brainstorm to create a spike like activity and use that to simulate data for EEG.
I used the same data, without even exporting it, to compute the inverse problem (so everything stays the same, (head model, etc.)
and I got the following result for the case of sLORETA! Do you guys think this results in sound right or my spike like simulation wasn't realistic enough? Where do you think is the problem? I tried to make noise and background activity close to zero to exam and see if sLORETA really has zero error in case of no noise ....
Now I understand that by using thresholding we can decrease the get closer answer to ground through but as I increase threshold, I use the second source (second image)
This looks like an acceptable result to me. You give your model 15000 sources to explain your spike, it uses them. Use a single dipole fitting if you want one single point.
Alternatively: use more EEG electrodes, add realistic noise to your recordings (computed from real EEG recordings), compute a min norm normalized with respect to a baseline (before your spike).
Yes, after using Real EEG as noise covariance for the simulation the result got much better.
Can you elaborate more on this point?
I simulated a spike like activity for 50 sources (2 patches of 25) and Noise Covariance from Real EEG, including the only factor of 0.2 for the noise (so very high SNR), 19 Electrodes and I got the following output. This is the best output I could get from sLORETA (setting a threshold by hand that provides the closest solution to the DATA and trying all regularization parameters to see which one gives better output and setting SNR high so the output is less Smooth). Do you think if I increase the number of electrodes it would result in a much better estimation of sLORETA?
I am performing this task so I can make sure that problem is not with my simulation (my simulation is realistic enough) , usually I see in articles that sLORETA have very good estimation with a bit of smoothness, so I think If I can generate the data in a way that gives me good result for sLORETA estimation, my simulation is good enough for evaluation of my method. Any suggestion?
If you keep increasing the amplitude threshold, does the sLORETA solutions peaks at the actual locations of the simulated sources? Also, using more electrodes can certainly help.
up to some thresholding yes it does make the solution better, however, after some threshold, it is going to lose one of the sources. Also, the second bar called min (which I assume put a threshold on the radius of the source rather than the amplitude of data ) would make the solution better as well.
Has there been any article that you are aware of on automating the thresholding and “min distance” values? It might be something useful to work on because I feel like not having automated thresholding may make the validation and comparison of methods unbiased …
