Choice of regularization parameter in sLORETA

Dear all,

I am working on some EEG data, and when computing the sources with sLORETA I obtain a generalized high activity at the deepest sources, which I suppose being driven by noise. This is reduced when using the MNE approach, which is known to be biased toward superficial sources.

I noticed that this phenomenon might be attenuated by changing the regularization parameter. Comparing values in the range from 3 to 0.5, smaller parameters reduces the (supposedly noisy) signal. This works to a certain lower limit, at which the overall signal is canceled.

It actually seems that reducing the regularization parameter makes the localization smoother, leading to more distributed activity.

Is this a sensible approach for selecting an appropriate value?

Many thanks

@Sylvain @John_Mosher : Any advice ?

(In general, if you are not too confident in the manipulation of these parameters, we recommend you keep them to their default values - these are advanced parameters for expert users with good knowledge of these inverse modeling methods)

Thank you for your recommendation.

The reason why I am testing different values is to check whether the likely noisy activity at deep sources might be somehow attenuated. I may eventually opt for normalizing the power values though.

It is also important to keep an eye on the fit to the sensor data. Indeed, when regularization is too strong, the source maps will look less noisy (smoother, typically) but the fit to sensor data will be poor. You can control these aspects with Brainstorm as you try different values for the regularization parameter. If the map changes drastically with close values, it is a bad sign that the source model is unstable. This should not happen with minimum-norm -like models such as sLORETA, dSPM and MN.

Many thanks.

By stronger regularization do you mean lower values of the parameter?

As far as I could understand, the regularization parameter weights the constraint (the functional) to the ill-posed minimization problem of source localization.

With my data the maps do not change drastically between close regularization values, which is a good sign. For very low values of the parameter (~0.5) I obtain instead smoother maps, i.e. more distributed power activation for the same sources. The sources look more focused for greater parameter, but the aforementioned noise is also emphasized. Would you generally suggest to opt for a compromise value or just trust the recommended value in the tutorial (= 1 for single trial analysis)?



1 Like

I understand, many thanks for your help!