Device Independent Signal Maxwell Filtering

💬 Support & Discussions
1

How can I obtain the device independent signal from SSS as desribed in Presentation of electromagnetic multichannel data: The signal space separation method - Section V Applications of SSS - Subsection B. Virtual signals and movement correction.

To be clear, I am not interested in movement correction, since the subjects head position is not monitored over time. (Head position and size is only roughly given by NAS, LPA, RPA and 2 points at the forehead.)
What I am interested in is the question: How can I compute equation 39 in the paper:

\phi_v = S_{v,in}\hat{x}_{in}

If I understand the documentation for maxwell_filter correctly then the method returns device dependent signals as given by eq. 38

\hat{\phi}_{in} = S_{in} \hat{x}_{in}

Background: I want to classify subjects into 3 classes. MEG data was recorded on 2 sites. One class (lets call it y3) was almost completly recorded on one site and the feature extraction/ classifier (CNN) is able to perfectly distinguish the recording sites which means that the classifier might not learn the typcial features for class y3.

2

Not your fault, it looks like we don’t have MathJax support enabled. I’ll see if I can fix it.

1 Like
3

I’ve enabled MathJax and fixed @Marc’s post to render correctly.

2 Likes
4

Any reason you don’t want to operate on \hat{x}_{in} directly?

Either way, S_{in} (and its pseudoinverse, which is used to estimate \hat{x}_{in}) can be obtained with the use of some private functions in mne.preprocessing.maxwell, but it would be good to perhaps make a version public so you could actually have a stable API.

Can you try this PR to see if it works for you?

5

Theoretically \hat{x}_{in} has same information as \hat{\phi}_{in} but the current classifier was developed for sensor signals and I don’t know how it will perform on multipole moments.

Thank you for the answer. I will try it out.

6

Here are my results.

sensortosensor
sensortosensor1500×1027 225 KB

In order to have a comparison my supervisor recommended to look for other methods which allow the transformation of MEG data from one device to another. I found Knösche 2002 but that was before SSS was published. Does MNE provide any methods for transforming data between similar MEG devices?
Fieldtrip has a method called ft_apply_montage for synthetic gradients – whatever that is.

For future readers this is the algorithm without respecting bad channels:

SA,pSA,reg_momentsA,n_use_inA = mne.preprocessing.maxwell.info_maxwell_basis(rawA.info,origin=origin,regularize=None)
SB,pSB,reg_momentsB,n_use_inB = mne.preprocessing.maxwell.info_maxwell_basis(rawB.info,origin=origin,regularize=None)
# assert reg_momentsB==reg_momentsA
# assert n_use_inA==n_use_inB 
dataBtoA = SA[:,:n_use_in] @ pSB[:n_use_in] @ rawB.get_data()[:,any_sample]
7

I’m not sure if it’s the same algorithm (can’t remember the details and don’t know that 2002 paper) but you can use _map_meg_or_eeg_channels but keep in mind the API is not stable or guaranteed because it’s currently private:

If you want to use this code for example in a publication we should come up with some public API for it.