[FieldTrip] Spatial Bias Dominates Variance in Distributed Source Modeling

[Schoffelen, J.M. (Jan Mathijs)]

Dear Alexander,

I am not sure whether I completely follow your diagnostic steps and your conclusions, but I would also check:

- the alignment between the electrodes, volume conduction model, and source model.
- are the electrode positions expressed in the same coordinate system as the volume conductor/source model?
- are the leadfields well-behaved? For instance: if the dipole locations are too close to the innermost mesh, numerical problems may arise.

Best wishes,
Jan-Mathijs


On 12 Oct 2018, at 18:34, Nakhnikian, Alexander <Alexander_Nakhnikian at hms.harvard.edu<mailto:Alexander_Nakhnikian at hms.harvard.edu>> wrote:

Dear All,

Apologies for the long post, I've tried to be as succinct as possible while describing my problem is sufficient detail.

I've been trying to solve a problem with source modeling for sometime. I've found that distributed source estimates (MNE, sLORETA, eLORETA) are heavily biased towards the ventral temporal lobe. This is the case with multiple data sets analyzed using both built-in field trip functions and imaging kernels generated by my own code. It occurs in within group grand averages and statistical contrasts between controls and patients. I've confirmed that sLORETA and eLORETA are unbiased for noiseless data by filtering point sources through the resolution kernel. I'm working on a Mac running OS 10.11.6 and the latest version of Field Trip. I'm using a standard 10/10 electrode layout (no individual sensor locations) with 4 custom locations (PO9/PO10, M1/M2) and Field Trip's template BEM. The forward model is restricted to the cortical mantle (I've had similar problems with whole brain forward models as well).

I recently ran PCA at the source level to explore the issue. The data were collecting during quiet rest and bandpassed to isolate a peak that accounts for a significant difference between controls and patients at the sensor level. The imaging kernel was applied to the sensor spectral matrix. To obtain the PCs, I analyzed the covariance of power among voxels. The rank of the voxel covariance matrix was 31 with the first 3 eigenvalues account for approximately 95% of the variance. Interestingly, the spatial distribution of the first three principal components exhibited the same bias as the sensor data. When I reconstructed the source data omitting the first 3 components, the control-patient contrast localized to a collection of canonical DMN nodes.

It seems extremely odd to me that removing so much of the information present in the original data returns a reasonable source-level contrast while the majority of variance is accounted for by what is clearly bias. I cannot complete this analysis and submit the results unless I can isolate the problem and correct it so I can run the analysis without reducing the rank of the source data. If anyone can speculate on possible reasons for this problem and/or potential solutions I would be grateful.

Thank you,

Alexander

Alexander Nakhnikian, Ph.D.
Research Investigator
VA Boston Healthcare System
Instructor in Psychiatry, Harvard Medical School
_______________________________________________
fieldtrip mailing list
https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
https://doi.org/10.1371/journal.pcbi.1002202

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20181016/07bc6756/attachment-0002.html>