Spatiotemporal decomposition specification for PCA

[rciric]

This issue regards the BIDS derivatives specification for spatiotemporal decompositions, in particular storage of metadata for components computed by principal component analysis (PCA).

• For PCA-based decompositions, it would be useful to have metadata specifications for eigenvalues (or singular values) and their derivatives. There was previous discussion about these measures; however, it appears that they might not have made it into the proposal.
  • In the absence of an explicit directive from BIDS, the fmriprep confounds PR implements the following fields (in addition to Method): SingularValue, VarianceExplained, CumulativeVarianceExplained, and Mask (denoting the source mask if more than one mask is used for component extraction).
• Metadata from components that are not ultimately selected for inclusion in the model matrix can nonetheless provide useful diagnostic information about the decomposition. For instance, in the fmriprep confounds PR, these metadata are used to plot the cumulative variance curve. Since discarded components won’t have corresponding column names in the ~regressors.tsv matrix, it would be useful if there were a standard name for such components to be used when entering their metadata in the decomposition JSON.

In any case, the derivatives spec appears to be frozen for now, so perhaps this can be revisited after it re-opens?

[tsalo]

Was this not included in the extension? I see it in the Google Doc but not in #265.

[effigies]

It's in #207 (https://bids-specification.readthedocs.io/en/derivatives/05-derivatives/05-functional-derivatives.html#spatiotemporal-decompositions), postponed until after common derivatives (#265) is released.

[tsalo]

Thank you @effigies! I guess I can circle back once #207 is reopened for input.