Adding a conservative elbow-free decision tree

[handwerkerd]

Summary

This is based on an OHBM2024 discussion with @moser297 that was discussed at our July 2024 developer call. The kappa and rho elbows do not robustly appear in all runs. In particular, an infant dataset seemed to often not generate clear elbows. This is worth investigating in general, but we could also add a conservative super-minimal decision tree that doesn't depend on any elbows or relative values that might cause issues in some datasets.

Additional Detail

The idea would be to make a decision tree wouldn't be as aggressive at removing noise as other trees, but any user could be reasonably confident that it never removes something they care about. My initial idea would be to just include the first 6 nodes of the MEICA Decision Tree ( https://tedana.readthedocs.io/en/stable/included_decision_trees.html ) which just exclude extreme cases, like rho>kappa.

Right now, this is a placeholder issue so that we don't forget this discussion. Anyone is welcome to take this on.

Next Steps

• Make & name a simple decision tree and test it on some data, possibly in collaboration with @moser297
• Check if the tree is too conservative (i.e. it's only removing things that would have already been removed by detrending regressors or other standard tools)
• Check if there are any criteria tweaks that could be used to make it a bit more agressive, but still reliable.

[moser297]

Thank you for tackling this topic! I currently don't have the capacity to be actively involved but I am happy to share some testing data with you if you like.

[handwerkerd]

@moser297 I'm not sure I'll be able to immediately test something, but if you share your data, I'll try to find time soonish. I just emailed your UMN address so that we can coordinate directly. Thank you.