This project aims to perform permutation testing using MaxT approach (https://doi.org/10.1002/hbm.1058) to control for family-wise errors(FWE).
Dataset: The dataset consists of 36 subject's task activation across 8 runs while performing NEXT behavioral task.
Method:
- Task activations were estimated using regularized ridge regression (https://github.com/alexhuth/ridge).
- Parcel-level functional connectivity was estimated using Graphical Lasso regression (Peterson, 2023).
- Two way networkwise button press decoding was performed using cross-validated pearson correlation-based minimum distance classifier on actual betas to estimate the ground truth.
- A generative multi-step activity flow model was used to create trial-wise predicted activations using FPN as task instruction encoding network and VIS1/VIS2 as GO probe encoding network.
- Two way networkwise button press decoding was performed on predicted betas and was tested for significance.
Hypothesis: The structure of functional connectivity (FC) and task activation enables above-chance decoding of somatomotor network (SMN) activity patterns during task performance. When FC and activation patterns are randomly shuffled, this information structure is disrupted, and decoding performance drops to chance.
- FC was shuffled (1000 permutations) and max t-stat null distribution of accuracy was estimated.
- Trial-wise activation was shuffled (1000 permutations) and max t-stat null distribution of accuracy was estimated.
Result: Both permutations showed non-shuffled decoding on SMN to be significantly different compared to null distribution (p < 0.05), hence confirming the hypothesis.