Split-Half Analysis of Motion Associated Networks
Motion impact score detects whether residual in-scanner head motion artifact spuriously inflates or obscures resting-state fMRI trait-connectivity correlations. [preprint]
Clone the repository and start matlab.
$ git clone https://github.com/DosenbachGreene/shaman.git
$ cd shaman
$ matlabGenerate simulated data.
>> cd simulate
>> simulate % writes simulated data to sub*.mat
>> cd ..Compute motion impact score for the simulated fMRI data and trait.
>> % Construct a DataProvider object.
>> % Point to the folder containing the simulated data.
>> data_provider = SimulatedDataProvider("simulate");
>> % Feed the data into the SHAMAN algorithm.
>> % Specify names of traits to analyze.
>> shaman = Shaman(data_provider, ["trait"]);
>> % Run a few quick permutations.
>> shaman.permutations.nperm = 32;
>> % Get a table of false positive motion impact scores.
>> tbl = shaman.get_scores_as_table("score_type", ScoreType.FalsePositive)tbl =
1×2 table
False Positive Score False Positive p Value
____________________ ______________________
trait 103.33 0
Read the API documentation and work through example.m.
>> doc ShamanThe motion impact score and SHAMAN algorithm are described in our preprint, which you can cite if you find this tool useful. Briefly, we assume that the functional connectivity related to a trait of interest (e.g. score on a standardized test) does not vary from second to second. This allows us to:
- Split each participant's timeseries fMRI data into a high-motion and low-motion half.
- Generate a connectivity matrix from each half of each participant's data.
Most regression approaches for functional connectivity include between-participant head motion as a covariate. In the next step, we allow this covariate to mop up some of the residual motion artifact.
- Regress out between-participant differences in head motion.
- Subtract each participant's high-motion matrix from her low-motion matrix.
Following step 4, it is important to realize that the each participant's difference matrix should be zero (or unstructured noise). Absent motion artifact, trait-related functional connectivity in the high- and low-motion matrices should be identical and should therefore cancel out during subtraction. Whatever is left must be due to motion. Thus, finally:
- Regress the trait of interest against the difference matrices to obtain motion impact score.
Please direct correspondence to Benjamin Kay. email | twitter