Skip to content
Masters project. Our objective is to characterize a mapping of the neural substrates of social interaction using advanced signal processing and graph/information theory
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


  • Fieldtrip (21092018 version)
  • spm12
  • EEGLAB (eeglab14_1_2b)
  • ICBM152 (T1, 0.5mm)



Remarks: The scripts has three "big" sections: computing the headmodel, electrode allignment with the template, and the actual preprocessing and decomposition of the data. Here are all the steps:

  1. Setup path, dependencies, and important variables
    • Use spm12 because Fieldtrip's default version is not compatible with my MATLAB version
  2. Create headmodel based on ICBM 152 (unbiased standard magnetic resonance imaging template brain volume for normal population; the most commonly used reference brain for neuroimaging)
    • Get MRI and add fiducials to it
    • Segment the volume (brain, skull, scalp)
    • Prepare mesh (triangulated surface mesh for the volume conduction model)
    • Create headmodel (Boundary Element Method)
  3. Electrode allignment with template
    • Two passes: first allign template fiducials with electrode fiducials then interactively make the fit better
    • Two electrodes of one subject were digitized incorrectly and corrected by taking those same coordinates from a glass head (g64cap.sfp)
    • The MNI coordinates for the fiducials (Nasion, Right pre-auricular point and left pre-auricular point) are taken from: Cutini S, Scatturin P, Zorzi M (2011): A new method based on ICBM152 head surface for probe placement in multichannel fNIRS
  4. Load data
  5. High pass filter @ 0.5 Hz (Hamming FIR)
  6. Load channel location file
  7. Trim data from third event to end of data (only analyze data when participants were actually playing)
    • This quantities were obtained from the MIDI files and Reaper sessions
  8. Take out line noise using spectral regression
  9. Run clean_rawdata
    • Channel rejection: channels with flat signal for over 5 seconds and (2) channels that are poorly correlated (r < 0.75) with adjacent channels (also channels with activity 8 standard deviations away from the whole mean are rejected)
    • Run ASR, which removes short-time high-amplitude artifacts in the cont data
  10. Spherical interpolation
  11. Common Average Reference
  12. Output EEG lab file and save it
  13. Process baseline using same steps
  14. Prepare the EEG file
  15. EEG to Fieldtrip structure (min duration for covariance matrix and one with full duration trials)
  16. Prepare the leadfield model (forward model for multiple dipole locations)
  17. Compute cortical patches basis functions (forward model)
    • We use a coarse version of the AAL atlas (19 regions)
    • Get the leadfield for each patch (n_channels x (3*points in the patch))
    • We assume Gamma = I, white noise
    • Do SVD of leadfield in that patch, and take n singular values that will give you a gamma representation accuracy of 0.85 (this respresentes a trade off between representation error and resolution)
  18. LCMV criterion spatial filters for each patch
    • Get basis for each patch
    • take the smallest eigenvalue of Yk to maximize power output (because we do not know the moment of the dipole)
    • compute each patch spatial filter weights
      • covariance matrix:
        • only take out electrodes interpolated in all trials
        • trial level covariance matrix (cut trials to min duration of the trials)
        • average cov matrices together
      • minimize the output power subject to a unit response constraint to patch location
  19. Get patches time series
    • We are using an unconstrained model; we maximize power output by taking the smalles eigenvalue of Yk
  20. Plot average power of each patch on MRI scan
  21. Plot patch resolution on anatomical image
    • Basic idea: ideally, this would be 1, but this is impossible to achieve on practice because we must balance representation error against resolution


Remarks This folder includes all the necessary files and scripts to run the STE analysis in a Compute Canada cluster (in this case, Graham). For details of the bundle Serial farm implementation, you can check out this video. Note this is called "fixed" because we had an issue with an earlier implementation in the way we were controlling the transfer entropy delays (you can find the older versions in the "old" folder; the things that changed is that we now keep the delay fixed for the target signal and only delay the source signal in the calculation of STE)

  1. Copy all the contents of this file to your Graham home directory

    • You can run this command
    scp -r /file/in/local/computer/ $
    • P0* directories include three files per subject: the patch time series (19 sources), the patch labels, and the trial labels
  2. Once in the server, run

    • This is a bash script that that does the most basic serial farming: sending all the jobs as a loop (and this is all we need)
    • You will need a table.dat file. This file is created by the Python script "". Each line of this file is the command that represents a "case" (or an individual job) in the serial farm.
    • As you can see, each job loads Python, then the SciPy Stack, then runs with certain parameters
  3. The script intakes 7 parameters: Pair, Source frequency, target frequency, subject a, subject b, delay. I wrote it this way to make it easier to create the serial farm on Graham. It does as follows:

    1. It get the parameters from the terminal input
    2. It gets the data from the .mat files (you will need to change the directory here!)
    3. Do the baseline time-frequency decomposition to get the normalizing factor
    4. Do trial time-frequency decomposition
    5. Do computation of symbolic transfer entropy and output a csv per frequency interaction (gamma - gamma; delta - beta; ...; N = 25 per pair per delay)
  4. Once it's done, copy the csv directory to your computer (run this bash code on your computer)

    scp -r $ /file/in/local/computer/


Remarks These files run essentially the same things as 2-STE_fixed_delayed BUT to the baseline files (for statistical purposes--we will compare playing against baseline). The file "baseline_start.csv" is used to synchronize both baselines (in the original trials this is done in the MATLAB script)


Remarks These files run essentially the same things as 2-STE_fixed_delayed BUT to scrambled participants (so we can get a "noise" baseline level of STE). The file "baseline_start.csv" is used to synchronize both baselines (in the original trials this is done in the MATLAB script)


Remarks These files is a Jupyter notebook where all the magic happens:

  1. First calculate descriptive statistics of participants' pairs to answer the question: did we do a good job at pairing them up?
  2. Take mean of all pieces and compare them to (1) baseline and (2) scrambled pair values of STE to answer the question: are the top between brain connections related to neural activity while playing? Or are they just noise?
  3. After this, we explored the possibility of there being differences due to our experimental manipulation in terms of graph theory statistics
  4. Finally, we look at some metrics as a function of time (i.e. small worldiness of hyperbrain networks...)


This folder includes all the scripts with the previous implementation of STE (i.e. delay both target and source signal as opposed to only source signal)

You can’t perform that action at this time.