Evidence Accumulation in Value-Based decisions

OpenNeuro curator note: This dataset was previously accessible at ds001219. The dataset was reuploaded due to privacy considerations.

EEG_data files contain EEG preprocessed data for each subject,session.
EEG_events contain two cells of relevant events for the two sessions of each subject

EEG_data
Y : [number of electrodes x number_of_times double] => EEG activity for all electrodes and all times

EEG_events fields for each cell

respt: [1xnumber_of_trials double] => response onset times (in ms)

rt: [1xnumber_of_trials double] => reaction times (ms)

tstim: [1xnumber_of_trials double] => stimulus onset times (in ms)

resptexcl: [1x13 double] => excluded response times

tstimexcl: 5.8806e+05 => excluded stimulus onset times

diffV: [1xnumber_of_trials double]=>item rating difference (absolute stimulus difficulty)

corr: [1xnumber_of_trials logical] =>accuracy (1:correct, 0:error)

ratL: [1xnumber_of_trials double]=>rating of item on the left of fixation cross

ratR: [1xnumber_of_trials double]=>rating of item on the right of fixation cross

chooseL: [1xnumber_of_trials logical]=>choosing left? (1:yes, 0:no)

chooseR: [1xnumber_of_trials logical]=>choosing right? (1:yes, 0:no)

t0: constant time to shift fMRI events to align to EEG onset times (see below)

METHODS OF EEG PREPROCESSING
We performed EEG pre-processing offline using MATLAB
(Mathworks, Natick, MA). EEG signals recorded inside an MR scanner are
contaminated with gradient and ballistocardiogram (BCG) artifacts due to
magnetic induction on the EEG leads. We first removed the gradient artifacts.
Specifically, from each functional volume acquisition we subtracted the average
artifact template constructed using the 80 volumes centred on the volume-ofinterest
using in-house MATLAB software. We repeated this process for as
many times as there were functional volumes in our data sets. We subsequently
applied a 10-ms median filter to remove any residual spike artifacts. Next, we
band-pass filtered the data by applying a 0.5-Hz high-pass filter to remove direct
current (DC) drifts and a 40Hz low-pass filter to remove high frequency artifacts
not associated with neurophysiological processes of interest. These filters were
applied together, non-causally to avoid distortions caused by phase delays.
BCG artifacts share frequency content with the EEG and as such are more
challenging to remove. To avoid loss of signal power in the underlying EEG
we adopted a conservative approach and removed a small number of BCG
components using principal component analysis in two steps. Firstly, four BCG
principal components were extracted from data that were initially low-pass filtered
at 4Hz to extract the signal within the frequency range where BCG artifacts are
observed. Secondly, the sensor weightings corresponding to those components
were projected onto the broadband (original) data and subtracted out.

fMRI_data files contain fMRI preprocessed data for each subject,session.

METHODS FOR fMRI PREPROCESSING
We discarded the first ten volumes from each fMRI run to
ensure a steady-state MR signal, and we used the remaining 307 volumes for the
statistical analysis presented in this study. Pre-processing of our data was
performed using the FMRIB�s Software Library (Functional MRI of the Brain,
Oxford, UK) and included: head-related motion correction, slice-timing correction,
high-pass filtering (4100 s), and spatial smoothing (with a Gaussian kernel of
8mm full-width at half maximum). To register our EPI image to standard space,
we first transformed the EPI images into each individual�s high-resolution space
with a linear six-parameter rigid body transformation. We then registered the
image to standard space (Montreal Neurological Institute, MNI) using FMRIB�s
Non-linear Image Registration Tool with a resolution warp of 10 mm. Finally, B0
unwarping was applied to correct for signal loss and geometric distortions due to
B0 field inhomogeneities in the EPI images.

METHODS TO CREATE fMRI REGRESSORS
We performed whole-brain statistical analyses of functional data
using a multilevel approach within the generalized linear model (GLM) framework,
as implemented in FSL through the FEAT module:
Y= Xb + E = b1X1+ b2X2 + b3X3 +b4X4 + E
where Y is the times series of a given voxel comprising T time samples and X is a
Tx4 design matrix with columns representing four different regressors
(see below) convolved with a canonical hemodynamic response function (double-g
function). The regressors times are shifted by the fMRI t0 (the EEG time at which the scanner started) which is
saved in the EEG events files.

b is a 4x1 column vector of regression coefficients and e a Tx1
column vector of residual error terms. We performed a first-level analysis to
analyse each participant�s individual runs, which were then combined using a
second-level analysis (fixed effects). Finally, we used a third-level, mixed-effects
model (FLAME 1) to combine data across subjects, treating participants as a
random effect. Time-series statistical analysis was carried out using FMRIB�s
improved linear model with local autocorrelation correction.

Our GLM model included an EEG-informed regressor capturing the trial-by-trial
dynamics of the process of EA. Specifically, for each trial we used the raw
EEG time-series (from the subject-specific sensor that was most predictive of the
model-derived EA profile) to parametrically modulate the regressor amplitudes.
We considered the entire trial duration (that is, RT) minus the subject-specific nDT
estimated by the model, which accounted for stimulus processing and motor
execution. More specifically, we split this nDT in two intervals by fixing the motor
preparation to 100 ms prior to the response (when a sudden increase in
corticospinal excitability occurs) and setting the average duration of the stimulus
encoding to nDT-100 ms . To absorb the variance associated with other
task-related processes we included three additional regressors: (1) an unmodulated
stick function regressor at the onset of the stimuli, (2) a stick function regressor at
the onset of stimuli that was parametrically modulated by the VD between the
decision alternatives and (3) a stick function regressor aligned at the time of
response and modulated by RT . As a control analysis we also removed the
RT and VD regressors from the GLM design to test if our EEG-informed regressor
absorbed additional activations. The only activation we found in the EEG-informed
regressor was the one capturing accumulation dynamics as in the main analysis
(that is, pMFC) with a marginal improvement in the statistical significance of the
area.