mshalvagal / bci_errps_2019 Goto Github PK

for i=1:size(dirinfo)
if contains(dirinfo(i).name, 'calibration')
path = [dirinfo(i).folder '/' dirinfo(i).name];
[calibration_signal, calibration_header] = sload([path '/' dirinfo(i).name '.gdf']);
calibration_behavior = single(dlmread([path '/' dirinfo(i).name '.txt']));
end
end

calibration_eeg = calibration_signal(:, 1:num_channels);
calibration_eog = calibration_signal(:, 17:19);

% EOG correction
if do_eog_correction
eog_b = filterEOG(calibration_eeg, calibration_eog);
eeg = eeg - eog * eog_b;
end

takes model and online data and outputs the posterior error probability

Add MCC also to the model selection

create the LDA model

Accuracy and bias plots for each training set size (the number of trials used for the training which will be an indication of the calibration time)

(Moving from issue #6 because this is a different task)

Add mean correction to the CCA fct and return mean to be used for testing data centering

1- Take it at multiple time points
2- use the mean over a window
3- (optional) separate 20-40-60 degrees

Create a dropbox so we all have everyone data, use Manu fct to concat them

plot the ErrERP and CorrERP together with their difference —> we should see two two early + and - peaks and two broader + and - peaks

add an option input for the same function

run a cross validation to decide the threshold of error detection as hyperparameter

Start recording performance on a common Google sheet https://docs.google.com/spreadsheets/d/1lYXYrcpfHNmf_E0zMpSmmp3nSiI33lCMQdfRiojoxq0/edit#gid=0

This will help us keep track of our results. Should help for the presentation as well

1. PCA false but still need to do feature extraction
2. Select based on Fisher score

Get other performance metrics like precision, recall, F-measure, etc

• Make a function to train models with different preprocessing steps
• accuracy and bias plots for each training set size(the number of trials used for the training which will be an indication of the calibration time)

If we decide to do CCA in the online decoding, we will need it:)

Probably it would be better to change CCA such that it only take train data and output W. Then in the model assessment we can apply W on test data. Like this we can use the same function in the online decoding too. Because for the online decoding we only need to find W on the train data and there is no test data.

create the LDA model

Move the cross-validation loop to main.m, so that we can evaluate different models with the same splits in case we want to.