P300 Accuracy V Copyright 2012 g.tec medical engineering GmbH

Transcription

1 g.tec medical engineering GmbH Sierningstrasse 14, A-4521 Schiedlberg Austria - Europe Tel.: (43) Fax: (43) office@gtec.at, TTThttp:// P300 Accuracy V Copyright 2012 g.tec medical engineering GmbH

2 Introduction This document describes how to use the P300 Accuracy function of g.bsanalyze. In a P300 experiment usually each stimulus is given several times before a classification is performed. This is done to reduce the error rate of the classifier. The number of needed repetitions for an accurate classification could vary from user to user and also from day to day. It is clear that the classification accuracy increases with a higher number of repetitions, but on the other side the spelling speed decreases. P300 Accuracy displays the accuracy of a given classifier as a function of used repetitions. Hence the output graph gives the experimenter a hint how to choose an optimal number of them for the online experiment. Be aware: The function only works correctly when the data is recorded in the same format as the Simulink model in g.p300 or the intendix software does. This means that the file must contain at first the recorded EEG channels. In the second last channel the flash ID is saved. This is the ID (number) of the actual presented flash (character). The target information is stored in the last channel. It is is set to one if the current flash is a target flash and set to zero if the flash is a nontarget flash. A raw datafile recorded within a P300 experiment Step 1: Create a Classifier Start g.bsanalyze and load the recorded P300 data file. If no dataset is available you can also open the file: C:\Program Files\gtec\gBSanalyze\testdata\P300Accuracy\P300data.mat This testdata is already prepared for use and contains therefore no timestamp. If you take a raw-data file recorded with intendix or g.p300 it contains the timestamp on the first channel. Remove it with the following procedure: (This step is not necessary if you have loaded the file above!) In the Transform menu click on Cut Trials & Channels. Now press on Select trials/chan and in the window that opens now, select the first channel to exclude (see screenshot). Click on OK! to confirm your selection. Press the Start button in the Cut Trials Channels window to finally perform the cutting. P300 Accuracy

3 Remove the timestamp from your raw data Select the first channel for removal Now, the classinfo of the dataset is needed. The classinfo describes which of the recorded trials are target-trials and which ones are nontarget-trials. To read out the classinfo copy and paste the following code into the Matlab command window (For later on you can save the code in a file). P300 Accuracy

4 global P_C; dat = P_C.Data; [NTrials,NSamples,NChannels] = size(dat); triggerdata = dat(:,:,nchannels - 1); targetdata = dat(:,:,nchannels); triggertime = find(diff(triggerdata) > 0); targettime = find(diff(targetdata) > 0); classinfo = zeros(2,length(triggertime)); for NT = 1:length(triggerTime) if isempty(intersect(triggertime(nt),targettime)) classinfo(:,nt) = [1;0]; else classinfo(:,nt) = [0;1]; end end clear dat; save('p300classinfo','classinfo'); This code saves the classinfo in the file P300classinfo.mat. The next step is to filter the data with a bandpass filter. The lower cutoff frequency is 0.1 Hz, the upper cutoff frequency is set to 30 Hz. Click in the menu on Pre-Processing and then on Filter. Check the box at Use MATLAB filtfilt function to filter data. Apply a filter on the data When you perform this calculation for the first time you need to define your filter. Click on Design filter and enter the following settings: Filter type: bandpass (BP) Filter realization: Butterworth Lower cutoff frequency (- 3 db): 0.1 [Hz] Filter order (max 500): 4 Upper cutoff frequency (- 3 db): 30 [Hz] Name: P300 Click on add to list -> to confirm your settings and click on Ok to close this window. Now you can select the P300 filter in the Select filter: window. At last click on Select trials/chan. and P300 Accuracy

5 exclude channels 9 and 10 to apply the filter only on the EEG data. When the filter is selected start filtering by pressing Start!. Set up a 4 th order butterworth filter Now apply a moving average filter onto the EEG data In the Pre-Processing menu select Moving Window Filter. Choose Moving average as filter type. Select 12 [samples] for the length of the window. Press Select trials/chan. and exclude channels 9 and 10. Click on Start! to apply the filter. P300 Accuracy

6 Apply a Moving average filter with the window length of 12 samples on the data Now the data could be triggered. Open the Trigger dialog in the Transform menu and set the following parameters: Time before trigger: 100 ms. Time after trigger: 700 ms. In the pane Specify TRIGGERS and ATTRIBUTES, select Physical channel and choose 10(CH9) from the list. As Threshold voltage choose 0.5 µv. Click on Accept overlap and add to list-> to acknowledge your selections. Click on Select channels and exclude the channels 9 and 10. Only the triggered EEG data remains now for further processing. Apply the moving window filter by pressing Start!. Specify parameters in the Trigger dialog P300 Accuracy

7 To load the previous saved classinfo, click on Load Class Information in the File menu. Press Import Wizard and select the file P300classinfo.mat that was created some steps before and Finish in the window that appears. Then enter the name NONTARGET in the Change name to field and click on the button Assign Attribute to assign the name to the class 3. Now select with the mouse the class 4 and enter TARGET into the Change name to field. Again click on Assign Attribute for changing the name of class 4 into TARGET. Click on OK to finally assign the new classinfo onto all trials. Load and apply the classinfo onto the triggered data Select the saved classinfo file Now perform a baseline correction onto every single trial. The mean value of each trial that appears 100 ms before the flashing event is subtracted from the whole trial. In the menu click on Pre- Processing and Baseline Correction. Select 1 [samples] in the Start at field and 26 [samples] (~100 ms) in the Stop at field. Press Start to apply the correction. P300 Accuracy

8 Apply the baseline correction to the data After that, the preprocessing of the data is finished and the classifier can be calculated. For that a feature matrix is defined and then the classifier is computed with this defined matrix of features. Start with selecting Classification and then Feature Matrix in the menu. Now choose the following parameters Start at: 27 [samples] Step: 12 [samples] Stop at: 205 [samples] Merge time points Select the classes NONTARGET and TARGET. Now click on Start to create the Feature Matrix. Create a feature matrix to train the classifier P300 Accuracy

9 Create and save your classifier The Linear Classifier window appears. Select Train 100 % - Test 0 % in the popup list for Training/test-sets. Check the Save results checkbox and enter the filename P300Classifier to save the file. By clicking on Start the Classifier is generated. P300 Accuracy

10 Step 2: Apply a Classifier on a P300 dataset In g.bsanalyze load your P300 dataset that you want to test, e.g. C:\Program Files\gtec\gBSanalyze\testdata\P300Accuracy\P300data.mat The file needs to contain the same number of EEG channels that were used for creating the classifier. It also must contain the triggers for the highlighted characters on the second last channel and the information which characters were selected as target characters on the last channel (see also the Introduction section). For a raw datafile the timestamp needs therefore to be removed. (Again this step is not necessary if you have loaded the file above). In the menu click on Transform and Cut Trials & Channels. Cut out the first channel by pressing on Select trials/chan. and choose the first channel to exclude. Click again on Start. Prepare you data for testing the classifier P300 Accuracy

11 Select the first channel to be cutted out Apply the same bandpass filter as for creating the classifier (see above) to the data. Now the classifier could be tested. Select in the Analyze menu P300 Accuracy. Click on Browse and select your previous created file P300Classifier. Load the classifier and apply it on the data Click on Start and the g.result2d window will open. P300 Accuracy

12 The g.result2d window shows the classification accuracy The graph shows the obtained accuracy level of the applied classifier as a function of used number of flashes for discrimination. The number of flashes is the quantity each row and each column (or each character for a SC speller) is highlighted before a classification is done. The highest number of flashes for the given dataset was at 15, hence each row and each column flashed fifteen times for each spelled character. For these 15 flashes the accuracy was at 100 %. But also when using less flashes (6 until 15), the classifier was able to reach 100 % of classification accuracy. Only with less than 6 flashes the accuracy decreases. P300 Accuracy

13 P300 Accuracy