Figure 9. 12 Channel montage used in the study with FPz as reference voI tage and left mastoid as ground.
3. DATA ANALYSIS
3.1. Data pre-processing Methods As each of the five stimuli (mVEP buttons) were a target for 60 trials for each game level, we were able to record a total of 300 trials per level from each subject. Data epochs were derived in association with each motion onset stimulus, beginning 200ms prior to the motion onset and lasting for 1200ms. All single trials were baseline corrected with respect to the mean voltage over the 200ms preceding motion onset. Data were digitally filtered using a low-pass Butterworth filter (order 5, with cut-off at 10Hz) and subsequently resampled at 20Hz. Features were extracted between the lOOms and 500ms epoch post stimulus which normally contains the most reactive mVEP components e.g. N200, P300 and N400. This yields nine features for each channel. Data were averaged over five trials yielding twelve feature vectors per stimulus for each level. Data were initially split into target vs. non-target where for each non-target feature vector five randomly selected nontarget trials were used. mY EP is time locked and phase locked to the motion onset stimulus therefore mVEP induced from the motion stimuli could be obtained through the above simple averaging procedure [15].
3.2. Channel Selection A Linear Discriminant Analysis (LDA) classifier was trained to discriminate target vs. non target feature vectors extracted from single channels in a Leave One Out (LOO) cross validation on 50% of the data (the remaining 50% was held out for final
testing). For each of the twelve channels the average LOO classification accuracy (LOO-CA) was determined and channels were ranked by accuracy. The most commonly highest ranked channels across all subjects consisted of 01, P7 and TP7. The top three ranked channels were concatenated to form a new feature vector (27 features per vector) and a further LOO cross validation was performed. The results of this are reported as LOO-CA3. A single trial test of target vs. non target is also applied on the training data (Target vs. Non Target - Single Trial).
3.3. m VEP Classification - 5 Class Using all the training data (50% of data) a new LDA classifier is produced to classify target vs. non target data. To classify individual symbols in a single trial test each feature vector associated with each stimulus in a trial is classified as either target or nontarget. The LDA classifier produced a distance value, D, reflecting the distance from the hyper plane separating target and non-target features (D>O for target and DO, however the value of D is normally maximal among the target stimulus i.e., the stimulus on which the user is focused). Single trial results for five class are reported for the training data and then the setup is applied on the remaining 50% of the data, unseen testing data. Offline analysis was performed using customised MATLAB code along with the BioSig [36] and LIBSVM [37] toolboxes.
4. RESULTS
4.1. Offline Testing Data from all ten subjects were analysed for each of the 5 game levels with the addition of the training level and the Crash Bandicoot game with the white background omitted from the button controller area. Four methods are used to analyse the subjects data namely, LOO-CA3 (test 1), target vs. non target single trial (training) (test 2), single trial 5 class (training) (test 3) and single trial 5 class (testing) (test 4). Fig. 10 shows the average test 1 result for all ten subjects