**3.3. P300-MI hybrid**

groups flickered at frequencies 6, 6.67, 7.5, and 8.57 Hz to evoke SSVEP, while 100 ms interval was used to intensify and change the color of large central button to elicit P300. SVM and discrete Fourier transform (DFT) were used for classification of P300 and SSVEP, respectively, with an overall classification accuracy of about 90% from eight healthy subjects (20–31 years)

A hybrid SSVEP-P300 BCI was proposed by [55] as mentioned earlier to improve spelling accuracy combining random flashing and periodic flickering to evoke P300 and SSVEP, respectively. All the cells of 6x6 matrix were flickered on black background with six frequencies 8.18, 8.97, 9.98, 11.23, 12.85, and 14.99 Hz, and selection of these frequencies were based on the higher SSVEP amplitude and easier target detection while orange crosses were flashed for 120 ms in random manner. Twelve healthy subjects (5 male and 7 female subjects, age 21–29 years) with good visions were used and the performance of the hybrid system was evaluated online using single trial. This experiment claimed to have the best performance. In [56], a BCI with shape changing and flickering speller was designed, rather than traditional color changing as in [67], combining SSVEP and P300 in which the classification accuracy was improved in each of the individual systems. Shape changing of red boxes to arrows was used for P300 and flickering of those four boxes with frequencies 6, 8, 9, and 10 Hz for SSVEP. Ten healthy right-handed subjects with normal vision (9 male and 1 female subjects, age 22–27 years) were used for five offline experimental sessions having 20 runs of each sessions lasting for 4 s so the one session was 40 min including 10 min rest. The subjects found

An asynchronous hybrid BCI combining P300 and SSVEP was proposed by [68] where the information transfer and control state (CS) detection was accomplished using P300 and SSVEP, respectively. This system operated in sequential manner in both offline and online experiments. Ten healthy subjects (7 males and 3 females aged 19–28 years) were participated in both experiments where P300 was elicited from flashing of a 6x6 matrix of 36 characters (A-Z and 0–9), and SSVEP was obtained from flickering of the characters black and white alternatively with frequency around 17.7 Hz. Two classifiers: Fisher's linear discriminant analysis (FLDA) for first three subjects and Bayesian linear discriminant analysis (BLDA) for remaining subjects were used for P300 classification. Inclusion of SSVEP into P300 improved

In [69], same target stimulus was used to elicit P300 and SSVEP blocking (SSVEP-B). A new speller of 3x3 matrix with characters 1–9 was proposed while all the characters flickered at a constant repetitive rate, and each character stopped flickering randomly within a short time. The repetitive flickering elicited SSVEP and the interrupted flickering of a character would elicit SSVEP-B and P300, a rare event at the same time which was detected simultaneously. Twelve right-handed healthy subjects (7 male and 5 female subjects, age 23–36 years) were participated in the experiment. The size and font of characters were changed with a variance of 0.49 ms for event-related potential, and brightness altered between light and dark with about 14.96 Hz for evoked potential. Stepwise linear discriminant analysis (SWLDA) was used for classification and the hybrid system produced better result than the individual systems.

performing "go/stop" control task using a real wheelchair.

124 Evolving BCI Therapy - Engaging Brain State Dynamics

the new hybrid less annoying.

the classification accuracy.

A hybrid P300-MI was used in [70] to control appliances in a virtual environment, in which P300 was used as a device control to operate control panel of virtual devices and MI as a navigation tool to turn left/right in the virtual environment, and the detection was sequential based on the activation of either MI or P300. Four healthy adults (1 male, 3 female subjects, age 23–25 years) participated and P300 was elicited using oddball paradigm, and MI using left-hand/right-hand movement imagination. The average online classification accuracy was achieved around 82% and authors claimed that more complex tasks in virtual environment could be performed compared to single pattern BCI.

The possibility of combining various brain signals for hybrid BCI was discussed by [71] merging P300 and ERD to control a robotic device such that additional features of one system could be used to improve classification accuracy of another. In this case, object selection from various options, called discrete decision, was done using P300 and movement control of the robot was done using ERD from motor imagination. EEG was filtered to a band 0–10 Hz for P300 classification. Optimum filter band varied with subject for ERD and the filter band was obtained from training data. Principal component analysis (PCA) and CSP was used in feature extraction for P300 and MI, respectively, and the discriminant classification was done by FLDA. Four healthy subjects were experimented with at least three MI trainings before the experiment by each subject with tasks: one out of five P300 symbols (1–5) and one out of two MI hand movements (left/right). Each subject performed one experimental session consisting of 60 trials, 30 trials for P300 in which targets were 1–5 numbers random flashings, and remaining 30 trials for MI where subjects needed to focus only one either on P300 or MI during each trial. Hybrid classification accuracy achieved was about 82% with average MI classification accuracy of 71%.

The selection of a target based on hybrid P300-MI BCI was developed by [72] on 1166×721 pixels monitor using two sequential tasks: cursor movement with the help of ERD from left-/right-hand movement imaginations and selection/rejection of the target from focused attention on flashing to evoke P300 keeping one system passive when other is active. Eleven healthy subjects (10 male, 1 female subjects, and age 22–32 years) were used in both online and offline analysis and each experiment was performed within 2 s with each trial duration for about 18 s. Two targets such as Green Square for selection and Blue Square for rejection were randomly appeared on the screen and the classification accuracy was about 93% in real time.

*x*(*n*) = \_\_<sup>1</sup>

**4.2. Linear discriminant analysis (LDA)**

each classes are represented as *μA*

*<sup>N</sup>* ) and N = length [*x(n)*].

*LDA* <sup>=</sup> (*μA* <sup>−</sup> *μB*)

technique for online classification of P300 [80].

the corresponding SVM is Gaussian SVM which is given as:

K(x, y) = exp(

where K(x, y) is kernel function and σ is kernel width.

**4.3. Support vector machine (SVM)**

where *WN* <sup>=</sup> *<sup>e</sup>* <sup>−</sup>*<sup>j</sup>*

( \_\_\_ 2*π*

classifiers in BCI research.

*<sup>N</sup>* ∑ *k*=0 *N*−1

Amplitude vs. frequency is plotted and the dominant amplitude at a particular frequency is obtained for SSVEP signal analysis in BCI. FFT classifier has been replaced by other better

LDA, also known as Fisher's LDA (FLDA), is a useful classification technique that transforms features into a low-dimensional space with high degree of separation. Suppose, there are a certain set of samples belonging to classes "A" and "B" whose mean and scatterings within

This method is simple to use, has low computational cost with high accuracy, and is widely used in number of BCI applications for P300 and MI-related tasks [77]; but, it also suffers from small sample size and linearity problems [78]. BLDA, a Bayesian version of LDA, is useful for P300 in which parameters are estimated with Bayesian regression whose probabilistic output value lies between zero and one. It gives good results for small number of train sets or noise contaminated data and may give poor result for a complex nonlinear EEG data [78, 79]. SWLDA is a stepwise LDA that performs space reduction by selecting suitable features and stepwise analysis to remove least significant features which is an effective classification

SVM is used for classification of linearly separable binary data sets that uses a discriminant hyperplane to identify classes and the best choice is the hyperplane that leaves maximum margin from both classes. The kernel generally used in P300 BCI is the Gaussian kernel and

SVM is simple and stable, and has a low variance which may be a key for low classification error for unstable and noisy P300 signals. SVM produces good results with high-dimension feature vectors and a small training set, thus outperforming LDA, but are generally slower than other classifiers [79]. BLDA is most robust for P300 application compared with LDA and SVM [81].

−||*x* − *y*|| 2 \_\_\_\_\_\_\_

2 \_\_\_\_\_\_\_ *SA* <sup>2</sup> + *SB*

, *μB* and *SA* 2 , *SB* 2 *X*(*k*) *WN*

<sup>−</sup>*kn* (2)

http://dx.doi.org/10.5772/intechopen.75132

127

Hybrid Brain-Computer Interface Systems: Approaches, Features, and Trends

, respectively, and LDA is calculated as [76].

<sup>2</sup> (3)

<sup>2</sup> *<sup>σ</sup>*<sup>2</sup> ) (4)

MI-based brain switch was merged with P300 sequentially by [73] in which right-/left-hand movement imaginations was used for control signal acting as a brain switch to turn ON/OFF P300 speller. Eleven healthy subjects (8 male, 3 female subjects, and age 23–30 years), with a few subjects having previous experience to P300 or MI, were experimented. Offline training was done for P300 and MI before real tests in a 22″ LED monitor containing 6×7 matrix of 26 English letters, 10 numbers, a few special characters, and commands. SWLDA was used for P300 classification and filtration in the alpha-beta rhythm range with two bands discrimination using CSP for MI and the classification accuracy achieved was about 92% for P300.

An efficient approach was proposed by [74] combining MI with P300 in a block diagonal matrix form to improve classification accuracy concatenating features of each paradigm where first-order information was used for P300 and second order for MI. Twelve volunteers (10 male, 2 female subjects, age 22–35 years) were experimented provided with eight flashing buttons and an arrow cue such that screen remained blank for initial 2.25 s and a cross appeared on the screen from 2.25 to 4 s to attract subject's visual attention. Arrow cue was displayed from 4 to 8 s such that up arrow was for P300 task to focus on the random flashing buttons without MI tasks and right arrow for MI task to make left-/right-hand imaginations. Test data was obtained from second experimental session and the classification accuracy was about 92%.
