A sample of discriminant features extracted by ROC analysis.
Electroencephalogram or electroencephalography (EEG) has been widely used in medical fields and recently in cognitive science and brain-computer interface (BCI) research. To distinguish metal tasks such as reading, calculation, motor imagery, etc., it is generally to extract features of EEG signals by dimensionality reduction methods such as principle component analysis (PCA), linear determinant analysis (LDA), common spatial pattern (CSP), and so on for classifiers, for example, k-nearest neighbor method (kNN), kernel support vector machine (SVM), and artificial neural networks (ANN). In this chapter, a novel approach of feature extraction of EEG signals with receiver operating characteristic (ROC) analysis is introduced.
- brain-computer interface (BCI)
- electroencephalogram or electroencephalography (EEG)
- artificial neural networks (ANN)
- support vector machine (SVM)
- receiver operating characteristic (ROC)
- Fourier transformation (FT)
The electrical activity of the brain can be measured by electrodes placed on the scalp and the observed signal is called electroencephalogram or electroencephalography (EEG). EEG is also called “brain wave” and it has been widely used in clinical diagnose of brain disease since the early time of last century .
Different mental tasks yield EEG signals in different patterns in the different observation values. For example, in the case of human brain, the resting state (relax state), the most prominent power spectra are 8–15 Hz EEG signals (so-called “alpha-wave”) observed in posterior sites, meanwhile, 16–31 Hz signals (beta-wave) appears in the mental tasks such as active thinking, high alert, anxious, etc. Gamma-wave, EEG with higher than 32 Hz, displays during cross-modal sensory processing such as combining the stimuli of visual and auditory. On the other hand, the location of electrodes on scalp records different EEG signals spatially, and they are called EEG signals in different “channels”. The allocation of electrodes is usually with the international 10–20 system. The name of 10–20 system comes from those adjacent electrodes that are allocated in distances of 10 or 20% of the total front-back or right-left of skull. More channels, more spatial features, may result in higher recognition rate of mental tasks. On the other hand, few channels give lower computational cost in the EEG classification systems.
In last decades, EEG has been utilized in the field of the brain-computer interface (BCI) for its ability of the mental task recognition [2, 3, 4, 5, 6]. Mental tasks indicate the state of activity of the brain with some specific tasks. For example, imagining writing a letter, counting, calculating, or raising a hand, a leg, etc. There are many classifiers for EEG recognition that have been proposed such as linear discriminant analysis (LDA), support vector machine (SVM), artificial neural networks (ANN), fuzzy inference systems, Bayesian graphical network (BGN), and so on. However, for the reasons of the complex nature of EEG signals, for example, noise and outliers, nonstationarity, high dimensionality, individual difference, etc., the pattern recognition (classification) problem of EEG signals is still a high hurdle for BCI realization.
To normalize the raw EEG signals, Nakayama and Inagaki proposed to reduce the number of the time series data of power spectrum of frequency given by fast Fourier transformation (FFT) with average values and normalize the FFT by a nonlinear normalization function . To extract discriminant features of EEG signals for mental task recognition, Li and Zhang proposed a regularized tensor discriminative feature space, which includes multichannels, power spectrum of frequency, and those data in time series: channel × frequency × time . Obayashi et al. applied Nakayama and Inagaki’s pre-processing method to their practical EEG recognition system with single channel information in . In , Jrad and Congedo used spatially weighted SVM (swSVM) to build a spatial filter for each temple feature. In the previous works of authors , discriminant temporal frequency data were utilized to reduce the flattening of different EEG patterns adopting the pre-processing method of , temporal spatial frequency concept, and average moving processing of  were adopted to obtain higher rate of mental task recognition.
Recently, we proposed to find the discriminant feature of temporal frequency by receiver operating characteristic (ROC) analysis in . The discriminant feature of temporal frequency indicates the power spectra of FFT in an interval of time series of EEG data, which are higher relative to a mental task comparing with other intervals (windows). ROC analysis has been widely utilized in medical & diagnostic science [10, 11], microarray classification , and recently in EEG classification . It is a stochastic criterion to classify two kinds of probability distributions and the details will be described in the next section.
In this chapter, discriminative feature extraction methods of EEG signals, which play an important role for classifiers, are discussed. Specially, an advanced temporal–spatial spectrum feature extraction method is introduced .
2. Discriminant feature extraction using ROC analysis
2.1. ROC analysis
Receiver operating characteristic (ROC) analysis was first used in radar signal detection in 1940s. The classification results of data in two kinds of distributions can be divided into four categories: true positive (TP), false positive (FP), true negative (TN), and false negative (FN). A curve is plotted by the rate of TP against the rate of FP and it can be a measure of classification accuracy.
Now, let the TP of class A be in the shadow area α, and FP in area 1−β, where β is the TP of class B (See Figure 1). When the dividing line between A and B is slid along x axis, a ROC curve is plotted indexing the divisibility of the two probability density functions (See Figure 2). If two distributions completely overlapped, α = 1−β.
In Figure 2, the area below the ROC curve is called “area under the curve” (AUC). This value takes from 0.0 to 1.0, and it is an indicator of the divisibility of the two distributions. If the value of AUC becomes 0.5, two distributions are completely overlapped. Conversely, when the value of AUC reaches 1.0 (or 0.0), it means that the two distributions are completely separated.
In the practice procedure of ROC analysis, the area of α, that is, the rate of TP, and 1−β, the rate of TN, can be calculated by the number of training samples, which are labeled data belonging to different classes.
2.2. Discriminant feature extraction of EEG signals
In , power spectrums of an interval of frequencies given by EEG signals FFT, which has a distinguish value to neighbors were used as discriminant features as the input vectors of classifiers. The flow chart of this method is depicted in Figure 3. Algorithm I shows the method in detail.
Step 1. Dividing (windowing) the original EEG signals into several intervals;
Step 2. Executing discrete Fourier transformation (DFT) in different intervals and normalizing the transformation results;
Step 3. Calculating the average power spectrum of banded (limited) frequencies in each phases;
Step 4. Finding a special (feature) interval, in which average power spectrum is the most different one from its neighborhoods;
Step 5. The power spectrum of FFT in the windowed frequencies and their average values are used as the feature data for classifiers.
A sample of the first processing (Step 1) is shown in Figure 4. In Figure 4, an EEG signal, which is a time series data (the potential of an electrode) of one channel, is divided into five intervals. DFT is executed in each interval at Step 2, and as a sample, the result of the second intervals (at time 30–60) is shown in Figure 5.
The normalization of DFT results is given by a nonlinear function .
where x(n) is the original DFT power spectrum of frequency n.
This nonlinear normalization reduces the vibration of time series of DFT results, avoiding the overfitting when classifiers are designed.
A frequency interval, which has distinguished power spectra for a certain mental task is chosen by Eq. (2).
where p = 1, 2, …, P is the number of intervals, Fph is the power spectrum on the frequency, h = hlow, hlow + 1, …, hup is the frequency, hlow and hup are bands of feature frequencies of mental tasks and they were 4 and 45 Hz, respectively in our experiments.
For ROC analysis, it gives a measure of the difference between two probability distributions, it is validly used to find the discriminant features for EEG signal classification. In , Nguyen et al. utilized the AUC of ROC curve to select the elite wavelet coefficients, and in , we adopted an algorithm that using high AUC values to select metal task-related frequencies of EEG signals in different channels, respectively, and using the power spectrum of these frequencies as discriminant features for various classifiers such as SVM, ANN [including multi-layer perceptron (MLP), and deep neural networks (DNN)], k-nearest neighbor, decision tree (DT), and so on. The discriminant feature extraction method using ROC analysis is given by Algorithm II.
Let the input signals be xkc, m, n (c = 1 or 2, k = 1, 2, …, K), where k indicates the kth EEG signal of a set of EEG data, and c indicates the class of mental task, m indicates the channel number, and n is the time of signal.
Step 1. Perform FFT to all the EEG signals xkc, m, n and let the result be power spectrum Em, p (p = 1, 2, …, P) corresponding to frequency Fkc, m, p, where p indicates the order number of frequencies.
Step 2. Obtain Pk1, m, p and Pk2, m, p, which are two probability density functions of Fkc, m, pat p frequency, where class c = 1 and 2 of K signals of channel m.
Step 3. Calculate the ROC curve and its AUC Am, pof Pk1, m, p and Pk2, m, p.
Step 4. Repeat Step 2 and Step 3 on all channels, a set AUCm, n of frequency p in channel m is obtained.
Step 5. Find P points of frequencies, in which Am, p is high.
Step 6. Power spectrum Em, p(p = 1, 2, …, P) of the unknown EEG signal are used as input feature vector of a classifier.
The main difference between Algorithm I and II is that the power spectra in a special interval of frequencies, which is mostly related to an event of brain activity, are chosen in the former, meanwhile the power spectra of special frequencies chosen by high AUC of ROC are chosen as discriminant features in the later algorithm. The flow chart of EEG signal classification using ROC analysis is depicted in Figure 6.
Figures 7–9 showed a sample of the processing. In Figure 7, a raw EEG signal and its FFT result are shown. Note that the number of horizontal axis indicates the order of frequencies, and the value of vertical axis is the power spectrum. In Figure 8, the distribution of the power spectrum of each frequency is calculated using the labeled samples. For example, there are K samples including N samples of class A and K−N samples of class B as shown in Figure 8. AUC of the power spectra on each frequency is shown in Figure 9. Additionally, frequencies with high AUC extracted by a threshold line are used as criteria of discriminant feature selection. For example, in the case of three input dimensions for a classifier, the input vector is the power spectra with high AUC of frequencies as shown in Table 1.
|Ordered AUC||Number of freq.||Power spectrum|
(input of classifiers)
To compare the performance of different feature extract methods for EEG signal classification, experiments with two kinds of EEG data were performed . One was a benchmark data set given by Brain-Computer Interfaces Laboratory, Colorado State University [14, 15], and another was from BCI competition II . Classifiers used in the comparison experiments for different feature extraction methods were kernel SVM, MLP, kNN, deep neural network (DNN), and DT, in which source coded are in a software package R  as shown in Table 2.
|ROCR||ROC analysis/AUC calculation|
|Kernlab||Support vector machine (kernel SVM)|
|nnet||Neural network (MLP)|
|class||k-nearest neighbor (kNN)|
|h2o(+JavaVM)||deep neural network (DNN)|
|rpart||decision tree (DT)|
The evaluation of the performance of different feature extraction methods uses the accuracy of classification, which is given by Eq. (3).
4. Benchmark data and experiment results
Open access free website of BCI laboratory of Colorado State University [14, 15] provides Benchmark EEG data with five kinds of mental tasks as shown in Table 3. The data were measured by six channels with EEG sensors (See Figure 10) and one channel data of an EOG sensor (to measure the movement of an eye). The sampling rate is 250 Hz, and EEG data are recorded in 10 seconds, that is, 2500 time series data obtained by one trial. EEG signals of each mental task are recorded in 10 trials of five subjects. For the ROC analysis classifies two classes data, “Baseline” (relaxing state) and “Multiplication” (Multiplication calculation mentally) data, were used in our experiment. Additionally, training samples and testing samples used EEG data of the same subject, which were chosen randomly with a ratio of 15:5.
The classification accuracies of Algorithm I , and Algorithm II  by different classifiers are shown in Table 4. In Table 4, it is also shown that different dimensionalities of the input vector influenced the classification accuracy. Feature extraction method using Algorithm II. (FFT and ROC analysis) had a prior performance especially in the case of 140-dimension input vector. The highest classification accuracy 97.5% was given by kernel SVM classifier, and DNN stood the second position with 95.37% using Algorithm II feature extraction method, respectively.
4.1. BCI competition II data and experiment results
BCI competition II data  were also used in the performance comparison of different feature extraction methods. There are two-class data named “Ia” and “Ib,” which are EEG data obtained by a healthy subject and an amyotrophic lateral sclerosis (ALS) patient. In each data set, two kinds of mental tasks were required, respectively. One was to move a cursor up (class A) and another was to move the cursor down (class B). Details of these EEG data descriptions are shown in Table 5. Additionally, training samples and testing samples were chosen randomly with a ratio of 240:28 for Ia and 180:20 for Ib.
|Baseline||Relaxing as much as possible|
|Multiplication||Calculating multiplication mentally.|
|Letter-composing||Considering the contents of a letter|
|Rotation||Imagining rotation of a 3-D object|
|Counting||Imagining writing a number in order|
|Classifier||Feature extraction method|
(FFT and ROC analysis)
|Deep neural network||61.67||71.67||95.37||94.58|
|Data set||Mental tasks||Trials||Channels||Samples/Ch.||Sampling freq.|
The accuracies of classification of Ia and Ib by different feature extraction methods and classifiers are shown in Tables 6 and 7, respectively. Algorithm II (FFT and ROC analysis) showed the highest classifications for all classifiers. The highest accuracy for data Ia was 91.23%, given by kernel SVM using 1120 dimensions of input vector, which were discriminant features extracted by Algorithm II, and the same methods yielded the highest classification rate 77.65% for data Ib. These accuracies are higher than the best classification rates 90.10 and 56.67%, which are the results of a state-of-the-art method of EEG signal recognition . The future work of the improvement of Algorithm II is to find the optimal dimensionality of the discriminant feature space. It is hard to consider higher dimensionality results higher classification accuracy as shown in these experiments. It was better to choose 140-D in the case of benchmark data (Table 4), and oppositely, 1120-D was more suitable for BCI competition II data (Tables 6 and 7).
|Classifier||Feature extraction method|
(FFT and ROC analysis)
|Deep neural network||57.72||62.08||83.48||86.10|
|Classifier||Feature extraction method|
(FFT and ROC analysis)
|Deep neural network||43.90||49.25||69.93||75.25|
To recognize the mental tasks by EEG signals, two kinds of temporal-spatial frequency–based feature extraction methods were introduced in this chapter. In Algorithm I, event-related intervals of the raw EEG time series data (temporal information) was extracted at first, and the averaged power spectra of frequencies given by FFT within the interval (frequency information) were used as the discriminant features. In Algorithm II, event-related frequencies of EEG’s FFT were extracted by ROC analysis with high AUCs. The input space for classifiers was composed by all features extracted by two algorithms from multiple channels, so the spatial information was also included in these feature extraction methods.
Pattern recognition of EEG signals has been studied for decades, and it plays an important role in the field of human robot interaction (HRI). So, we expect that the feature extraction methods introduced in this chapter can be adopted in the real HRI systems in the near future.
We would like to thank dear Editors for their appropriate advices during the revision of this paper. This work was supported by Grant-in-Aid for Scientific Research (JSPS No. 26330254 & No. 25330287).