Support-vector-machine-based Meditation Experience Evaluation ...

Journal of Medical and Biological Engineering, 34(6): 589-597

589

Support-vector-machine-based Meditation Experience Evaluation Using Electroencephalography Signals Yu-Hao Lee1

Sharon Chia-Ju Chen2

Ming-Shing Young1

Chung-Yao Hsu5

Yung-Jong Shiah3 Geng Qiu Jia Cheng 6

Shih-Feng Wang4 Chih-Lung Lin1,*

1 Department of Electrical Engineering, National Cheng Kung University, Tainan 701, Taiwan, ROC Department of Medical Imaging and Radiological Sciences, Kaohsiung Medical University, Kaohsiung 807, Taiwan, ROC 3 Graduate Institute of Counseling Psychology and Rehabilitation Counseling, National Kaohsiung Normal University, Kaohsiung 802, Taiwan, ROC 4 Department of Aviation & Communication Electronics, Air Force Institute of Technology, Kaohsiung 820, Taiwan, ROC 5 Department of Neurology, Kaohsiung Medical University, Kaohsiung 807, Taiwan, ROC 6 Tibetan Nyingmapa Kathok Buddhist Organization, Sichuan 627350, China 2

Received 17 Dec 2013; Accepted 7 May 2014; doi: 10.5405/jmbe.1776

Abstract Meditation is used to improve psychological well-being, but there is no scientific quantitative evidence to prove the relation between them. Therefore, in this study, an effective classifier, namely a support vector machine (SVM), is applied to classify meditation experiences and help validate the interaction between emotional stability and a meditation experience. Three groups (10 subjects in each), created based on practice experience in meditation (S group with 10-30 years, J group with 1-7 years, and N group with 0 years of experience in Tibetan Nyingmapa meditation), were recruited to receive visual stimuli in the form of affective pictures. The images shown were selected from the International Affective Pictures System (IAPS), a confidential database. The response signals were acquired through physiological examination via electroencephalography (EEG). The subjects’ data were entered into two classification systems, namely those based on the classification and regression tree (CART) method and the SVM method, respectively, and the outcomes were compared. From the classification results, SVM had a higher accuracy rate (98%) than that of CART (79%). The stability and robustness of SVM are higher than those of CART, as determined by performing over 100 repetitions and using various variable numbers. An evaluator based on SVM can thus assess a meditation experience through visual emotional stimulation. The results can help explain emotional stability during meditation. Keywords: Electroencephalography (EEG), Classification and regression tree (CART), Support vector machine (SVM), meditation experience, emotional stability

1. Introduction Emotional stability is an important index of psychological well-being and a necessary part of health management [1-3]. Generally, emotional stability means that a person is able to remain stable and balanced even under stressful circumstances. Emotional stability has been shown to improve the quality of life [4]. In addition, it has an indirect but significant effect on job performance [5-8]. In contrast, neuroticism, characterized by anxiety, depression, and schizophrenia, damages personal health [9], and leads to disrupted immune function [10] and abnormal cardiac function [11]. Hence, in order to prevent neuroticism, promoting emotional stability can be used to * Corresponding author: Chih-Lung Lin Tel: +886-6-2757575 ext. 62338; Fax: +886-6-2345482 E-mail: [email protected]

strengthen psychological health. Emotional stability is an abstract condition which can be observed as calmness of mind and low emotional fluctuation in daily life. To efficiently detect the state of calmness of mind, Archa and Craske [12] investigated the influence of visual stimulation on arousal degree through a focused breathing method (an introduction to meditation). They found that focused breathing successfully induces positive effects on emotional stability. Another study found that consistently performing mediation reduces emotional variation of negative mood [13]. However, these studies used subjective and qualitative analyses (questionnaires and interviews). Although qualitative analysis can be used to observe the positive effect of meditation on emotion management, it cannot provide the exact degree of influence on psychological health. Therefore, in order to further understand the variation of physical responses caused by meditation, adopting quantitative analysis is necessary.

J. Med. Biol. Eng., Vol. 34 No. 6 2014

Some studies that have conducted a quantitative analysis on emotional stability have found that low affective reactivities can be measured through physiological measurements [14-18]. For example, Schmidtke and Heller [14] assessed the relation between brain activities and emotional stability with a series of questionnaires and electroencephalography (EEG) physical examinations. They found that the degree of emotional stability observed in EEG is proportional to the increase in alpha power in the left frontal lobe. Aftanas and Golosheykin [15] detected a reduced signal of arousal visual stimuli through an EEG spectrum and inferred that mediation improves emotional stability. The extra benefits for long-term meditation users include lower sensitivity to EEG event-related potentials (ERPs) under adverse emotional load caused by visual manipulation. Similar findings have been reported by fMRI studies with visual [17] and auditory [18] stimuli. They found that long-term meditation leads to emotional stability. Furthermore, long-term meditators show less affective reactivity to emotional arousal events physiologically. These studies revealed that emotional stability helps one to adapt to changing situations. Therefore, meditation practice is regarded as an effective approach for achieving calmness of mind. Evaluating the experience of meditation is thus an important issue for predicting emotional stability. Several algorithms can classify emotional reactivity into various categories according to the physiological response. They have been used to estimate the meditation experience of participants. Classification has been conducted using Bayes’ rule [19], the k-nearest neighbor (kNN) method [20], artificial neural networks (ANNs) [21], classification and regression tree (CART) [22-27], and support vector machine (SVM) [28-34]. Among these methods, CART and SVM are used most frequently. CART, which can rapidly differentiate between groups, has been extensively used in numerous applications, including public health [22], violence risk assessment [23], chronic heart failure [24-27], and detection of epilepsy [27]. The advantage of CART is that the analysis process can be understood without complex mathematical calculation and its outcomes can be easily interpreted. This research adopts SVM and CART and compares their operation. SVM is a supervised machine learning algorithm that is effective in either linear or nonlinear discrimination between groups. In addition, SVM is suitable for some complicated problems that involve many confounding factors such as brain-computer interfaces [28,29], detection of epilepsy [30], bioinformation [31], and emotion assessment [32-34]. SVM and CART are used to determine the affective reactivity of various meditation experiences. The present study evaluates the experience of meditation via the response to emotional arousal stimuli analyzed using SVM and CART. The emotional arousal stimuli were obtained from the International Affective Picture System (IAPS), which provides normative visual stimuli and has been widely used in studies of emotion [35]. Moreover, CART and SVM are used as classifiers to categorize the experience of meditation according to the physical response of the participants. This study hypothesizes that the experience of meditation can be identified using CART and SVM.

590

2. Methods 2.1 Participant preparation Thirty participants were recruited for this experiment. They were divided into the following 3 groups: the senior experience group (S) included ten participants (7 males and 3 females; mean age = 52 years and standard deviation (SD) = 12 years) with 10-30 years of experience (mean = 17, SD = 8) of Tibetan Nyingmapa meditation. The junior experience group (J) included ten participants (5 males and 5 females; mean age = 55 years and SD = 12) with 1-7 years of experience (mean = 5, SD = 2 years) of Tibetan Nyingmapa meditation. The inexperienced group (N) included ten participants (3 males and 7 females; mean age = 52 years and SD = 10) with no experience in any meditation. To ensure uniformity of mediation practice, all participants were asked to use the focused breathing method, which is the most common and basic form of concentrative meditation [36,37]. The inexperienced participants were trained to perform focused breathing and were asked to practice the skill for 2 hours before the day of the experiment. All participants were free from cardiac, pulmonary, metabolic, and any other disease that could cause autonomic nerves system (ANS) dysfunction. Furthermore, all participants were medication-free and none were habitual drinkers or smokers. They were asked not to consume caffeine or alcoholic beverages for 12 hours and not to exercise 24 hours prior to the experiment, and they are also asked to refrain from eating and drinking anything for at least 3 hours prior to the experiment. This study was approved by Institutional Review Board-I of Kaohsiung Medical University Chung-Ho Memorial Hospital (Protocol Number: KMUH-IRB980185). 2.2 Experiment design Sixty color pictures selected from IAPS [35], Center for the Study of Emotion & Attention, University of Florida, were divided into 6 categories, namely snakes, non-threatening animals, neutral people, mutilations, erotica, and neutral scenes, with 10 pictures in each category. The 10 pictures in each category were divided into 2 sets, each with 5 pictures. Snakes, mutilations, and erotica were grouped as arousal stimuli, with IAPS normative bipolar arousal ratings (1-9 scale: 1 = low arousal, 9 = high arousal) averaging 6.70 (SD = 0.38); non-threatening animals, neutral people, and neutral scenes were grouped as neutral stimuli, with IAPS normative bipolar arousal ratings averaging 3.55 (SD = 0.89). A t-test revealed a significant difference between the 2 groups of stimuli under a threshold for significance set at p < 0.001. The experimental procedure is illustrated in Fig. 1. For each participant, three-stage sessions comprised initial baseline, meditation, and visual stimulation. An initial physiological baseline measurement period consisted of 4 60-s blocks of ongoing activity with a counterbalanced random order of openeyed and closed-eyed conditions within each block, with EEG recorded simultaneously. Afterwards, to ensure that the

591

Meditation Evaluated Using EEG Signals

meditators maintained emotional stable before the task, they were asked to practice meditation for 5 min 3 times with a 2-min interval between meditations after the baseline period.

frequency bands: theta (θ, 4.25-8 Hz), alpha (α, 8.25-12 Hz), beta (β, 12.25-30 Hz), and gamma (γ, 30.25-50 Hz). The average power of the four frequency bands in natural logarithm format was calculated for the baseline and aroused periods for each channel. The difference in power between the arousal stimulation and the baseline for a certain band in a certain channel was regarded as a feature. 64 features (4 frequency bands × 16 EEG channels) were obtained in total. 2.5. Statistical analysis

Figure 1. Experiment flow chart. Starts with baseline session (4 min) with eyes-open and then eyes-closed order, and then meditation for 20 min in 3 cycles with focused breathing method and finally task session for 20 min.

After meditation, the participants were subjected to the visual stimuli. The procedure of presenting visual stimuli was similar to that used by Sabatinelli et al. [38]. The stimulation period was divided into 2 9-min blocks separated by approximately 1 min of rest. Thirty pictures were presented in each block, 5 from each of the 6 categories (3 arousal and 3 neutral). For each trial, the picture was presented for 6 s, and then a black screen with a white cross in the center was shown for 12 s. The picture categories were presented in counterbalanced order, so the frequencies of each category were almost the same in the first and second runs. 2.3 EEG signal acquisition EEG signals were acquired from 16 electrodes placed according to the standard 10-20 system, i.e., Fp1, Fp2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T3, T4, T5, and T6, with 2 reference electrodes attached to the earlobes. These physical signals were recorded using the Nicolet Clinical EEG acquisition system with a sampling rate of 250 Hz. The impedance of each electrode was less than 10 kΩ. Before the EEG signals were analyzed, artifacts from various sources, namely eye and body movement and power line interference, were removed. Artifacts caused by body and eye movement are very-low-frequency noise ( < 4 Hz), and can be filtered by a high-pass filter with a cut-off frequency at 4 Hz [39,40]. Signals between 4 and 70 Hz were filtered by a digital bandpass filter. A digital notch filter was applied at 60 Hz to remove external noise caused by the power line output. 2.4 Feature extraction

The difference in the EEG spectrum power between the arousal stimulation and the baseline periods was compared among the three levels of meditation experience for each frequency band across all channels. To determine the differences among the three groups, one-way analysis of variance (ANOVA) was used to check if the distributions of the three groups varied before and after systematic classification. When significant effects were observed, post-hoc analysis implemented using a paired t-test was used to evaluate between-group variation from the three groups with multiple-comparison correction. Data are presented as means ± SD. The threshold for significance was set at p < 0.001. 2.6. Classifier This work examines the influence of meditation experience on the response to arousal visual stimuli in EEG signals. In order to distinguish the levels of emotional stability, two classifiers were used to evaluate the effect on arousal visual stimulation. All subjects’ data (16 EEG channels × 4 frequency bands × 3 groups) were input into CART and SVM classifiers. In addition, in order to examine the predictive accuracy rate, this study employed k-fold cross validation. In k-fold cross validation, all data sets are randomly divided into k subsets with the counterbalanced conditions. In this study, each subset comprised three data points from each group under different arousal categories, and each data point was selected for testing once. One subset was regarded as the test set and the remaining sets were used for training. The CART and SVM classifiers were applied to determine the experience of meditation according to the physiological response to visual stimulation. 2.6.1. Classification and regression tree The CART algorithm creates a decision tree that partitions the data into categories (c) based on the data’s categorical characteristics. Then the algorithm decides a proper splitting point to divide the data into two groups with the minimum impurity. This procedure is applied recursively to each sub-group until CART detects that no further gain of the impurity measurement can be made. In the CART algorithm, Gini impurity (IG) is used to check how often an item (P) in the input data set (D) is misclassified into a sub-category (j): c

For each EEG channel, a power spectrum was computed using the fast Fourier transform (FFT) every 4 s with a 2-s overlap. The EEG frequency spectrum was divided into 4

I G ( D)  1   Pj 2 j 1

(1)


The Gini impurity reaches its minimum (zero) when all cases fall into a single category. Therefore, when dataset D is split up into two sub-sets D1 and D2, the cost of Gini impurity, IG_split, can be expressed as:

I G _ split ( D) 

K1 K I G ( D1 )  2 I G ( D2 ) K K

(2)

where K is the number of data in set D, and K1 and K2 are the numbers of data in sets D1 and D2, respectively. The smallest IG_split across all the possible splitting points is chosen to split the dataset. The cost of Gini impurity is computed with all possible splitting nodes across all extracted data sets (4 frequency bands × 16 EEG channels). The whole tree expansion process stops when the impurity measures cannot be improved by the splitting point. When the tree expansion process completed, all data are assigned to a labeled category (with “^” on the top of group mark). 2.6.2. Support vector machine SVM is a kernel-based supervised machine learning classifier that attempts to decide a functional margin to group data by transforming the input data to higher-dimensional space where the separation of two class sets can be achieved by a linear separation with hyperplane [41]. The SVM algorithm finds the largest distance between the decided hyperplane and the nearest training points of any data class because if the margin is larger the generalization error of the classifier will be lower. Thus, the training dataset is first used to set the classifying parameters. The hyperplane decided by the largest margin is applied to examine the test dataset. Under this arrangement, given a training set of instance-label pairs (xi,yi), i = 1,…,l, where xin is the training vector, and yi[1,-1]l denotes the class of xi, the SVM attempts to solve the following optimization problem: l 1  min  wT w  c   j  w,b , j 1 2  T yi ( w  ( xi )  b)  1  i ,  j  0

592

using grid search with cross validation. Various pairs of (C,γ) values were tested, in which C was searched from 0.5 to 8 and γ was searched from 0.0625 to 2. The one with the best validation accuracy rate was picked, namely (C,γ) = (1,0.0625). Because SVM is suitable for confounder-mixing problems, the authors are convinced that this method can deal with the subtle difference of meditation experience between senior and junior meditators. In this study, the algorithm adapted from the LIBSVM library [43] implements the “one-against-one” strategy to classify the three types of meditator.

3. Results To evaluate the reproducibility of classifiers, the estimated accuracy rate was computed by averaging the classification rates of 10-fold cross validation. Then, the system robustness between two classifiers was tested by evaluating the characteristic curves according to the number of features. 3.1 Data distribution One-way ANOVA was used to test the feature differences of frequency bands among the three groups. Significant group differences were found for all EEG frequency bands (theta [F(2,1437) = 218.3, p < 0.001], alpha [F(2,1437) = 239.5, p < 0.001], beta [F(2,1437) = 10.6, p < 0.001], gamma [F(2,1437) = 20.5, p < 0.001]). In Fig. 2, the histogram of the three groups shows significant group differences between any two groups obtained with the post-hoc t-test (p < 0.001). In Figs. 2(a) and 2(b), in low-frequency bands, theta and alpha, the distribution can be distinguished between any two groups (p < 0.001). However, for the high-frequency bands, beta and gamma, there is no significant difference between S and J groups (see Figs. 2(c) and 2(d)). The details of frequency bands for each group are presented in Table 1.

(3)

where w is a vector orthogonal to the hyperplane , C > 0 is a penalty factor of the error term, ξi, and b is the bias term. Here, xi is mapped into a higher-dimensional space by the function 𝜙 when it is not possible to determine which group a participant belongs to by using linear separation. The kernel function, K(xi,xj) 𝜙(xi)T𝜙(xj), used in this study is the radial basis function (RBF) K(xi,xj) exp(-γ‖xi-xj‖2), γ > 0, which obtains the highest classification accuracy for our case. With a given penalty factor C and kernel parameter γ, the trade-off between errors of the SVM from data and the boundary margin is determined. For a multiclass classification problem, the one-againstone method gives good results [42]. This method is used to train k(k-1)/2 classifiers. Each of the classifiers is trained to determine two different classes of a total k classes. Testing data are determined to be in a specific class from voting by the training classifiers. The penalty factor C and kernel parameter γ were set

(a)

(b)

(c)

(d)

Figure 2. Data distribution of S, J, and N groups and statistical comparisons with post-hoc t-test on every pair in (a) theta band, (b) alpha band, (c) beta band, and (d) gamma band. (* denotes statistical significance at p < 0.001)


593

Table 1. Mean (SD) values of EEG features and p-values from one-way ANOVA among 3 groups for each EEG frequency band. S J N

Theta 1.258 ± 1.328 0.554 ± 0.964 0.047 ± 0.149 F(2,1437) = 218.3, p < 0.001

Alpha 0.427 ± 0.712 -0.228 ± 0.521 -0.837 ± 1.168 F(2,1437) = 239.5, p < 0.001

3.2 CART classification Regarding CART results, the decision tree of a certain round in 10-fold cross validation is shown according to the value of the Gini index in Fig. 3. The feature T5_alpha represents the minimum Gini index (Eq. (2)), as shown in the first node of the tree. At the first step, all data of the three groups were split by the first node. The results reveal that 85% (23/27) of senior meditators are assigned to the same category ( S ). Then, the next decision criterion, C3_alpha, indicates that the second minimum Gini index successfully divided 74% (20/27) of junior meditators ( J ) and 78% (21/27) of inexperienced meditators ( N ). The whole tree expansion process continues until the Gini error rate is lower than the tolerance level. The final misclassification rate of the training data is 4.9% (4/81, 2 N group members were misassigned to after the P4_theta split, and 2 N group members were misassigned to S and J , respectively, after the P3_alpha split, where the mark “^” means the label of the new assigned group). The testing data (with “plus” sign at the upper-right side of the label) were then input into the built tree to estimate the accuracy rate of the tree system. The testing data indicate a 66% (2/3) accuracy rate for the S group, as shown in Fig. 3. One member of the testing S’ group was misassigned to the J group after P3_alpha, and 2 members of S’ were correctly assigned to S after F8_theta. For the other two testing groups (J’ and N’), all J’ members were correctly assigned to J after the P3_alpha and Fp1_beta splitting nodes, and all N’ members were correctly assigned to N after the Fp1_theta splitting node.

Beta 0.023 ± 0.355 0.085 ± 0.928 -0.121 ± 0.234 F(2,1437) = 10.6, p < 0.001

Gamma -0.097 ± 0.682 0.092 ± 1.409 -0.282 ± 0.370 F(2,1437) = 20.5, p < 0.001

In order to get more details of the clustering information, Fig. 4 shows the map of the decision boundary for this trial on the illustrated plane according to the first two splitting nodes (T5_alpha and C3_alpha). In Fig. 4, the misclassified data are indicated in black. Several subjects seem to be misclassified, but they are actually assigned to the right group if viewed in high dimension space.

Figure 4. Decision boundaries of CART according to schema in Fig. 3. S, J, and N denote whole data sets for senior, junior, and inexperienced meditators. The filled markers denote misclassified data of training set.

After 10-fold cross validation, the average accuracy rate of the three groups was 79% (24+22+25/90). The associated confusion matrix is given in Table 2. Subsequently, the distribution of the reorganized groups was evaluated by ANOVA. The ANOVA results reveal significant group differences in all EEG frequency bands (theta [F(2,1437) = 187.9, p < 0.001], alpha [F(2,1437) = 148.2, p < 0.001], beta [F(2,1437) = 10.7, p < 0.001], gamma [F(2,1437) = 16.7, p < 0.001]). Figure 5 shows the reorganized group variants in the high-frequency bands and it indicates impaired expression in beta band between S and J groups, and also in gamma band between J and N groups. The details of each reorganized group from CART for each frequency band are presented in Table 3. 3.3. SVM classification

Figure 3. Demonstration of the decision tree from one certain round in 10-fold cross validation by CART algorithm. S’, J’ and N’ denote the training sets for three groups (senior, junior and un- experienced meditators); and S, J and N denote the testing set for 3 groups; and S , J and N denote the reorganized group labels for 3 groups. The numerical number indicates the cut-off value at each splitting node.

In the SVM system, a one-against-one strategy was implemented to classify the three groups. The classification procedure from one trial of 10-fold cross validation produced several curvy boundaries from three groups on the hyperplane domain. Next, the reorganized groups were identified by conjoining these boundaries. After the group boundary environment was set up via a series of training data, a set of testing data was used to test the accuracy rate. In this trial, the training and testing accuracy rate were 100%. In Fig. 6, a grouping map is projected on the T5_alpha and C3_alpha domains. Several subjects seem to be misassigned to a wrong category in Fig. 6, but they are actually assigned to the right group if viewed in high dimension space.


(a)

(b)

(c)

(d)

Figure 5. Data distribution of the three groups after CART reorganization ( S , J , and N ) and statistical comparisons with post-hoc t-test on every one of two groups pair in (a) theta band, (b) alpha band, (c) beta band, and (d) gamma band. (* denotes statistical significance at p < 0.001) Table 2. Confusion matrices of CART and SVM from 10-fold cross validation. S, J, and N denote the original group labels and S , J , and N denote the reorganized group labels.

S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 J1 J2 J3 J4 J5 J6 J7 J8 J9 J10 N1 N2 N3 N4 N5 N6 N7 N8 N9 N10

A1 J J S N J J -

CART A2 A3 N J J J S S N N N N J S S -

total

SVM A3 -

A1 -

A2 -

S :3 J :22 N :5

-

N -

-

S :2 J :3 N :25

J -

-

-

S :24 J :5 N :1

total

594

Figure 6. Decision regions of SVM classification with training data used for Fig. 4. S, J, and N denote whole data sets for senior, junior, and inexperienced meditators.

Over all, the estimated accuracy rate of the classifiers for 10-fold cross validation was 98%. The associated confusion matrix indicates that the misclassification rate is 3% in the J and N groups, as shown in Table 1. The signal distribution of the reorganized groups was evaluated by ANOVA. The histograms of signal distribution of the three groups are shown in Fig. 7. The ANOVA results reveal significant group differences in all EEG frequency bands (theta [F(2,1437) = 225.1, p < 0.001], alpha [F(2,1437) = 251.1, p < 0.001], beta

S :30 J :0 N :0

(a)

(b)

(c)

(d)

S :0 J :29 N :1

S :0 J :1 N :29

Figure 7. Data distribution of the three groups after SVM reorganization ( S , , J and N ) and statistical comparisons with post-hoc t-test on every pair in (a) theta band, (b) alpha band, (c) beta band, and (d) gamma band. (* denotes statistical significance at p < 0.001)

Table 3. Mean (SD) values of EEG features and p-values from one-way ANOVA among 3 groups reorganized using CART for each EEG frequency band. S J N

Theta

Alpha

Beta

Gamma

1.244 ± 1.323 0.550 ± 0.951 0.103 ± 0.284 F(2,1437) = 187.9, p < 0.001

0.352 ± 0.654 -0.265 ± 0.654 -0.690 ± 1.324 F(2,1437) = 148.2, p < 0.001

0.057 ± 0.628 0.060 ± 0.645 -0.124 ± 0.257 F(2,1437) = 10.7, p < 0.001

-0.025 ± 0.998 0.030 ± 1.114 -0.284 ± 0.382 F(2,1437) = 16.7, p < 0.001


595

Table 4. Mean (SD) values of EEG features and p-values from one-way ANOVA among 3 groups reorganized using SVM for each EEG frequency band. S J N

Theta 1.258±1.328 0.571±0.954 0.030±0.142 F(2,1437)=225.1, p < 0.001

Alpha 0.427±0.712 -0.205±0.503 -0.860±1.158 F(2,1437)=251.1, p < 0.001

[F(2,1437) = 26.8, p < 0.001], gamma [F(2,1437) = 20.5, p < 0.001]. The reorganized data reveal that the group differences between S and J are enhanced in high-frequency bands and the data also report statistical significance in the gamma band (p < 0.001) in the post-hoc test. Even though the post-hoc test reports that the statistical significance is not obvious, the lower p value indicates that the differentiation between S and J in the beta band is improved (from p = 0.12 to p = 0.07). Compared with CART’s results in Fig. 5, the distribution arranged by SVM is similar to the unclassified but in statistical analyses the group difference is obvious between S and J in the beta and gamma bands. The details of each reorganized group from SVM for each frequency band are presented in Table 4.

Beta 0.023±0.355 0.097±0.932 -0.132±0.225 F(2,1437)=26.8, p < 0.001

Gamma -0.097±0.682 0.118±1.411 -0.308±0.347 F(2,1437)=20.5, p < 0.001

presents higher stability and robustness even with changing amount features, so SVM is a reliable tool for classifying meditation experience.

3.4. Evaluation of algorithm robustness To evaluate the reproducibility of the system, the confusion matrices of the classifiers in 100 trials are summarized (100 times 10-fold cross validation). The results indicate that CART misclassifies some participants, i.e., 15% and 7% of the S group members are assigned to J and N , respectively; 8% and 9% of the J group members are assigned to S and N , respectively; 14% and 14% of the N group members are assigned to S and J , respectively. SVM performs much better, with 2% of S assigned to J , 4% of J assigned to N , and 4% of N assigned to J . To sum up, in the classification of the three groups, the average classification rate of SVM is over 95%, which is 20% higher than that of CART. The performance of the system was also evaluated using characteristic curves under gradual increased number of features. The used features (16 EEG channels × 4 frequency bands in total) were increased according to the Gini index from small to large Gini values. The accuracy rate with number of features is shown in Fig. 8. As the number of feature increases, the average accuracy rate (100 times 10-fold cross validation) is enhanced and the SD decreases for all classifiers. When the number of features is from 10 to 64, the overall accuracy rate improves from 89 ± 2.3% to 96 ± 2.0% for SVM and from 79 ± 3.3% to 79 ± 3.0% for CART.

4. Discussion In this study, the EEG responses triggered by affective visual stimuli were used to recognize the calmness of mind of participants with various levels of meditation experience. The results show that the meditation experience can be categorized according to its effect on emotional stability. SVM was found to be more effective in classifying the three groups than CART analysis. SVM had the highest accuracy rate (98% versus 79% for CART) after 10-fold cross validation. Moreover, SVM

Figure 8. Average accuracy variation with number of features picked according to the Gini index from small to large Gini values. The error bar denotes the SD of the estimated accuracy rate after 100 times 10-fold validation.

Decision accuracy was highly correlated to the rule of classification at each classifying step. In decision making process, CART provides specific value to identify groups, but it does not always assign the participant to the right group. As shown in Fig. 3, the first two splitting nodes can mostly classify the meditation experience. However, CART has flaw. Once a participant has been assigned to the wrong group at a certain node, there is no chance to change the pathway, and then the outcome will ruin the reliability of system. Therefore, the influence on group distribution of CART is less significant in high frequency performance (as shown in Fig. 5) and lower accuracy rate of classification in grouping (as shown in Table 2). Of course, this unstable phenomenon also creates higher fluctuation among test-retest process (as shown in Fig. 8). Unlike CART, SVM determines the categories based on signal characteristics. As a result, the boundaries decided by the signals are curvy on the edge of decision area of groups. SVM can reliably classify data for differentiating meditation experiences. Generally speaking, it is easy to tell the difference in emotional stability between experienced meditators and inexperienced meditators. Distinguishing the levels of meditation experience is more difficult. As mentioned in Frantzidis’s study [29,30], SVM is a methodology to infer emotional states of participants through neurophysiological signals. Furthermore, SVM could carry out its outstanding ability in solving this difficulty in identifying the experience of meditation. As expected, SVM had a higher accuracy rate than CART and SVM is more stable under various numbers of


features, as shown in Fig. 8. SVM is a more efficient classifier compared to CART in classifying meditation experience. The lower variation of the SVM accuracy rate demonstrates that SVM is a more assertive classifier in grouping data than is CART. Moreover, the responding curves show that CART accuracy rate is around 79% with the number of features above 10, but SVM accuracy rate is up to 96%. A possible reason for these differences is that CART separates the groups based on Gini impurity (Eq. (2)), which is highly sensitive to the performance of each feature, whereas SVM determines the groups based on the most prominent features among the members of the group (Eq. (3)). SVM provides a perception-sensitive technique to identify which category the data belong to in this study. However, there still exist other types of meditation practices (compassion meditation for example [44-47]). The different physiological responses induced by different meditation practice types, so we are not sure whether SVM can also work on other meditation practice types or not. Different meditation practice might induce different responses even though the subject is subjected to the same emotional stimuli. For example, several studies have investigated focused-attention tasks that elicit EEG activity in the theta and alpha frequency ranges [36,37,47], so it can be assumed that SVM can classify categories according to the frequency variation. However, compared with concentrative meditation, different physiological responses to compassion meditators induce much higher response in gamma frequency [45,46]. Therefore, a well-trained SVM system for a certain type of meditation practice might not be applicable to other meditation practices. Neuroimaging studies have been conducted on the effect of meditation on emotional management [17,18]. Several brain areas, including the amygdala and medial prefrontal and posterior cingulate cortices, indicate that the emotional managements of experienced meditators have improved [17]. Applying SVM to more than one-dimensional data (threedimensional + time functional MRI data for example) would be another challenge. Under such circumstances, a great motivation of understanding the networking of emotional management is aroused to further study on whether SVM can also undertake classification on image data. Authors are convinced that this significant breakthrough will lead us even deeper into the understanding of the brain function and possibly extend this field of research to becoming an alternative therapy in the future. This study adopted CART and SVM for classification of meditation experience through the responses to emotional stimuli. It provides comparable evidence that experienced meditators maintain the state of calmness of mind in data mining approach. The results of this study directly indicate that meditation can improve psychological well-being, as calmness of mind is a representation of emotional stability. Moreover, SVM also shows stable accuracy rate and robustness. SVM helps to classify meditation experience according to responses of participant to visual stimuli.

596

5. Conclusion This study compared CART and SVM classifiers in term of classifying meditation experience by physiological response to emotional visual stimuli. The results show that meditation improves emotional stability. Moreover, SVM greatly and precisely identifies the experience of meditation in terms of psychological responses.

Acknowledgments The authors are especially indebted to Dr. Chung-Yao Hsu of Kaohsiung Medical University and the staff of the Sleep Disorders Center at the affiliated Chung-Ho Memorial Hospital for their support of this study. The authors also thank the Tibetan Nyingmapa Kathok Organization for supplying the participants for our study. This work was supported by the National Science Council of Taiwan under grants NSC 1002221-E-006-160 and NSC 101-2221-E-006-221-MY3.

References [1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9] [10]

[11]

[12]

[13]

K. M. DeNeve and H. Cooper, “The happy personality: a meta-analysis of 137 personality traits and subjective wellbeing,” Psychol. Bull., 124: 197-229, 1998. P. Hills and M. Argyle, “Emotional stability as a major dimension of happiness,” Pers. Indiv. Differ., 31: 1357-1364, 2001. P. Hills and M. Argyle, “The Oxford happiness questionnaire: a compact scale for the measurement of psychological well-being,” Pers. Indiv. Differ., 33: 1073-1082, 2002. C. E. Brett, A. J. Gow, J. Corley, A. Pattie, J. M. Starr and I. J. Deary, “Psychosocial factors and health as determinants of quality of life in community-dwelling older adults,” Qual. Life Res., 21: 505-516, 2012. K. Bontempo and J. Napier, “Evaluating emotional stability as a predictor of interpreter competence and aptitude for interpreting,” Interpreting, 13: 85-105, 2011. T. A. Judge and J. E. Bono, “Relationship of core selfevaluations traits-self-esteem, generalized self-efficacy, locus of control, and emotional stability-with job satisfaction and job performance: a meta-analysis,” J. Appl. Psychol., 86: 80-92, 2001. T. A. Judge, A. E. M. Van Vianen and I. E. De Pater, “Emotional stability, core self-evaluations, and job outcomes: a review of the evidence and an agenda for future research,” Hum. Perform., 17: 325-346, 2004. P. H. Lysaker, M. D. Bell, E. Kaplan and G. Bryson, “Personality and psychosocial dysfunction in schizophrenia: the association of extraversion and neuroticism to deficits in work performance,” Psychiatry Res., 80: 61-68, 1998. B. B. Lahey, “Public health significance of neuroticism,” Am. Psychol., 64: 241-256, 2009. A. L. Bouhuys, F. Flentge, A. J. Oldehinkel and M. D. van den Berg, “Potential psychosocial mechanisms linking depression to immune function in elderly subjects,” Psychiatry Res., 127: 237-245, 2004. S. D. Barger and S. J. Sydeman, “Does generalized anxiety disorder predict coronary heart disease risk factors independently of major depressive disorder?” J. Affect. Disord., 88: 87-91, 2005. J. J. Arch and M. G. Craske, “Mechanisms of mindfulness: emotion regulation following a focused breathing induction,” Behav. Res. Ther., 44: 1849-1858, 2006. P. Broderick, “Mindfulness and coping with dysphoric mood: contrasts with rumination and distraction,” Cogn. Ther. Res., 29: 501-510, 2005.

597


[14] J. I. Schmidtke and W. Heller, “Personality, affect and EEG: predicting patterns of regional brain activity related to extraversion and neuroticism,” Pers. Indiv. Differ., 36: 717-732, 2004. [15] L. Aftanas and S. Golosheykin, “Impact of regular meditation practice on EEG activity at rest and during evoked negative emotions,” Int. J. Neurosci., 115: 893-909, 2005. [16] A. Sobolewski, E. Holt, E. Kublik and A. Wrobel, “Impact of meditation on emotional processing-a visual ERP study,” Neurosci. Res., 71: 44-48, 2011. [17] V. A. Taylor, J. Grant, V. Daneault, G. Scavone, E. Breton, S. Roffe-Vidal, J. Courtemanche, A. S. Lavarenne and M. Beauregard, “Impact of mindfulness on the neural responses to emotional pictures in experienced and beginner meditators,” Neuroimage, 57: 1524-1533, 2011. [18] A. Lutz, J. Brefczynski-Lewis, T. Johnstone and R. J. Davidson, “Regulation of the neural circuitry of emotion by compassion meditation: effects of meditative expertise,” PLoS One, 3: e1897, 2008. [19] A. B. Benevides, T. F. Bastos Filho and M. Sarcinelli Filho, “Pseudo-online classification of mental tasks using Kullback-Leibler symmetric divergence,” J. Med. Biol. Eng., 32: 411-416, 2012. [20] S. Arya, D. M. Mount, N. S. Netanyahu, R. Silverman, A. Y. Wu, “An optimal algorithm for approximate nearest neighbor searching in fixed dimensions,” J. ACM, 45: 891-923, 1998. [21] J. Khan, J. S. Wei, M. Ringner, L. H. Saal, M. Ladanyi, F. Westermann, F. Berthold, M. Schwab, C. R. Antonescu, C. Peterson and P. S. Meltzer, “Classification and diagnostic prediction of cancers using gene expression profiling and artificial neural networks,” Nat. Med., 7: 673-679, 2001. [22] S. C. Lemon, J. Roy, M. A. Clark, P. D. Friedmann and W. Rakowski, “Classification and regression tree analysis in public health: methodological review and comparison with logistic regression,” Ann. Behav. Med., 26: 172-181, 2003. [23] B. Rosenfeld and C. Lewis, “Assessing violence risk in stalking cases: a regression tree approach,” Law Hum. Behav., 29: 343357, 2005. [24] P. Melillo, N. De Luca, M. Bracale and L. Pecchia, “Classification tree for risk assessment in patients suffering from congestive heart failure via long-term heart rate variability,” IEEE J. Biomed. Health Inform., 17: 727-733, 2013. [25] L. Pecchia, P. Melillo and M. Bracale, “Remote health monitoring of heart failure with data mining via CART method on HRV features,” IEEE Trans. Biomed. Eng., 58: 800-804, 2011. [26] L. Pecchia, P. Melillo, M. Sansone and M. Bracale, “Discrimination power of short-term heart rate variability measures for CHF assessment,” IEEE Trans. Inf. Technol. Biomed., 15: 40-46, 2011. [27] R. J. Martis, U. R. Acharya, J. H. Tan, A. Petznick, R. Yanti, C. K. Chua, E. Y. Ng and L. Tong, “Application of empirical mode decomposition (emd) for automated detection of epilepsy using EEG signals,” Int. J. Neural Syst., 22: 1250027, 2012. [28] T. N. Lal, M. Schroder, T. Hinterberger, J. Weston, M. Bogdan, N. Birbaumer and B. Scholkopf, “Support vector channel selection in BCI,” IEEE Trans. Biomed. Eng., 51: 1003-1010, 2004. [29] M. Thulasidas, G. Cuntai and W. Jiankang, “Robust classification of EEG signal for brain-computer interface,” IEEE Trans. Neural Syst. Rehabil. Eng., 14: 24-29, 2006. [30] L. Chisci, A. Mavino, G. Perferi, M. Sciandrone, C. Anile, G. Colicchio and F. Fuggetta, “Real-time epileptic seizure prediction using AR models and support vector machines,” IEEE Trans. Inf. Technol. Biomed., 57: 1124-1132, 2010.

[31] E. Byvatov and G. Schneider, “Support vector machine applications in bioinformatics,” Appl. Bioinformatics, 2: 67-77, 2003. [32] C. A. Frantzidis, C. Bratsas, M. A. Klados, E. Konstantinidis, C. D. Lithari, A. B. Vivas, C. L. Papadelis, E. Kaldoudi, C. Pappas and P. D. Bamidis, “On the classification of emotional biosignals evoked while viewing affective pictures: an integrated datamining-based approach for healthcare applications,” IEEE Trans. Inf. Technol. Biomed., 14: 309-318, 2010. [33] C. A. Frantzidis, C. Bratsas, C. L. Papadelis, E. Konstantinidis, C. Pappas and P. D. Bamidis, “Toward emotion aware computing: an integrated approach using multichannel neurophysiological recordings and affective visual stimuli,” IEEE Trans. Inf. Technol. Biomed., 14: 589-597, 2010. [34] P. C. Petrantonakis and L. J. Hadjileontiadis, “Emotion recognition from EEG using higher order crossings,” IEEE Trans. Inf. Technol. Biomed., 14: 186-197, 2010. [35] P. J. Lang, M. M. Bradley and B. N. Cuthbert, International affective picture system (IAPS): affective ratings of pictures and instruction manual. (Technical Report A-8). University of Florida, 2008. [36] Y. Kubota, W. Sato, M. Toichi, T. Murai, T. Okada, A. Hayashi and A. Sengoku, “Frontal midline theta rhythm is correlated with cardiac autonomic activities during the performance of an attention demanding meditation procedure,” Brain Res. Cogn. Brain Res., 11: 281-287, 2001. [37] T. Takahashi, T. Murata, T. Hamada, M. Omori, H. Kosaka, M. Kikuchi, H. Yoshida and Y. Wada, “Changes in EEG and autonomic nervous activity during meditation and their association with personality traits,” Int. J. Psychophysiol., 55: 199-207, 2005. [38] D. Sabatinelli, M. M. Bradley, J. R. Fitzsimmons and P. J. Lang, “Parallel amygdala and inferotemporal activation reflect emotional intensity and fear relevance,” Neuroimage, 24: 1265-1270, 2005. [39] K. Coburn and M. Moreno, “Facts and artifacts in brain electrical activity mapping,” Brain Topogr., 1: 37-45, 1988. [40] M. Fatourechi, A. Bashashati, R. K. Ward and G. E. Birch, “EMG and EOG artifacts in brain computer interface systems: a survey,” Clin. Neurophysiol., 118: 480-494, 2007. [41] C. Cortes and V. Vapnik, “Support-vector networks,” Mach. Learn., 20: 273-297, 1995. [42] S. Knerr, L. Personnaz and G. Dreyfus, “Single-layer learning revisited: a stepwise procedure for building and training a neural network,” in: F. Soulié and J. Hérault (Ed.), Neurocomputing: algorithms, architectures and applications, Springer-Verlag, 41-50, 1990. [43] C. C. Chang and C. J. Lin, “LIBSVM: a library for support vector machines,” ACM Trans. Intell. Syst. Technol., 2: 1-27, 2011. [44] B. H. Chang, J. A. Dusek and H. Benson, “Psychobiological changes from relaxation response elicitation: long-term practitioners vs. novices,” Psychosomatics, 52: 550-559, 2011. [45] D. Lehmann, P. L. Faber, P. Achermann, D. Jeanmonod, L. R. R. Gianotti and D. Pizzagalli, “Brain sources of EEG gamma frequency during volitionally meditation-induced, altered states of consciousness, and experience of the self,” Psychiatry Res., 108: 111-121, 2001. [46] A. Lutz, L. L. Greischar, N. B. Rawlings, M. Ricard and R. J. Davidson, “Long-term meditators self-induce high-amplitude gamma synchrony during mental practice,” Proc. Natl. Acad. Sci. U. S. A., 101: 16369-16373, 2004. [47] S. Baijal and N. Srinivasan, “Theta activity and meditative states: spectral changes during concentrative meditation,” Cogn. Process., 11: 31-38, 2010.