c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 4 ( 2 0 1 4 ) 50–59

journal homepage: www.intl.elsevierhealth.com/journals/cmpb

Classification of intelligence quotient via brainwave sub-band power ratio features and artificial neural network A.H. Jahidin ∗ , M.S.A. Megat Ali, M.N. Taib, N.Md. Tahir, I.M. Yassin, S. Lias Faculty of Electrical Engineering, Universiti Teknologi MARA, 40450 Shah Alam, Selangor, Malaysia

a r t i c l e

i n f o

a b s t r a c t

Article history:

This paper elaborates on the novel intelligence assessment method using the brainwave

Received 6 February 2013

sub-band power ratio features. The study focuses only on the left hemisphere brainwave in

Received in revised form

its relaxed state. Distinct intelligence quotient groups have been established earlier from

21 January 2014

the score of the Raven Progressive Matrices. Sub-band power ratios are calculated from

Accepted 23 January 2014

energy spectral density of theta, alpha and beta frequency bands. Synthetic data have been generated to increase dataset from 50 to 120. The features are used as input to the artificial

Keywords:

neural network. Subsequently, the brain behaviour model has been developed using an

Intelligence quotient (IQ)

artificial neural network that is trained with optimized learning rate, momentum constant

EEG

and hidden nodes. Findings indicate that the distinct intelligence quotient groups can be

Sub-band power ratio

classified from the brainwave sub-band power ratios with 100% training and 88.89% testing

Artificial neural network (ANN)

accuracies. © 2014 Elsevier Ireland Ltd. All rights reserved.

White Gaussian noise

1.

Introduction

Cognitive performance, described in terms of intelligence is characterized by information processing capabilities of the brain [1]. The ability differs between individuals [2] and is influenced by time [3] and experiential learning [4]. The underlying model of intelligence has been established through the neural efficiency theory [5]. The theory states that brighter individuals are less mentally active than people with average intellectual ability [6]. This is achieved through focused use of resting brain and disuse of irrelevant brain area for good task performance [7]. The phenomenon is reflected by lower glucose metabolism in brighter individuals [8]. Although receiving criticism [9], the theory has been reinforced through a variety of studies employing different neurophysiological

∗

measurement methods and a broad range of cognitive task demands [10]. Mental processing is performed by the frontal cortex. It performs executive functions of the brain which include the ability to regulate emotion, anticipate and plan for the future, make rational decisions and shape behaviour towards attainment of motivational goals. The functionality of the frontal cortex can further be divided into the left and right hemispheres where different regions exhibits different cognitive ability. The left hemisphere is involved in sequential and logic processes. The right hemisphere however, is specialized for social interactions and emotional capabilities [11]. Psychometric tests have been established to gauge an individual’s ability on various aspects of intelligence. This would include the intelligence quotient (IQ), working memory capacity, non-reasoning tasks, problem solving, learning

Corresponding author. Tel.: +60 12 6194590. E-mail address: [email protected] (A.H. Jahidin). 0169-2607/$ – see front matter © 2014 Elsevier Ireland Ltd. All rights reserved. http://dx.doi.org/10.1016/j.cmpb.2014.01.016

51

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 4 ( 2 0 1 4 ) 50–59

[12], mental rotation [13], and verbal reasoning [14]. Majority of studies relating intelligence with brain activity have utilized problem solving and non-reasoning tasks to stimulate intense mental activity. Brain activities are measured using various measurement modalities such as the functional magnetic resonance imaging (fMRI), positron emission tomography (PET) [5], and the electroencephalogram (EEG) [15]. All the findings consistently support the neural efficiency hypothesis [5]. In non-reasoning tasks, results show that although no reasoning was required, information processing differs between individuals of varying intelligence [12]. The EEG is a recording of electrical fluctuations of the brain, commonly known as brain waves. The signal exhibits an irregular and continuous oscillating pattern with frequencies varying from 0.5 to 100 Hz. The frequency characteristics heavily rely on the intensity of mental activity in the brain. Distinct patterns can be observed under varying conditions through the delta, theta, alpha and beta waves [16]. Delta wave with very low frequency range 0.4–4 Hz prevails in deep sleep and is commonly a precursor in comatose condition. The theta waves, however, are present in light sleep and are often associated with emotions and creativity. The signal exhibits a frequency range 4–8 Hz [17]. In relaxed state, the rhythmic alpha wave with frequency range 8–12 Hz is prevalent. In the event of intense mental activity however, the alpha wave would be replaced by the beta wave that exhibit higher frequency of between 12 and 30 Hz [16]. In the past, characterization of brainwave features has taken numerous approaches. In relation with mental performance, several studies have utilized band power to observe the effects on the alpha and beta waves. It has been observed that during mental stimulation, a less intelligent individual would demonstrate higher beta with lower alpha power as compared to the brighter ones [7,13,15,18–22]. These have been confirmed through the power ratio method that looks into the theta, alpha and beta sub-band inter-relationships [23–25]. In addition, similar studies have also directed to brainrelated disorders such as the attention-deficit/hyperactivity disorder (ADHD) [26–28], chronic fatigue syndrome [29], and Alzheimer’s disease [30]. The artificial neural network (ANN) is a machine learning method that is inspired by the working of biological neurons in the brain. The technique enables supervised learning through back-propagation update procedure on the weights of neuron connections to minimize the output error [31–33]. ANN has established itself in a variety of biomedical applications, which include disease recognition [34–36], chromosome detection [37], physiological analysis and modelling [38,39]. Previous works on modelling of brain behaviour using EEG have looked into characterization of epilepsy [40] and vigilance level [38]. In addition, application of ANN on EEG has also been implemented for brain–computer interface [41]. Conventionally, intelligence assessments are only conducted using established psychometric tests such as the Wechsler Intelligence Scale and Raven Progressive Matrices (RPM). Recent studies in variety of neurophysiological, many researches were done to correlate intelligence with brain activity [6,13,42–45]. Evaluation of IQ based on scientific technique however, is new. Although there are attempts to estimate IQ using ANN, the study focuses on estimating IQ

Raw EEG signals

Pre-processing and filtering (Noise removal)

Generate synthetic data (30 dBSNR White Gaussian noise)

Feature extraction (ESD and power ratio)

Feature selection (Observation of significant pattern and removal of outliers)

Is dataset enough to train network?

No

Yes Develop and train ANN classifier Fig. 1 – Flow chart of research work.

from different psychometric tools instead of EEG brainwaves [46]. There is still a gap in knowledge connecting complete systematic approach on IQ measurement via intelligent technique and EEG brainwaves. Hence the paper presents an intelligent approach to evaluate IQ from resting EEG. The IQbrain behaviour model is developed using sub-band power ratio features from the left brain hemisphere in closed eyes condition and ANN. The use of resting EEG in closed eyes condition to classify IQ was based on a previous study which revealed that brain activity pertaining to intelligence can be distinguished in its resting state (tonic), and in the absence of cognitive demanding task [47–49]. It was also discovered that pre-task resting conditions reflected state differences within non-clinical young adults [50].

2.

Methods

This section explains extensively on the overall processes that were implemented for this research work. It consist of EEG data collection, pre-processing, feature extraction, feature selection, generation of synthetic data and finally development of brain behaviour model using ANN as illustrated in Fig. 1.

2.1.

EEG data collection and IQ test

Approval from the University Ethics Committee was obtained and all volunteers were required to complete the consent form

52

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 4 ( 2 0 1 4 ) 50–59

prior to data collection. A total of 50 healthy and drug-free university students (mean age/standard deviation (SD) = 23.9/3.5, ranging in age from 20 to 40 years) from various disciplines were involved in the study. The EEG signals were recorded from the left and right side of the prefrontal area in unipolar configuration using g-MOBIlab+ from g.tec Guger Technologies and analyzed offline in MATLAB 2010b. Gold-plated electrodes were placed at Fp1 (left) and Fp2 (right) side of the forehead which conforms to the International 10/20 EEG standard. The earlobes linked to Fpz at the centre of the forehead acts as reference montage. The sampling rate was set to 256 Hz. During recording, subjects were required to relax in comfortably seated position with eyes closed for 3 min. The subjects were then required to answer an RPM-based IQ test with scores ranging from 0 to 150. The RPM was designed to specifically gauge intelligence in terms of logical thinking and non-verbal task performance. The scores obtained were used to establish IQ groups as a benchmark. Subjects were then grouped into different IQ levels (high, medium and low) based on the mean and SD of the normally distributed IQ scores.

2.2. Signal pre-processing, feature extraction and feature selection It is well established that the left hemisphere is involved in sequential and logic processes. Hence, the study focuses on the features extracted from the left hemispheric brainwave that will be the input to the brain behaviour model. In order to obtain an artefact-free EEG, all data were pre-processed offline. Any signal component exceeding a range of +100 ␮V to −100 ␮V are considered as noise and thus rejected. The EEG signals were filtered using bandpass Hamming filter with cutoff frequencies at 0.5 Hz and 30 Hz. The signals were then divided into the delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz) and beta (13–30 Hz) [51]. This can be done by applying Fast Fourier Transform (FFT) algorithm to obtain its discrete Fourier transform (DFT) prior to the extraction of features [52–60]. Thus, a similar successful method which has been practiced by many other researchers in filtering of EEG signals was adopted in this work. The term of FFT is widely used instead of DFT. The FFT algorithm that computes the DFT of vector X can be mathematically expressed as in Eq. (1).

X(k) =

N  n=1

 x(n) exp

2(k − 1)(n − 1) −j N

 ,

1 ≤ k ≤ N,

(1)

where X(k) is the discrete Fourier coefficient, x(n) represents the input signal in the time domain, N is the length of input vector x(n), n is the sampling instances in time domain and k is the sampling instances in frequency domain. The estimation of non-parametric power spectral density (PSD) curve for the respective sub-band was then calculated by implementing FFT. It is capable of capturing the frequency content of a stochastic process and can be calculated using Welch method to estimate the power spectrum of a time sequence [51,61]. The signals were divided into 50% overlapping epochs, with length of 1024 and window size of 256. The processing method

involved implementation of Hamming windowing function. Finally, the energy spectral density (ESD), which refers to the energy distribution specific to each frequency band was obtained by computing the area of the PSD curve. Box plots of ESD feature were used to further analyze the significant patterns among IQ groups for every sub-band. The study further looks into the inter-relationship between the sub-band components using the power ratio technique. Theta, alpha and beta ratio can each be represented by Eqs. (2)–(4). Delta sub-band was not considered as a feature because it is highly correlated with deep sleep, not intelligence. Theta ratio =

 ˛+

(2)

Alpha ratio =

˛ ˛+ˇ

(3)

Beta ratio =

ˇ ˛+ˇ

(4)

where  is theta ESD sub-band, ˛ is alpha ESD sub-band and ˇ is beta ESD sub-band. Similarly to ESD, the power ratio feature for each sub-band is analyzed for any significant patterns.

2.3.

Synthetic EEG signals

This section explains on the technique applied to generate synthetic signals from the original EEG. This is necessary to enhance the training dataset that will increase classification accuracy [62,63]. EEG is a stochastic in nature. Thus, a random white Gaussian noise can be implemented to generate synthetic signal. The signal-to-noise ratio (SNR) needs to be carefully selected to obtain a synthetic version that is similar to the original ones. If the value of SNR is too low, the noise content will modify the characteristics of the original signal [64]. Hence, this will lead to misclassification by the ANN classifier [65]. For these reasons, 30 dBSNR was selected. Noise array, Vnoise , was generated from multiplication between random white Gaussian noise, Wnoise , and the noise voltage, Vattn , where Vattn is the attenuated voltage from SNRdB relationship. Formerly, the noise voltage was derived from the input of SNRdB [66]. At 30 dB, the noise power, Pnoise , was calculated via Pnoise = Psignal /SNR, where Psignal is the averaged power for the raw EEG signal. Vattn was computed through the square root of Pnoise . The synthetic EEG, Vsynt , was then obtained by adding the generated noise, Vnoise , to original EEG, VEEG . The relationship can be expressed by Eqs. (5) and (6). Vnoise = Wnoise × Vattn

(5)

Vsynt = VEEG + Vnoise

(6)

2.4.

Brain behaviour model using ANN

The brain behaviour model was developed using ANN. The network comprises of an input layer, a single hidden layer and an output layer. The sub-band power ratios were used as input to the neural network structure. Meanwhile, the network output corresponds to the IQ index values. Tangent sigmoid and

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 4 ( 2 0 1 4 ) 50–59

53

Fig. 2 – PSD estimation of (a) delta, (b) theta (c) alpha and (d) beta sub-bands.

linear activation function was implemented to the hidden and output layers, respectively. Prior to development of the model, data were randomized and divided into three datasets, in which 70% of the data were used for training, while 15% for testing and another 15% for validation [67]. The training set is used for network training, while the testing set is used to evaluate the generalization ability of the trained network. However, only the training set is used for updating the network weights and biases. Meanwhile, the validation set is needed to periodically monitor the generalization performance of the ANN during learning. When the validation error increases for a specified number of iterations, the training is stopped to avoid over-fitting. Error with respect to testing data subset is not monitored during training, but is quantified to assess the final performance of a trained ANN model [68]. These are standard practice as many researchers have used it in their work [47,50,67–73]. In the training stage, iterative backpropagation training was adopted via the Levenberg–Marquardt algorithm. Prior to training, optimization was required to obtain the number of hidden nodes, learning rate and momentum constant. The final model was selected based on the best accuracy with the lowest error in training, testing, and validation. In general, the error is quantified in the form of mean squared error (MSE) for inspection of error convergence [67,74–76]. Classification matrix was employed to gauge the performance of the model in terms of accuracy (Acc), sensitivity (Se) and precision (Pp). Each of the parameters can be expressed by Eqs. (7)–(9), where TP is true positive, TN is true negative, FP is false positive, and FN is false negative. Acc =

Se =





TP + TN TP + TN + FP + FN

TP TP + FN



× 100%



× 100%

(7)

(8)

Pp =



TP TP + FP



× 100%

3.

Results and discussion

3.1.

Development of three distinct IQ groups

(9)

The IQ scores from the EEG data were analyzed offline. Samples were divided into three distinct IQ groups based on the mean (96.9) and SD (24.8). 39 samples were grouped under Index 2 (medium IQ), six samples were grouped under Index 3 (high IQ), and five samples were grouped under Index 1 (low IQ). Due to the unbalanced distribution of samples in each group, the use of synthetic data is necessary to enhance the performance of classifier. Synthetic data were generated from the original dataset (N = 50), amounting to 40 data per group (N = 120). This fulfils the minimum statistical value of 30 samples per group [77] for further analysis.

3.2. Feature extraction, feature selection and synthetic EEG signal Fig. 2 depicted the estimated PSD plots for delta, theta, alpha and beta sub-bands. Results indicate that distinct behaviour can be clearly observed for the corresponding sub-band using Welch method. The box plots in Fig. 3(a)–(d) indicate pattern-less distribution of ESD features for all sub-band groups. In addition, three extreme outliers can be observed. Hence, the power ratio feature was implemented to find significant pattern which can be correlated with different IQ groups. Results shown in Fig. 4(a)–(c) indicate that significant pattern can be observed among the three IQ groups for theta, alpha and beta power ratios. The box plots illustrate distinct level of mean sub-band power ratio values for low, medium

54

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 4 ( 2 0 1 4 ) 50–59

Fig. 3 – Box plots of ESD for sub-band (a) delta, (b) theta, (c) alpha and (d) beta with respect to low, medium and high IQ groups. The study focuses on the left brain hemisphere in its relaxed state, N = 50 samples.

and high IQ groups respectively. The mean for low IQ group is higher compared to medium IQ group for sub-band theta ratio, followed by high IQ group. Similar pattern can be observed in the sub-band beta ratio. Contrariwise low IQ group displays the lowest mean for sub-band alpha ratio, followed by the medium IQ group. The high IQ group exhibits the highest mean sub-band alpha ratio. Hence, implementation of power ratio technique can indirectly overcome the presence of extreme outliers from the ESD features. The observed power ratios were then used as inputs to the ANN classifier. To increase the classification accuracy, the two outliers were removed and the original dataset were enhanced with synthetic data. The two outliers that were removed are still less than 10% of the original dataset [78]. By removing them, it will not affect

robustness of the ANN classifier [78–80]. Moreover, in order to obtain optimal results, the errors should be independent and normally distributed, which is not the case when the training data contain the outliers [74,78,81]. The outliers can be maintained if this work implements any robust backpropagation algorithm [81–84]. Fig. 5(a) shows the random array of white Gaussian noise, Wnoise which was added to the original EEG. Meanwhile, Fig. 5(b) displays the probability density function (PDF) of white Gaussian noise which is normally distributed with maximum peak at zero value and decays away from zero. Subsequently, Fig. 6(a) and (b) illustrate the original EEG signal, VEEG and its synthetic signal, Vsynt . For a clear view, these signals were plotted for 1000 sampling instances. However, the histogram was plotted using full signal range.

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 4 ( 2 0 1 4 ) 50–59

Fig. 4 – Box plots of (a) theta, (b) alpha and (c) beta power ratio with respect to low, medium and high IQ groups, N = 50 samples.

Fig. 5 – (a) Generated random array of white Gaussian noise and (b) its histogram.

55

56

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 4 ( 2 0 1 4 ) 50–59

Fig. 6 – (a) Original EEG and (b) its synthetic signal with 30 dBSNR white Gaussian noise.

Fig. 7 – Box plots of (a) theta, (b) alpha and (c) beta power ratios with respect to low, medium and high IQ groups after removed of outliers and enhancement with synthetic data (40 data in each IQ group, N = 120).

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 4 ( 2 0 1 4 ) 50–59

Table 1 – Parameters of final network. Parameters

Value

Input nodes Hidden nodes Output node Learning rate Momentum rate Error goal Epochs

3 7 1 0.6 0.8 0 1000

Table 2 – Classification accuracy and mean squared error in training, testing and validation of final network model. Training

Testing

Validation

100 0.01

88.89 0.05

100 0.02

Acc (%) MSE

Table 3 – Performance matrix parameters which are precision and sensitivity in training, testing and validation for three IQ groups. IQ group

1 2 3

Training

Testing

Pp (%)

Se (%)

Pp (%)

Se (%)

Pp (%)

Se (%)

100 100 100

100 100 100

83.3 100 88.9

100 60 100

100 100 100

100 100 100

Acknowledgements Authors would like to express gratitude to the Ministry of Higher Education, Malaysia for the financial support through the MyPhD scholarship and the Fundamental Research Grant Scheme (600-RMI/FRGS 5/3 (72/2012)). Our gratitude extends also to the reviewers for valuable comments and constructive criticism.

references

Brain behaviour model

Learning rate was optimized by setting a constant value of momentum at 50% and 5 hidden nodes. Learning rate was varied from 0 to 1 with step size of 0.1. In order to optimize the momentum constant, number of hidden nodes was set to 5 and optimized learning rate at 0.6. Next, the number of hidden nodes was optimized by setting the learning rate and momentum constant to optimized values of 0.6 and 0.8, respectively. 7 hidden nodes were selected because it yields better accuracy in training. Based on the training results, the ANN architecture and training parameters is finalized in Table 1. Subsequently, Table 2 tabulates the summary of classification performance and error of final model with the optimized parameters. It has been discovered that final network gives the best training accuracy of 100%, testing at 88.89%, and validation at 100%. Error is minimal for training, testing and validation. The sensitivity and precision of the final network corresponding to three IQ groups are as summarized in Table 3.

4.

compared to the low IQ group. In contrast for theta and beta sub-bands, the ratio is low in high IQ group but higher in low IQ group. The results confirm the neural efficiency theory that states brighter individuals are less mentally active than the less intelligent people. Experimental results also revealed that sub-band power ratio from the left hemisphere can be classified using ANN with 100% training accuracy and 88.89% testing accuracy. Results also indicate that analysis on the left hemisphere from the frontal region is adequate for IQ recognition. As a conclusion, results show good classification performance with optimized network parameters in the ANN. In the future, comparison with other modelling techniques can be considered by employing different structure of neural network such as support vector machine (SVM), radial basis function (RBF) or incorporating fuzzy reasoning technique with neural network. In addition, further work using power ratio feature and ANN model may include cross-disciplinary study which can correlate IQ with other cognitive-related measures such as brain asymmetry, learning style and emotion.

Validation

Results show that similar pattern in the power ratio distribution can be observed from the box plots. The addition of synthetic data has improved on the distribution and are outliers-free, as shown in Fig. 7.

3.3.

57

Conclusion

The critical findings of this work revealed that sub-band power ratio features from the left hemisphere can be correlated with IQ. Hence, it can be applied as a form of feature to recognize the IQ groups. Alpha ratio is higher in high IQ group

[1] R.H. Grabner, A.C. Neubauer, E. Stern, Superior performance and neural efficiency: the impact of intelligence and expertise, Brain Research Bulletin 69 (2006) 422–439. [2] A.C. Neubauer, R.H. Grabner, H.H. Freudenthaler, J.F. Beckmann, J. Guthke, Intelligence and individual differences in becoming neutrally efficient, Acta Psychologica 116 (2004) 55–74. [3] O. Must, J. te Nijenhuis, A. Must, A.E.M. van Vianen, Comparability of IQ scores over time, Intelligence 37 (2009) 25–33. [4] S.B. Kaufman, C.G. DeYoung, J.R. Gray, J. Brown, N. Mackintosh, Associative learning predicts intelligence above and beyond working memory and processing speed, Intelligence 37 (2009) 374–382. [5] A.C. Neubauer, A. Fink, Intelligence and neural efficiency, Neuroscience & Biobehavioral Reviews 33 (2009) 1004–1023. [6] A.C. Neubauer, A. Fink, Intelligence and neural efficiency: measures of brain activation versus measures of functional connectivity in the brain, Intelligence 37 (2009) 223–229. [7] A. Fink, R.H. Grabner, M. Benedek, G. Reishofer, V. Hauswirth, M. Fally, C. Neuper, F. Ebner, A.C. Neubauer, The creative brain: investigation of brain activity during creative problem solving by means of EEG and fMRI, Human Brain Mapping 30 (2009) 734–748. [8] R.J. Haier, Neuro-intelligence, neuro-metrics and the next phase of brain imaging studies, Intelligence 37 (2009) 121–123. [9] D.L. Robinson, A test of the hendrickson postulate that reduced EEG response variance causes increased aep contour length: implications for the ‘neural transmission

58

[10]

[11]

[12]

[13]

[14]

[15]

[16] [17]

[18]

[19]

[20]

[21]

[22]

[23]

[24]

[25]

[26]

[27]

[28]

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 4 ( 2 0 1 4 ) 50–59

errors’ theory of intelligence, Personality and Individual Differences 22 (1997) 173–182. U. Fidelman, Neural transmission-errors, cerebral arousability and hemisphericity: some relations with intelligence and personality, Kybernetes 28 (1999) 695–725. L.J. van der Knaap, I.J.M. van der Ham, How does the corpus callosum mediate interhemispheric transfer? A review, Behavioural Brain Research 223 (2011) 211–221. N. Jauˇsovec, K. Jauˇsovec, Differences in induced brain activity during the performance of learning and working-memory tasks related to intelligence, Brain and Cognition 54 (2004) 65–74. ˇ ´ S. Katina, Induced EEG alpha oscillations are I. Riecansk y, related to mental rotation ability: the evidence for neural efficiency and serial processing, Neuroscience Letters 482 (2010) 133–136. A.F. Primrose, M. Fuller, M. Littledyke, Verbal reasoning test scores and their stability over time, Educational Research 42 (2000) 167–174. S. Micheloyannis, E. Pachou, C.J. Stam, M. Vourkas, S. Erimaki, V. Tsirka, Using graph theoretical analysis of multi channel EEG to evaluate the neural efficiency hypothesis, Neuroscience Letters 402 (2006) 273–277. J.W. Clark Jr., Medical Instrumentation: Application and Design, Wiley, New York, 2010. T.L. Huang, C. Charyton, A comprehensive review of the psychological effects of brainwave entrainment, Alternative Therapies in Health and Medicine 14 (2008) 38–50. R. Grabner, A. Fink, A. Stipacek, C. Neuper, A. Neubauer, Intelligence and working memory systems: evidence of neural efficiency in alpha band ERD, Cognitive Brain Research 20 (2004) 212–225. N. Jauˇsovec, Differences in cognitive processes between gifted, intelligent, creative, and average individuals while solving complex problems: an EEG study, Intelligence 28 (2000) 213–237. R. Thatcher, D. North, C. Biver, EEG and intelligence: relations between EEG coherence, EEG phase delay and power, Clinical Neurophysiology 116 (2005) 2129–2141. M. Doppelmayr, W. Klimesch, W. Stadler, D. Pöllhuber, C. Heine, EEG alpha power and intelligence, Intelligence 30 (2002) 289–302. T. Harmony, T. Fernández, J. Silva, J. Bernal, L. Díaz-Comas, A. Reyes, E. Marosi, M. Rodríguez, EEG delta activity: an indicator of attention to internal processing during performance of mental tasks, International Journal of Psychophysiology 24 (1996) 161–171. A.T. Pope, E.H. Bogart, D.S. Bartolome, Biocybernetic system evaluates indices of operator engagement in automated task, Biological Psychology 40 (1995) 187–195. O.M. Bazanova, Alpha EEG activity depends on the individual dominant rhythm frequency, Journal of Neurotherapy 16 (2012) 270–284. T. Egner, J.H. Gruzelier, The temporal dynamics of electroencephalographic responses to alpha/theta neurofeedback training in healthy subjects, Journal of Neurotherapy 8 (2004) 43–57. A.R. Clarke, R.J. Barry, EEG activity in subtypes of attention-deficit/hyperactivity disorder, Journal of Neurotherapy 8 (2005) 43–62. M.M. Lansbergen, M. Arns, M. van Dongen-Boomsma, D. Spronk, J.K. Buitelaar, The increase in theta/beta ratio on resting-state EEG in boys with attention-deficit/hyperactivity disorder is mediated by slow alpha peak frequency, Progress in Neuro-Psychopharmacology and Biological Psychiatry 35 (2011) 47–52. S.M. Bresnahan, R.J. Barry, A.R. Clarke, S.J. Johnstone, Quantitative EEG analysis in dexamphetamine-responsive

[29]

[30]

[31]

[32]

[33]

[34]

[35]

[36]

[37]

[38]

[39]

[40]

[41]

[42] [43]

[44]

[45]

[46]

adults with attention-deficit/hyperactivity disorder, Psychiatry Research 141 (2006) 151–159. K.M. Billiot, T.H. Budzynski, F. Andrasik, E.E.G Patterns, Chronic fatigue syndrome, Journal of Neurotherapy 2 (1997) 20–30. A.A. Fingelkurts, Short-term EEG spectral pattern as a single event in EEG phenomenology, Open Neuroimaging Journal 4 (2010) 130. A.E. Zadeh, A. Khazaee, V. Ranaee, Classification of the electrocardiogram signals using supervised classifiers and efficient features, Computer Methods and Programs in Biomedicine 99 (2010) 179–194. F. Ibrahim, M.N. Taib, W.A.B.W. Abas, C.C. Guan, S. Sulaiman, A novel dengue fever (DF) and dengue haemorrhagic fever (DHF) analysis using artificial neural network (ANN), Computer Methods and Programs in Biomedicine 79 (2005) 273–281. ˘ M.F. Talu, M. Gül, N. Alpaslan, B. Yigitcan, Calculation of melatonin and resveratrol effects on steatosis hepatis using soft computing methods, Computer Methods and Programs in Biomedicine 111 (2013) 498–506. R. Alizadehsani, J. Habibi, M.J. Hosseini, H. Mashayekhi, R. Boghrati, A. Ghandeharioun, B. Bahadorian, Z.A. Sani, A data mining approach for diagnosis of coronary artery disease, Computer Methods and Programs in Biomedicine 111 (2013) 52–61. J. Kuruvilla, K. Gunavathi, Lung cancer classification using neural networks for CT images, Computer Methods and Programs in Biomedicine 113 (2014) 202–209. J.L.M. Amaral, A.J. Lopes, J.M. Jansen, A.C.D. Faria, P.L. Melo, Machine learning algorithms and forced oscillation measurements applied to the automatic identification of chronic obstructive pulmonary disease, Computer Methods and Programs in Biomedicine 105 (2012) 183–193. E. Poletti, E. Grisan, A. Ruggeri, A modular framework for the automatic classification of chromosomes in Q-band images, Computer Methods and Programs in Biomedicine 105 (2012) 120–130. M. Duta, C. Alford, S. Wilson, L. Tarassenko, Neural network analysis of the mastoid EEG for the assessment of vigilance, International Journal of Human-Computer Interaction 17 (2004) 171–195. J. Fernandez de Canete, S. Gonzalez-Perez, J.C. Ramos-Diaz, Artificial neural networks for closed loop control of in silico and ad hoc type 1 diabetes, Computer Methods and Programs in Biomedicine 106 (2012) 55–66. A. Subasi, E. Erc¸elebi, Classification of EEG signals using neural network and logistic regression, Computer Methods and Programs in Biomedicine 78 (2005) 87–99. M. Aberg, J. Wessberg, Evolutionary optimization of classifiers and features for single-trial EEG discrimination, BioMedical Engineering OnLine 6 (2007) 32. N. Jauˇsovec, Differences in EEG alpha activity related to giftedness, Intelligence 23 (1996) 159–173. A. Gevins, M.E. Smith, Neurophysiological measures of working memory and individual differences in cognitive ability and cognitive style, Cerebral Cortex 10 (2000) 829–839. I.J. Deary, L. Penke, W. Johnson, The neuroscience of human intelligence differences, Nature Reviews Neuroscience 11 (2010) 201–211. B. Staudt, A.C. Neubauer, Achievement, underachievement and cortical activation: a comparative EEG study of adolescents of average and above-average intelligence, High Ability Studies 17 (2006) 3–16. A.G. Di Nuovo, S.D. Nuovo, S. Buono, Intelligent quotient estimation of mental retarded people from different psychometric instruments using artificial neural networks, Artificial Intelligence in Medicine 54 (2012) 135–145.

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 4 ( 2 0 1 4 ) 50–59

[47] N. Langer, A. Pedroni, L.R.R. Gianotti, J. Hänggi, D. Knoch, L. Jäncke, Functional brain network efficiency predicts intelligence, Human Brain Mapping 33 (2012) 1393–1406. [48] L. Wang, M. Song, T. Jiang, Y. Zhang, C. Yu, Regional homogeneity of the resting-state brain activity correlates with individual intelligence, Neuroscience Letters 488 (2011) 275–278. [49] T.W. Lee, Y.T. Wu, Y.W.Y. Yu, H.C. Wu, T.J. Chen, A smarter brain is associated with stronger neural interaction in healthy young females: a resting EEG coherence study, Intelligence 40 (2011) 38–48. [50] E. Angelakis, J.F. Lubar, S. Stathopoulou, Electroencephalographic peak alpha frequency correlates of cognitive traits, Neuroscience Letters 371 (2004) 60–63. [51] B. Hosseinifard, M.H. Moradi, R. Rostami, Classifying depression patients and normal subjects using machine learning techniques and nonlinear features from EEG signal, Computer Methods and Programs in Biomedicine 109 (2013) 339–345. [52] G. Bettus, F. Wendling, M. Guye, L. Valton, J. Régis, P. Chauvel, F. Bartolomei, Enhanced EEG functional connectivity in mesial temporal lobe epilepsy, Epilepsy Research 81 (2008) 58–68. [53] D. Abásolo, R. Hornero, C. Gómez, M. García, M. López, Analysis of EEG background activity in Alzheimer’s disease patients with Lempel–Ziv complexity and central tendency measure, Medical Engineering & Physics 28 (2006) 315–322. [54] F. Howells, V. Ives-Deliperi, N. Horn, D. Stein, Mindfulness based cognitive therapy improves frontal control in bipolar disorder: a pilot EEG study, BMC Psychiatry 12 (2012) 15. ´ [55] J. Kaminski, A. Brzezicka, A. Wróbel, Short-term memory capacity (7 ± 2) predicted by theta to gamma cycle length ratio, Neurobiology of Learning and Memory 95 (2011) 19–23. [56] A.J. Tomarken, R.J. Davidson, R.E. Wheeler, L. Kinney, Psychometric properties of resting anterior EEG asymmetry: temporal stability and internal consistency, Psychophysiology 29 (1992) 576–592. [57] I.H. Gotlib, EEG alpha asymmetry, depression, and cognitive functioning, Cognition & Emotion 12 (1998) 449–478. [58] K. Polat, S. Günes¸, Classification of epileptiform EEG using a hybrid system based on decision tree classifier and fast Fourier transform, Applied Mathematics and Computation 187 (2007) 1017–1026. [59] J.A. Cannon, P.A. Krokhmal, R.V. Lenth, R. Murphey, An algorithm for online detection of temporal changes in operator cognitive state using real-time psychophysiological data, Biomedical Signal Processing and Control 5 (2010) 229–236. [60] U. Rajendra Acharya, S. Vinitha Sree, G. Swapna, R.J. Martis, J.S. Suri, Automated EEG analysis of epilepsy: a review, Knowledge-Based Systems 45 (2013) 147–165. [61] P. Welch, The use of fast Fourier transform for the estimation of power spectra: a method based on time averaging over short, modified periodograms, IEEE Transactions on Audio and Electroacoustics 15 (1967) 70–73. [62] F. Lotte, Generating artificial EEG signals to reduce BCI calibration time, in: 5th International Brain-Computer Interface Workshop, 2011, pp. 176–179. [63] H. Hoessler, C. Wöhler, F. Lindner, U. Kreßel, Classifier training based on synthetically generated samples, in: Proceedings of the 5th International Conference on Computer Vision Systems (ICCV), 2007. [64] M.J. Roberts, Signals and Systems, McGraw-Hill, New York, 2004. [65] H. Azami, K. Mohammadi, B. Bozorgtabar, An improved signal segmentation using moving average and Savitzky–Golay filter, Journal of Signal and Information Processing 3 (2012) 39–44.

59

[66] L. Sevgi, Synthetic radar-signal environment: computer generation of signal, noise and clutter, IEEE Antennas and Propagation Magazine 49 (2007) 192–198. ´ R. Doroslovacki, ˇ [67] A. Kupusinac, E. Stokic, Predicting body fat percentage based on gender, age and BMI by using artificial neural networks, Computer Methods and Programs in Biomedicine 13 (2013) 610–619. [68] J. Farifteh, F. Van der Meer, C. Atzberger, E.J.M. Carranza, Quantitative analysis of salt-affected soil reflectance spectra: a comparison of two adaptive methods (PLSR and ANN), Remote Sensing of Environment 110 (2007) 59–78. [69] C.W. Anderson, E.A. Stolz, S. Shamsunder, Multivariate autoregressive models for classification of spontaneous electroencephalographic signals during mental tasks, IEEE Transactions on Biomedical Engineering 45 (1998) 277–286. [70] J.T. Wei, Z. Zhang, S.D. Barnhill, K.R. Madyastha, H. Zhang, J.E. Oesterling, Understanding artificial neural networks and exploring their potential applications for the practicing urologist, Urology 52 (1998) 161–172. [71] B.O. Peters, G. Pfurtscheller, H. Flyvbjerg, Automatic differentiation of multichannel EEG signals, IEEE Transactions on Biomedical Engineering 48 (2001) 111–116. [72] G. Chanel, C. Rebetez, M. Bétrancourt, T. Pun, Emotion assessment from physiological signals for adaptation of game difficulty, IEEE Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans 41 (2011) 1052–1063. [73] N. Jauˇsovec, K. Jauˇsovec, Differences in resting EEG related to ability, Brain Topography 12 (2000) 229–240. [74] H.R. Maier, G.C. Dandy, Neural networks for the prediction and forecasting of water resources variables: a review of modelling issues and applications, Environmental Modelling & Software 15 (2000) 101–124. [75] B.M. Wilamowski, Neural network architectures and learning algorithms, IEEE Industrial Electronics Magazine 3 (2009) 56–63. [76] O. Faust, U.R. Acharya, J. Ma, L.C. Min, T. Tamura, Compressed sampling for heart rate monitoring, Computer Methods and Programs in Biomedicine 108 (2012) 1191–1198. [77] A. Field, Discovering Statistics Using SPSS, SAGE Publications, Singapore, 2009. [78] K. Bovis, S. Singh, J. Fieldsend, C. Pinder, Identification of masses in digital mammograms with MLP and RBF nets, in: Neural Networks, 2000. Proceedings of the IEEE-INNS-ENNS International Joint Conference on IJCNN 2000, 2000, pp. 342–347. [79] S.-Q. Shi, W.L. Maner, L.B. Mackay, R.E. Garfield, Identification of term and preterm labor in rats using artificial neural networks on uterine electromyography signals, American Journal of Obstetrics and Gynecology 198 (2008), 235.e231–235.e234. [80] C. Alexakis, H. Nyongesa, R. Saatchi, N. Harris, C. Davies, C. Emery, R. Ireland, S. Heller, Feature extraction and classification of electrocardiogram (ECG) signals related to hypoglycaemia, Computers in Cardiology 30 (2003) 537–540 (IEEE, 2003). [81] D.S. Chen, R.C. Jain, A robust backpropagation learning algorithm for function approximation, IEEE Transactions on Neural Networks 5 (1994) 467–479. [82] K. Liano, Robust error measure for supervised neural network learning with outliers, IEEE Transactions on Neural Networks 7 (1996) 246–250. [83] C.-C. Chuang, S.-F. Su, C.-C. Hsiao, The annealing robust backpropagation (ARBP) learning algorithm, IEEE Transactions on Neural Networks 11 (2000) 1067–1077. [84] L. Huang, B.-L. Zhang, Q. Huang, Robust interval regression analysis using neural networks, Fuzzy Sets and Systems 97 (1998) 337–347.