523764 research-article2014
EEGXXX10.1177/1550059414523764Clinical EEG and NeuroscienceErguzel et al.
Article
Feature Selection and Classification of Electroencephalographic Signals: An Artificial Neural Network and Genetic Algorithm Based Approach
Clinical EEG and Neuroscience 1–6 © EEG and Clinical Neuroscience Society (ECNS) 2014 Reprints and permissions: sagepub.com/journalsPermissions.nav DOI: 10.1177/1550059414523764 eeg.sagepub.com
Turker Tekin Erguzel1, Serhat Ozekes1, Oguz Tan2,3, and Selahattin Gultekin4
Abstract Feature selection is an important step in many pattern recognition systems aiming to overcome the so-called curse of dimensionality. In this study, an optimized classification method was tested in 147 patients with major depressive disorder (MDD) treated with repetitive transcranial magnetic stimulation (rTMS). The performance of the combination of a genetic algorithm (GA) and a back-propagation (BP) neural network (BPNN) was evaluated using 6-channel pre-rTMS electroencephalographic (EEG) patterns of theta and delta frequency bands. The GA was first used to eliminate the redundant and less discriminant features to maximize classification performance. The BPNN was then applied to test the performance of the feature subset. Finally, classification performance using the subset was evaluated using 6-fold cross-validation. Although the slow bands of the frontal electrodes are widely used to collect EEG data for patients with MDD and provide quite satisfactory classification results, the outcomes of the proposed approach indicate noticeably increased overall accuracy of 89.12% and an area under the receiver operating characteristic (ROC) curve (AUC) of 0.904 using the reduced feature set. Keywords genetic algorithm, artificial neural network, cordance, major depressive disorder, rTMS Received November 15, 2013; revised December 18, 2013; accepted January 18, 2014.
Introduction MDD is a chronic, relapsing and remitting illness. A large percentage of patients (30%-50%) fail to respond to an initial course of antidepressant treatment.1 For this reason, there is a clear need for methods that select the right treatment for the right patient.2 rTMS has been proposed as one such alternative,3 with its less invasive and painful process compared with electrical brain stimulation.4 With establishment of the efficacy of rTMS, there has been increased interest in finding potential predictors of clinical response. The value of clinical factors in predicting treatment outcomes in patients with MDD is very limited, and a shift toward biomarkers is noticeable. In light of this “personalized medicine” approach to depression, recently, both genetic and neuroimaging biomarkers have been explored, and are showing promising results in aiding treatment prediction using pretreatment measures.5 Studies have been conducted, mainly with neurophysiologic EEG6,7 and functional neuroimaging biomarkers,8,9 and have demonstrated a predictive effect of a change in frontal quantitative EEG (QEEG) cordance in theta and delta. Coutin-Churchman et al10 analyzed EEG data to compare normal subjects with those who had various mental disorders. The investigators found that a change in delta- or
theta- EEG power can be regarded as a specific sign of brain dysfunction.11 A considerable number of applications underline that the effects of antidepressant medication are physiologically detectable on EEG, and QEEG cordance is one auspicious biomarker used to predict treatment response that has generated research interest. In addition to its valuable contribution as a biomarker, EEG patterns, with optimized subsets using GAs, minimize the number of features while maximizing classification performance. In one study, a genetic neural mathematic method was applied to 2 problems of EEG channel selection and classification.12 In another study, the selection of relevant 1
Department of Computer Engineering, Faculty of Engineering and Natural Sciences, Uskudar University, Istanbul, Turkey 2 Department of Psychiatry, NPIstanbul Hospital, Istanbul, Turkey 3 Department of Psychology, Faculty of Humanities and Social Sciences, Uskudar University, Istanbul, Turkey 4 Department of Bioengineering, Faculty of Engineering and Natural Sciences, Uskudar University, Istanbul, Turkey Corresponding Author: Turker Tekin Erguzel, Uskudar University, Faculty of Engineering and Natural Sciences, Department of Computer Engineering, Altunizade Mah. Haluk Turksoy Sk. No:14, Uskudar, 34662, Istanbul, Turkey. Full-color figures are available online at http://eeg.sagepub.com
Downloaded from eeg.sagepub.com at GEORGIAN COURT UNIV on December 13, 2014
2
Clinical EEG and Neuroscience
features for EEG signal classification of schizophrenic patients also benefited from the optimization process.13 A combined GA and neural network structure was applied to find the minimal set of dominant features that were most efficient in classifying into 2 groups.14 Another study focused on optimal EEG feature selection using a GA for classification of imagination of hand movement.15 Several studies have presented feature selection methods, comparing the performance of each, and underlined the efficacy of a heuristic engine, the GA.16-18 The machinelearning paradigm has been applied using an artificial neural network (ANN) fed with EEG data to differentiate 3 classes of subjects: those with schizophrenia, those with depression, and healthy subjects.19 Some other studies have focused on the selection of relevant features for EEG signal classification of patients with various mental diagnoses and have yielded promising results when combining GAs and classification methods such as ANN, a support vector machine, and linear discriminant analysis.12-14,20 In this paper, an artificial intelligence approach combining a GA, for feature selection of slow bands with 6 electrodes, and a BPNN classification method is proposed using a baseline resting-state QEEG pattern.
Materials and Methods Participants This research was conducted at the Neuropsychiatry Istanbul Hospital to determine the ability of QEEG to classify patients with MDD as responders or nonresponders to rTMS before undergoing treatment. This retrospective study was formally approved by the local medical research ethics committee. Participants first visited a psychiatrist to determine whether they met the inclusion criteria. All subjects had abstained from psychotropic medications for ≥2 weeks before enrollment. Subjects with nonpsychotic depressive disorder, as defined by the International Statistical Classification of Diseases and Related Health Problems criteria, and determined using the 17-item Hamilton Depression Rating Scale (HAM-D; score > 14) were eligible. A total of 147 patients with MDD, resistant to medication treatment, completed the protocols and were examined for the study. Responder and nonresponder groups did not differ with respect to their psychopharmacologic treatment. To minimize potential confusing outcomes of pharmacologic withdrawal symptoms, all subjects were on a monotherapy regimen and received concurrent selective serotonin reuptake inhibitor antidepressant medications during their 3 weeks (20 sessions) of rTMS therapy. No patients were receiving lithium, mood stabilizers, or benzodiazepines. On the day before rTMS treatment a baseline clinical assessment was conducted by a psychiatrist using the HAM-D. Twice during the study patients were assessed clinically, neuropsychologically, and with QEEG. Laboratory studies (complete blood count, chemistry, and thyroid-stimulating hormone), urine toxicology screening, and electrocardiography were performed at screening, and subjects were medically stable before entry.
Patients with organic brain disorders, pacemakers, psychotic symptoms, dementia, delirium, substance-related disorders, cluster A or B axis II disorders, electroconvulsive therapy in the prior 6 months, history of craniotomy, skull fracture, seizures, significant neurologic illness, suicidal intent, planning, or attempts were ineligible.
EEG Recordings During pretreatment QEEG, subjects were at rest, eyes-closed and maximally alert, in a quiet room with subdued lighting. Technicians monitored the QEEG data and alerted subjects every minute as needed to avoid drowsiness. EEG was for 3 minutes of eyes-closed rest, using a Scan LT EEG amplifier and electrode cap (Compumedics, Charlotte, NC), at a sampling rate of 250 Hz. Nineteen sintered Ag/AgCl electrodes were positioned according to the 10-20 system, with binaural reference. EEG signals were received from 6 electrodes (Fp1, Fp2, F3, F4, F7, and F8) in slow bands (delta and theta). The raw EEG signal was filtered through a band-pass (0.15-30 Hz) before artifact elimination. Manually selected (minimum, 2 minutes) artifact-free EEG data with a minimum split-half reliability ratio of 0.95 and a minimum test-retest reliability ratio of 0.90 were used for cordance calculations. Fast Fourier Transformation was used to calculate absolute and relative power in each of 2 nonoverlapping frequency bands: delta (1-4 Hz) and theta (4-8 Hz). The EEG cordance method was initially developed by Leuchter et al21 to provide a measure with face validity for the detection of cortical deafferentation. They observed that often, the EEG signal over a white-matter lesion exhibited decreased absolute theta power but increased relative theta power, which they termed “discordant.” Therefore, EEG cordance combines both absolute and relative EEG power, and negative values of this measure (discordance), specifically in theta or beta, are believed to reflect low perfusion or metabolism, whereas positive values (concordance), specifically in alpha, are thought to reflect high perfusion or metabolism. In a subsequent study,22 the investigators further confirmed this by comparing EEG cordance with simultaneously recorded positron emission tomographic scans reflecting perfusion.
rTMS Session Procedures and Ratings rTMS was applied using the Magstim Super Rapid2 stimulator (Magstim, Whitland, United Kingdom) with a figure-of-eight shaped Air Film Coil in an open-label manner. rTMS intensity was set at 100% of the motor threshold, which was determined by visual inspection. Stimulations were given to the left prefrontal cortex, deemed to be located anterior to the cortical motor area of the abductor pollicis brevis of which the motor threshold was determined. The treatment schedule was 6 days each week, from Monday to Saturday, for 3 weeks. Stimulation at 25 Hz for 2 seconds was delivered 20 times with 30-second intervals. A full course comprised 1,000 magnetic pulses. Subjects were classified as responders if their HAM-D scores at 3 weeks showed ≥50% improvement over their pretreatment
Downloaded from eeg.sagepub.com at GEORGIAN COURT UNIV on December 13, 2014
3
Erguzel et al. scores. In this study, the HAM-D percentage change value was discretized into 2 values (or classes), corresponding to responders when ≥50%, and to nonresponders otherwise.23
BPNN ANN are appropriate for solving problems in analyzing biomedical signals, because of their variety of applicability and their ability to learn complex and nonlinear relations. ANNs are trained by example rather than rules. When used in the diagnosis of neuromuscular disorders, they are not affected by factors such as human fatigue, emotional state, and habituation. They are capable of rapid identification, analysis of conditions, and diagnosis in real time.24 The most frequently used training algorithm in classification problems is the BP algorithm, which was used in this work also. There are many different types and architectures of neural networks, varying fundamentally in how they learn, the details of which are well documented in the literature.25,26 A BPNN is a typical multilayer feed-forward network, trained according to the BP algorithm. A BPNN uses a parallel distributed processing approach, to handle both qualitative and quantitative knowledge. It has strong robustness, fault tolerance, and adaptability, and can fully approximate any complex nonlinear relationship.27 Because of these advantages, BPNNs are more appropriate for processing EEG data that are possibly noisy, unstable, and nonlinear. The architecture of the network is a layered, feed-forward neural network, and the information flows unidirectionally from input layer to output layer, through the hidden layer(s).28 Input data are received from 6 electrodes as QEEG cordance. Ten neurons were used in the hidden layer, and a sigmoid transfer function was used in each neuron because of its nonlinear behavior. The learning factor was set to 0.03 and the momentum factor to 0.2. To minimize the error between the model output and a reference value, the mean square error was used as the cost function, given in equation 1. The cost function is minimized by the GA. J(w ) =
1 N 2 ∑ ( yk − z k ) , N k =1
Table 1. Simulation Parameters of the GA. Parameter
Value
Population size Number of generations Selection scheme Crossover type Crossover probability Early stopping
10 20 Tournament, size 0.2 One point 0.5 2 generations without improvement Stochastic uniform Gaussian
Selection Mutation operator
Table 2. rTMS Treatment Responder Results Using the GA. Feature Selection Method None GA
Accuracy (%)
Sensitivity (%)
AUC
80.25 89.12
84.44 94.44
0.863 0.904
natural selection and genetics. Traditional gradient-based optimization techniques search for an optimal point in a multidimensional optimization surface by iteratively refining a single solution. On the other hand, the GA operates on a collection of candidate solutions in parallel, which is one of its main strengths. Through this approach, the GA has a higher probability of locating the global optimal point than the traditional techniques, which are more likely to get stuck at a local optimum around the initial guess. The parallel approach of the GA also makes it less sensitive to initial conditions.31 In this study, the GA was applied to features of all selected channels to reduce the dimension of feature vectors using the neural network evaluation function. Choosing a proper fitness function is very important for the effectiveness and efficiency of the GA. As the objective function, the GA aims to optimize the classification error. Other simulation parameters of the GA are also given in Table 1.
(1)
where yk is the output of the model and zk is the reference output.
Feature Selection With the GA Feature selection and dimension reduction29 are important steps in pattern recognition tasks. In this study, although the feature set was not excessive and provided satisfactory outcomes, using the most informative features increased the classification rate. Reducing the number of features also enabled the classifier to learn a more robust solution and achieved better generalization performance. To obtain an optimal subset of features, optimization algorithms were used. Algorithms for feature selection can be divided into 3 main groups.30 The GA is an adaptive heuristic search algorithm inspired by the laws of
Results In this study, classification of 90 responders and 57 nonresponders from 147 subjects was performed using feature selection and a neural network based on GAs. The features of the EEG data set were first optimized using the GA, and a reduced subset with fewer input parameters was used to obtain a better classification result. Six-fold cross-validation was performed to train and to test the classifier with stratified sampling. The classification results before and after the feature selection process with the ANN are given along with overall accuracy, sensitivity, and AUC in Table 2. The ROC curves for the compared approaches are plotted in Figure 1. After the frequency band and channel selection phase, the GA was used to reduce the feature dimension by considering the classification error as a fitness function. The contribution of
Downloaded from eeg.sagepub.com at GEORGIAN COURT UNIV on December 13, 2014
4
Clinical EEG and Neuroscience
Figure 1. ROC curves for ANN and ANN with GA.
the feature selection process to accuracy was quite satisfactory. The ANN classified responders and nonresponders with 89.12% overall accuracy and 94.44% sensitivity to detect responders. The AUC was also used to evaluate the performance of the GA. The AUC after feature selection increased from 0.863 to 0.904. The features selected by the optimization algorithm were Fp1, Fp2, F7, F8, and F3 for the theta frequency band; 7 features were eliminated from the 12-feature subset. Frontal electrical activity in the theta frequency band has been associated with the function of these structures, and previous research has linked pretreatment theta activity of the anterior cingulate with clinical response.2,5,32 The results of this study support these earlier findings, and the focus on the prefrontal region and theta frequency band for patients with MDD.
Discussion One goal of clinical research on MDD is to predict the response of MDD to rTMS therapy using QEEG cordance to enhance diagnostic accuracy. This is crucially important for proper medical treatment and for slowing the progress of the illness. In this study, an ANN-based model, combined with an optimization algorithm, was designed to decrease the feature set while increasing prediction accuracy. The ANN was first studied for slow bands using frontal pretreatment cordance. Classification performance was evaluated using 6-fold cross-validation, and satisfactory results were obtained. Sleep disturbances are an integral part of the diagnostic criteria for major depression, and sleep may provide biomarkers for treatment response to
antidepressant medication. Since the initial investigation of the therapeutic benefit of sleep deprivation in depression, many studies have attempted to understand the link between sleep and depression. Because abnormalities in slow-wave sleep are among the most consistent biologic markers of depression, it is plausible that the antidepressant effects of sleep deprivation are due to the effects on slow-wave homeostasis.33 Similar studies using slow bands of depression also underlined theta power change after rTMS treatment compared with delta, which might be associated with clinical and cognitive improvement.34-36 In previous research, similar methods have been used combining machine-learning and feature selection approaches to classify patients with MDD treated with rTMS. A study of 90 depressed patients treated with rTMS predicted nonresponders by combining P300 amplitude, prefrontal delta and beta cordance, and anterior individual alpha peak frequency biomarkers and obtained an AUC of 0.814.5 A nonlinear EEG study focused on nonlinear metrics using rTMS (average, 21 sessions) for 90 depressed patients and 17 healthy participants. The ROC curves for different discriminant analyses with nonlinear EEG metrics alone, and analysis using the alpha peak frequency measure produced an AUC of 0.697. The total discriminant analysis, with the nonlinear measures added, resulted in an increase in AUC from 0.814 to 0.835.37 In another study of 30 depressed patients, the power spectra of the alpha, beta, theta frequency bands and all EEG bands were used as features, showing that a support vector machine classifier, with a GA for feature selection, achieved accuracy of 88.6% in classifying patients with MDD.38 A similar study was also designed to
Downloaded from eeg.sagepub.com at GEORGIAN COURT UNIV on December 13, 2014
5
Erguzel et al. examine, in 108 patients with MDD, whether functional connectivity strengths of resting EEG could be potential biomarkers in predicting response at 8 weeks of treatment. ROC curves of the connectivity showed that frontotemporal connectivity for slow bands differentiated responders and nonresponders. The AUC for the theta frequency band was 0.667.39 A genetic neural mathematic method was also used to perform effective EEG channel selection and reduction and underlined the optimization process for neurologic patients with cerebral dysfunction. The study achieved 0.86 for response classification and 0.82 for the actual responding hand classification with a test data set using k-fold cross-validation.12 In our study, 147 patients with MDD were classified using a GA-based BPNN. The method achieved 89.12% responder prediction accuracy, and an AUC of 0.904 using a reduced feature set. Because the number of features used in this study did not include all frequency bands and electrodes, further evaluation is recommended for a much larger feature set before any definitive conclusions are drawn. The findings support the potential utility of GAs and ANNs in combination as clinical tools in rTMS therapy of MDD. Our results also help elucidate the complex relationship between rTMS and regional cerebral perfusion compared with stand-alone measures, representing a step forward in promoting the application of computer simulation results for determining the optimal clinical application of rTMS. Acknowledgment We thank NPIstanbul Hospital for providing the EEG data.
Declaration of Conflicting Interests The author(s) declared no conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding The author(s) received no financial support for the research, authorship, and/or publication of this article.
References 1. Trivedi MH, Morris DW, Grannemann BD, Mahadi S. Symptom clusters as predictors of late response to antidepressant treatment. J Clin Psychiatry. 2005;66(8);1064-1070. 2. Bares M, Brunovsky M, Novak T, et al. The change of prefrontal QEEG theta cordance as a predictor of response to bupropion treatment in patients who had failed to respond to previous antidepressant treatments. Eur Neuropsychopharmacol. 2010;20(7);459-466. 3. O’Reardon J, Solvason HB, Janicak PG, et al. Efficacy and safety of transcranial magnetic stimulation in the acute treatment of major depression: a multisite randomized controlled trial. Biol Psychiatry. 2007;62(11):1208–1216. 4. Im C, Lee C. Computer-aided performance evaluation of a multichannel transcranial magnetic stimulation system. IEEE Trans Magn. 2006;42;3803-3808. 5. Arns M, Drinkenburg W, Fitzgerald G, Kenemans J. Neurophysiological predictors of non-response to rTMS in depression. Brain Stimul. 2012;5(4);569-576.
6. Price G, Lee JW, Garvey C, Gibson N. Appraisal of sessional EEG features as a correlate of clinical changes in an rTMS treatment of depression. Clin. EEG Neurosci. 2008;39(3):131-138. 7. Micoulaud J, Richieri R, Cermolacce M, Loundou A, Lancon C, Vion-Dury J. Parieto-temporal alpha EEG band power at baseline as a predictor of antidepressant treatment response with repetitive transcranial magnetic stimulation: a preliminary study. J Affect Disord. 2012;137(1-3):156-160. 8. Kito S, Hasegawa T, Koga Y. Cerebral blood flow ratio of the dorsolateral prefrontal cortex to the ventromedial prefrontal cortex as a potential predictor of treatment response to transcranial magnetic stimulation in depression. Brain Stimul. 2012;5(4): 547-553. 9. Richieri R, Boyer L, Farisse J, et al. Predictive value of brain perfusion SPECT for rTMS response in pharmacoresistant depression. Eur J Nucl Med Mol Imaging. 2011;38(9):1715-1722. 10. Coutin-Churchman P, Añez Y, Uzcátegui M, et al. Quantitative spectral analysis of EEG in psychiatry revisited: drawing signs out of numbers in a clinical setting. Clin Neurophysiol. 2003;114(12):2294-2306. 11. Khodayari A. Diagnosis of psychiatric disorders using EEG data and employing a statistical decision model. Presented at: International Conference of the IEEE EMBS; Buenos Aires, Argentina; 2010. 12. Yang J, Singh H, Hines EL, et al. Channel selection and classification of electroencephalogram signals: an artificial neural network and genetic algorithm-based approach. Artif Intell Med. 2012;55(2):117-126. 13. Sabeti M, Boostani R, Katebi SD, Price GW. Selection of relevant features for EEG signal classification of schizophrenic patients. Biomed Signal Proc Contr. 2007;2(2):122-134. 14. Hyun TK, Kim BY, Park EH, et al. Computerized recognition of Alzheimer disease-EEG using genetic algorithms and neural network. Future Gen Comput Syst. 2005;21(7):1124-1130. 15. Chum P, Park S-M, Ko K-E, et al. Optimal EEG feature selection by genetic algorithm for classification of imagination of hand movement. Presented at: 38th Annual Conference on IEEE Industrial Electronics Society; Montreal, QC, Canada; 2012. 16. Ocak H. Optimal classification of epileptic seizures in EEG using wavelet analysis and genetic algorithm. Signal Proc. 2008;88(7):1858-1867. 17. Flotzinger D, Pregenzer M, Pfurtscheller G. Feature selection with distinction sensitive learning vector quantisation and genetic algorithms. Presented at: IEEE World Congress on Computational Intelligence; Beijing, China; 1994. 18. Zhiping H, Chen G, Chen C, Xu H, Zhang J. A new EEG feature selection method for self-paced brain-computer interface. Presented at: 10th International Conference on Intelligent Systems Design and Applications; Cairo, Egypt; 2010. 19. Li Y, Fan FY. Classification of schizophrenia and depres sion by EEG with ANNs. Conf Proc IEEE Eng Med Biol Soc. 2005;3:2679-2682. 20. Ito S, Mitsukura Y, Fukumi M, et al. An EEG feature detection system using the neural networks based on genetic algorithms. Presented at: IEEE International Symposium on Computational Intelligence in Robotics and Automation; Kobe, Japan; 2003. 21. Leuchter AF, Cook IA, Lufkin RB, et al. Cordance: a new method for assessment of cerebral perfusion and metabolism using quantitative electroencephalography. Neuroimage. 1994;1(3): 208-219.
Downloaded from eeg.sagepub.com at GEORGIAN COURT UNIV on December 13, 2014
6
Clinical EEG and Neuroscience
22. Leuchter AF, Uijtdehaage SH, Cook IA, O’Hara R, Mandelkern M. Relationship between brain electrical activity and cortical perfusion in normal subjects. Psychiatry Res. 1999;90(2):125-140. 23. Khodayari A, Reilly J, Hasey G, DeBruin H, MacCrimmon D. Using pre-treatment electroencephalography data to predict response to transcranial magnetic stimulation therapy for major depression. Presented at: 33rd Annual International Conference of the IEEE EMBS; Boston, MA; August 2011. 24. Micheli-Tzanakou E. Neural networks in biomedical signal processing. In: Bronzino JD, ed. The biomedical engineering handbook. Boca Raton, FL: CRC Press; 2000:426-430. 25. Fausett L. Fundamentals of neural networks: architectures, algorithms, and applications. Englewood Cliffs, NJ: Prentice Hall; 1994. 26. Haykin S. Neural networks: a comprehensive foundation. New York: Macmillan; 1994. 27. Hong Z. Recognition of epileptic EEG signals based on BP neural networks. Sci Technol Inform. 2009;35:18. 28. Lek S, Guegan J. Artificial neural networks as a tool in ecological modelling: an introduction. Ecol Model. 1999;120(2-3):65-73. 29. Yom-Tov E, Inbar GF. Feature selection for the classification of movements from single movement related potentials. IEEE Trans Neural Syst Rehabil Eng. 2002;10(3):170-177. 30. Blum AL, Langley P. Selection of relevant features and examples in machine learning. Artif Intell. 1997;97(1-2):245-271. 31. Yao X. Evolving artificial neural networks. Proc IEEE. 1999;87(9):694-713. 32. Khodayari-Rostamabad A, Hasey GM, Maccrimmon DJ, Reilly JP, de Bruin H. A pilot study to determine whether machine learn-
ing methodologies using pre-treatment electroencephalography can predict the symptomatic response to clozapine therapy. Clin Neurophysiol. 2010;121(12):1998-2006. 33. Landsness EC, Goldstein MR, Peterson MJ, Tononi G, Benca RM. Antidepressant effects of selective slow wave sleep deprivation in major depression: a high-density EEG investigation. J Psychiatr Res. 2011;45(8):1019-1026. 34. Leuchter A, Cook IA, Marangell LB, et al. Comparative effectiveness of biomarkers and clinical indicators for predicting outcomes of SSRI treatment in major depressive disorder: results of the BRITE-MD study. Psychiatry Res. 2009;169(2):124-131. 35. Cook IA. Midline and right frontal brain function as a physiologic biomarker of remission in major depression. Psychiatry Res. 2009;174(2):152-157. 36. Noda Y, Nakamura M, Saeki T, Inoue M, Iwanari H, Kasai K. Potentiation of quantitative electroencephalograms following prefrontal repetitive transcranial magnetic stimulation in patients with major depression. Neurosci Res. 2013;77(1-2):70-77. 37. Arns M, Cerquera A, Gutiérrez RM, Hasselman F, Freund JA. Non-linear EEG analyses predict non-response to rTMS treatment in major depressive disorder. Clin Neurophysiol. In press. 38. Hosseinifard B, Moradi MH, Rostami R. Classifying depression patients and normal subjects using machine learning techniques. Presented at: 19th Iranian Conference on Electrical Engineering; Tehran, Iran; 2011. 39. Lee TW, Wu YT, Yu YW, Chen MC, Chen TJ. The implication of functional connectivity strength in predicting treatment response of major depressive disorder: a resting EEG study. Psychiat Res. 2011;194(3);372-377.
Downloaded from eeg.sagepub.com at GEORGIAN COURT UNIV on December 13, 2014
EEGXXX10.1177/1550059414523764Clinical EEG and NeuroscienceErguzel et al.
Article
Feature Selection and Classification of Electroencephalographic Signals: An Artificial Neural Network and Genetic Algorithm Based Approach
Clinical EEG and Neuroscience 1–6 © EEG and Clinical Neuroscience Society (ECNS) 2014 Reprints and permissions: sagepub.com/journalsPermissions.nav DOI: 10.1177/1550059414523764 eeg.sagepub.com
Turker Tekin Erguzel1, Serhat Ozekes1, Oguz Tan2,3, and Selahattin Gultekin4
Abstract Feature selection is an important step in many pattern recognition systems aiming to overcome the so-called curse of dimensionality. In this study, an optimized classification method was tested in 147 patients with major depressive disorder (MDD) treated with repetitive transcranial magnetic stimulation (rTMS). The performance of the combination of a genetic algorithm (GA) and a back-propagation (BP) neural network (BPNN) was evaluated using 6-channel pre-rTMS electroencephalographic (EEG) patterns of theta and delta frequency bands. The GA was first used to eliminate the redundant and less discriminant features to maximize classification performance. The BPNN was then applied to test the performance of the feature subset. Finally, classification performance using the subset was evaluated using 6-fold cross-validation. Although the slow bands of the frontal electrodes are widely used to collect EEG data for patients with MDD and provide quite satisfactory classification results, the outcomes of the proposed approach indicate noticeably increased overall accuracy of 89.12% and an area under the receiver operating characteristic (ROC) curve (AUC) of 0.904 using the reduced feature set. Keywords genetic algorithm, artificial neural network, cordance, major depressive disorder, rTMS Received November 15, 2013; revised December 18, 2013; accepted January 18, 2014.
Introduction MDD is a chronic, relapsing and remitting illness. A large percentage of patients (30%-50%) fail to respond to an initial course of antidepressant treatment.1 For this reason, there is a clear need for methods that select the right treatment for the right patient.2 rTMS has been proposed as one such alternative,3 with its less invasive and painful process compared with electrical brain stimulation.4 With establishment of the efficacy of rTMS, there has been increased interest in finding potential predictors of clinical response. The value of clinical factors in predicting treatment outcomes in patients with MDD is very limited, and a shift toward biomarkers is noticeable. In light of this “personalized medicine” approach to depression, recently, both genetic and neuroimaging biomarkers have been explored, and are showing promising results in aiding treatment prediction using pretreatment measures.5 Studies have been conducted, mainly with neurophysiologic EEG6,7 and functional neuroimaging biomarkers,8,9 and have demonstrated a predictive effect of a change in frontal quantitative EEG (QEEG) cordance in theta and delta. Coutin-Churchman et al10 analyzed EEG data to compare normal subjects with those who had various mental disorders. The investigators found that a change in delta- or
theta- EEG power can be regarded as a specific sign of brain dysfunction.11 A considerable number of applications underline that the effects of antidepressant medication are physiologically detectable on EEG, and QEEG cordance is one auspicious biomarker used to predict treatment response that has generated research interest. In addition to its valuable contribution as a biomarker, EEG patterns, with optimized subsets using GAs, minimize the number of features while maximizing classification performance. In one study, a genetic neural mathematic method was applied to 2 problems of EEG channel selection and classification.12 In another study, the selection of relevant 1
Department of Computer Engineering, Faculty of Engineering and Natural Sciences, Uskudar University, Istanbul, Turkey 2 Department of Psychiatry, NPIstanbul Hospital, Istanbul, Turkey 3 Department of Psychology, Faculty of Humanities and Social Sciences, Uskudar University, Istanbul, Turkey 4 Department of Bioengineering, Faculty of Engineering and Natural Sciences, Uskudar University, Istanbul, Turkey Corresponding Author: Turker Tekin Erguzel, Uskudar University, Faculty of Engineering and Natural Sciences, Department of Computer Engineering, Altunizade Mah. Haluk Turksoy Sk. No:14, Uskudar, 34662, Istanbul, Turkey. Full-color figures are available online at http://eeg.sagepub.com
Downloaded from eeg.sagepub.com at GEORGIAN COURT UNIV on December 13, 2014
2
Clinical EEG and Neuroscience
features for EEG signal classification of schizophrenic patients also benefited from the optimization process.13 A combined GA and neural network structure was applied to find the minimal set of dominant features that were most efficient in classifying into 2 groups.14 Another study focused on optimal EEG feature selection using a GA for classification of imagination of hand movement.15 Several studies have presented feature selection methods, comparing the performance of each, and underlined the efficacy of a heuristic engine, the GA.16-18 The machinelearning paradigm has been applied using an artificial neural network (ANN) fed with EEG data to differentiate 3 classes of subjects: those with schizophrenia, those with depression, and healthy subjects.19 Some other studies have focused on the selection of relevant features for EEG signal classification of patients with various mental diagnoses and have yielded promising results when combining GAs and classification methods such as ANN, a support vector machine, and linear discriminant analysis.12-14,20 In this paper, an artificial intelligence approach combining a GA, for feature selection of slow bands with 6 electrodes, and a BPNN classification method is proposed using a baseline resting-state QEEG pattern.
Materials and Methods Participants This research was conducted at the Neuropsychiatry Istanbul Hospital to determine the ability of QEEG to classify patients with MDD as responders or nonresponders to rTMS before undergoing treatment. This retrospective study was formally approved by the local medical research ethics committee. Participants first visited a psychiatrist to determine whether they met the inclusion criteria. All subjects had abstained from psychotropic medications for ≥2 weeks before enrollment. Subjects with nonpsychotic depressive disorder, as defined by the International Statistical Classification of Diseases and Related Health Problems criteria, and determined using the 17-item Hamilton Depression Rating Scale (HAM-D; score > 14) were eligible. A total of 147 patients with MDD, resistant to medication treatment, completed the protocols and were examined for the study. Responder and nonresponder groups did not differ with respect to their psychopharmacologic treatment. To minimize potential confusing outcomes of pharmacologic withdrawal symptoms, all subjects were on a monotherapy regimen and received concurrent selective serotonin reuptake inhibitor antidepressant medications during their 3 weeks (20 sessions) of rTMS therapy. No patients were receiving lithium, mood stabilizers, or benzodiazepines. On the day before rTMS treatment a baseline clinical assessment was conducted by a psychiatrist using the HAM-D. Twice during the study patients were assessed clinically, neuropsychologically, and with QEEG. Laboratory studies (complete blood count, chemistry, and thyroid-stimulating hormone), urine toxicology screening, and electrocardiography were performed at screening, and subjects were medically stable before entry.
Patients with organic brain disorders, pacemakers, psychotic symptoms, dementia, delirium, substance-related disorders, cluster A or B axis II disorders, electroconvulsive therapy in the prior 6 months, history of craniotomy, skull fracture, seizures, significant neurologic illness, suicidal intent, planning, or attempts were ineligible.
EEG Recordings During pretreatment QEEG, subjects were at rest, eyes-closed and maximally alert, in a quiet room with subdued lighting. Technicians monitored the QEEG data and alerted subjects every minute as needed to avoid drowsiness. EEG was for 3 minutes of eyes-closed rest, using a Scan LT EEG amplifier and electrode cap (Compumedics, Charlotte, NC), at a sampling rate of 250 Hz. Nineteen sintered Ag/AgCl electrodes were positioned according to the 10-20 system, with binaural reference. EEG signals were received from 6 electrodes (Fp1, Fp2, F3, F4, F7, and F8) in slow bands (delta and theta). The raw EEG signal was filtered through a band-pass (0.15-30 Hz) before artifact elimination. Manually selected (minimum, 2 minutes) artifact-free EEG data with a minimum split-half reliability ratio of 0.95 and a minimum test-retest reliability ratio of 0.90 were used for cordance calculations. Fast Fourier Transformation was used to calculate absolute and relative power in each of 2 nonoverlapping frequency bands: delta (1-4 Hz) and theta (4-8 Hz). The EEG cordance method was initially developed by Leuchter et al21 to provide a measure with face validity for the detection of cortical deafferentation. They observed that often, the EEG signal over a white-matter lesion exhibited decreased absolute theta power but increased relative theta power, which they termed “discordant.” Therefore, EEG cordance combines both absolute and relative EEG power, and negative values of this measure (discordance), specifically in theta or beta, are believed to reflect low perfusion or metabolism, whereas positive values (concordance), specifically in alpha, are thought to reflect high perfusion or metabolism. In a subsequent study,22 the investigators further confirmed this by comparing EEG cordance with simultaneously recorded positron emission tomographic scans reflecting perfusion.
rTMS Session Procedures and Ratings rTMS was applied using the Magstim Super Rapid2 stimulator (Magstim, Whitland, United Kingdom) with a figure-of-eight shaped Air Film Coil in an open-label manner. rTMS intensity was set at 100% of the motor threshold, which was determined by visual inspection. Stimulations were given to the left prefrontal cortex, deemed to be located anterior to the cortical motor area of the abductor pollicis brevis of which the motor threshold was determined. The treatment schedule was 6 days each week, from Monday to Saturday, for 3 weeks. Stimulation at 25 Hz for 2 seconds was delivered 20 times with 30-second intervals. A full course comprised 1,000 magnetic pulses. Subjects were classified as responders if their HAM-D scores at 3 weeks showed ≥50% improvement over their pretreatment
Downloaded from eeg.sagepub.com at GEORGIAN COURT UNIV on December 13, 2014
3
Erguzel et al. scores. In this study, the HAM-D percentage change value was discretized into 2 values (or classes), corresponding to responders when ≥50%, and to nonresponders otherwise.23
BPNN ANN are appropriate for solving problems in analyzing biomedical signals, because of their variety of applicability and their ability to learn complex and nonlinear relations. ANNs are trained by example rather than rules. When used in the diagnosis of neuromuscular disorders, they are not affected by factors such as human fatigue, emotional state, and habituation. They are capable of rapid identification, analysis of conditions, and diagnosis in real time.24 The most frequently used training algorithm in classification problems is the BP algorithm, which was used in this work also. There are many different types and architectures of neural networks, varying fundamentally in how they learn, the details of which are well documented in the literature.25,26 A BPNN is a typical multilayer feed-forward network, trained according to the BP algorithm. A BPNN uses a parallel distributed processing approach, to handle both qualitative and quantitative knowledge. It has strong robustness, fault tolerance, and adaptability, and can fully approximate any complex nonlinear relationship.27 Because of these advantages, BPNNs are more appropriate for processing EEG data that are possibly noisy, unstable, and nonlinear. The architecture of the network is a layered, feed-forward neural network, and the information flows unidirectionally from input layer to output layer, through the hidden layer(s).28 Input data are received from 6 electrodes as QEEG cordance. Ten neurons were used in the hidden layer, and a sigmoid transfer function was used in each neuron because of its nonlinear behavior. The learning factor was set to 0.03 and the momentum factor to 0.2. To minimize the error between the model output and a reference value, the mean square error was used as the cost function, given in equation 1. The cost function is minimized by the GA. J(w ) =
1 N 2 ∑ ( yk − z k ) , N k =1
Table 1. Simulation Parameters of the GA. Parameter
Value
Population size Number of generations Selection scheme Crossover type Crossover probability Early stopping
10 20 Tournament, size 0.2 One point 0.5 2 generations without improvement Stochastic uniform Gaussian
Selection Mutation operator
Table 2. rTMS Treatment Responder Results Using the GA. Feature Selection Method None GA
Accuracy (%)
Sensitivity (%)
AUC
80.25 89.12
84.44 94.44
0.863 0.904
natural selection and genetics. Traditional gradient-based optimization techniques search for an optimal point in a multidimensional optimization surface by iteratively refining a single solution. On the other hand, the GA operates on a collection of candidate solutions in parallel, which is one of its main strengths. Through this approach, the GA has a higher probability of locating the global optimal point than the traditional techniques, which are more likely to get stuck at a local optimum around the initial guess. The parallel approach of the GA also makes it less sensitive to initial conditions.31 In this study, the GA was applied to features of all selected channels to reduce the dimension of feature vectors using the neural network evaluation function. Choosing a proper fitness function is very important for the effectiveness and efficiency of the GA. As the objective function, the GA aims to optimize the classification error. Other simulation parameters of the GA are also given in Table 1.
(1)
where yk is the output of the model and zk is the reference output.
Feature Selection With the GA Feature selection and dimension reduction29 are important steps in pattern recognition tasks. In this study, although the feature set was not excessive and provided satisfactory outcomes, using the most informative features increased the classification rate. Reducing the number of features also enabled the classifier to learn a more robust solution and achieved better generalization performance. To obtain an optimal subset of features, optimization algorithms were used. Algorithms for feature selection can be divided into 3 main groups.30 The GA is an adaptive heuristic search algorithm inspired by the laws of
Results In this study, classification of 90 responders and 57 nonresponders from 147 subjects was performed using feature selection and a neural network based on GAs. The features of the EEG data set were first optimized using the GA, and a reduced subset with fewer input parameters was used to obtain a better classification result. Six-fold cross-validation was performed to train and to test the classifier with stratified sampling. The classification results before and after the feature selection process with the ANN are given along with overall accuracy, sensitivity, and AUC in Table 2. The ROC curves for the compared approaches are plotted in Figure 1. After the frequency band and channel selection phase, the GA was used to reduce the feature dimension by considering the classification error as a fitness function. The contribution of
Downloaded from eeg.sagepub.com at GEORGIAN COURT UNIV on December 13, 2014
4
Clinical EEG and Neuroscience
Figure 1. ROC curves for ANN and ANN with GA.
the feature selection process to accuracy was quite satisfactory. The ANN classified responders and nonresponders with 89.12% overall accuracy and 94.44% sensitivity to detect responders. The AUC was also used to evaluate the performance of the GA. The AUC after feature selection increased from 0.863 to 0.904. The features selected by the optimization algorithm were Fp1, Fp2, F7, F8, and F3 for the theta frequency band; 7 features were eliminated from the 12-feature subset. Frontal electrical activity in the theta frequency band has been associated with the function of these structures, and previous research has linked pretreatment theta activity of the anterior cingulate with clinical response.2,5,32 The results of this study support these earlier findings, and the focus on the prefrontal region and theta frequency band for patients with MDD.
Discussion One goal of clinical research on MDD is to predict the response of MDD to rTMS therapy using QEEG cordance to enhance diagnostic accuracy. This is crucially important for proper medical treatment and for slowing the progress of the illness. In this study, an ANN-based model, combined with an optimization algorithm, was designed to decrease the feature set while increasing prediction accuracy. The ANN was first studied for slow bands using frontal pretreatment cordance. Classification performance was evaluated using 6-fold cross-validation, and satisfactory results were obtained. Sleep disturbances are an integral part of the diagnostic criteria for major depression, and sleep may provide biomarkers for treatment response to
antidepressant medication. Since the initial investigation of the therapeutic benefit of sleep deprivation in depression, many studies have attempted to understand the link between sleep and depression. Because abnormalities in slow-wave sleep are among the most consistent biologic markers of depression, it is plausible that the antidepressant effects of sleep deprivation are due to the effects on slow-wave homeostasis.33 Similar studies using slow bands of depression also underlined theta power change after rTMS treatment compared with delta, which might be associated with clinical and cognitive improvement.34-36 In previous research, similar methods have been used combining machine-learning and feature selection approaches to classify patients with MDD treated with rTMS. A study of 90 depressed patients treated with rTMS predicted nonresponders by combining P300 amplitude, prefrontal delta and beta cordance, and anterior individual alpha peak frequency biomarkers and obtained an AUC of 0.814.5 A nonlinear EEG study focused on nonlinear metrics using rTMS (average, 21 sessions) for 90 depressed patients and 17 healthy participants. The ROC curves for different discriminant analyses with nonlinear EEG metrics alone, and analysis using the alpha peak frequency measure produced an AUC of 0.697. The total discriminant analysis, with the nonlinear measures added, resulted in an increase in AUC from 0.814 to 0.835.37 In another study of 30 depressed patients, the power spectra of the alpha, beta, theta frequency bands and all EEG bands were used as features, showing that a support vector machine classifier, with a GA for feature selection, achieved accuracy of 88.6% in classifying patients with MDD.38 A similar study was also designed to
Downloaded from eeg.sagepub.com at GEORGIAN COURT UNIV on December 13, 2014
5
Erguzel et al. examine, in 108 patients with MDD, whether functional connectivity strengths of resting EEG could be potential biomarkers in predicting response at 8 weeks of treatment. ROC curves of the connectivity showed that frontotemporal connectivity for slow bands differentiated responders and nonresponders. The AUC for the theta frequency band was 0.667.39 A genetic neural mathematic method was also used to perform effective EEG channel selection and reduction and underlined the optimization process for neurologic patients with cerebral dysfunction. The study achieved 0.86 for response classification and 0.82 for the actual responding hand classification with a test data set using k-fold cross-validation.12 In our study, 147 patients with MDD were classified using a GA-based BPNN. The method achieved 89.12% responder prediction accuracy, and an AUC of 0.904 using a reduced feature set. Because the number of features used in this study did not include all frequency bands and electrodes, further evaluation is recommended for a much larger feature set before any definitive conclusions are drawn. The findings support the potential utility of GAs and ANNs in combination as clinical tools in rTMS therapy of MDD. Our results also help elucidate the complex relationship between rTMS and regional cerebral perfusion compared with stand-alone measures, representing a step forward in promoting the application of computer simulation results for determining the optimal clinical application of rTMS. Acknowledgment We thank NPIstanbul Hospital for providing the EEG data.
Declaration of Conflicting Interests The author(s) declared no conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding The author(s) received no financial support for the research, authorship, and/or publication of this article.
References 1. Trivedi MH, Morris DW, Grannemann BD, Mahadi S. Symptom clusters as predictors of late response to antidepressant treatment. J Clin Psychiatry. 2005;66(8);1064-1070. 2. Bares M, Brunovsky M, Novak T, et al. The change of prefrontal QEEG theta cordance as a predictor of response to bupropion treatment in patients who had failed to respond to previous antidepressant treatments. Eur Neuropsychopharmacol. 2010;20(7);459-466. 3. O’Reardon J, Solvason HB, Janicak PG, et al. Efficacy and safety of transcranial magnetic stimulation in the acute treatment of major depression: a multisite randomized controlled trial. Biol Psychiatry. 2007;62(11):1208–1216. 4. Im C, Lee C. Computer-aided performance evaluation of a multichannel transcranial magnetic stimulation system. IEEE Trans Magn. 2006;42;3803-3808. 5. Arns M, Drinkenburg W, Fitzgerald G, Kenemans J. Neurophysiological predictors of non-response to rTMS in depression. Brain Stimul. 2012;5(4);569-576.
6. Price G, Lee JW, Garvey C, Gibson N. Appraisal of sessional EEG features as a correlate of clinical changes in an rTMS treatment of depression. Clin. EEG Neurosci. 2008;39(3):131-138. 7. Micoulaud J, Richieri R, Cermolacce M, Loundou A, Lancon C, Vion-Dury J. Parieto-temporal alpha EEG band power at baseline as a predictor of antidepressant treatment response with repetitive transcranial magnetic stimulation: a preliminary study. J Affect Disord. 2012;137(1-3):156-160. 8. Kito S, Hasegawa T, Koga Y. Cerebral blood flow ratio of the dorsolateral prefrontal cortex to the ventromedial prefrontal cortex as a potential predictor of treatment response to transcranial magnetic stimulation in depression. Brain Stimul. 2012;5(4): 547-553. 9. Richieri R, Boyer L, Farisse J, et al. Predictive value of brain perfusion SPECT for rTMS response in pharmacoresistant depression. Eur J Nucl Med Mol Imaging. 2011;38(9):1715-1722. 10. Coutin-Churchman P, Añez Y, Uzcátegui M, et al. Quantitative spectral analysis of EEG in psychiatry revisited: drawing signs out of numbers in a clinical setting. Clin Neurophysiol. 2003;114(12):2294-2306. 11. Khodayari A. Diagnosis of psychiatric disorders using EEG data and employing a statistical decision model. Presented at: International Conference of the IEEE EMBS; Buenos Aires, Argentina; 2010. 12. Yang J, Singh H, Hines EL, et al. Channel selection and classification of electroencephalogram signals: an artificial neural network and genetic algorithm-based approach. Artif Intell Med. 2012;55(2):117-126. 13. Sabeti M, Boostani R, Katebi SD, Price GW. Selection of relevant features for EEG signal classification of schizophrenic patients. Biomed Signal Proc Contr. 2007;2(2):122-134. 14. Hyun TK, Kim BY, Park EH, et al. Computerized recognition of Alzheimer disease-EEG using genetic algorithms and neural network. Future Gen Comput Syst. 2005;21(7):1124-1130. 15. Chum P, Park S-M, Ko K-E, et al. Optimal EEG feature selection by genetic algorithm for classification of imagination of hand movement. Presented at: 38th Annual Conference on IEEE Industrial Electronics Society; Montreal, QC, Canada; 2012. 16. Ocak H. Optimal classification of epileptic seizures in EEG using wavelet analysis and genetic algorithm. Signal Proc. 2008;88(7):1858-1867. 17. Flotzinger D, Pregenzer M, Pfurtscheller G. Feature selection with distinction sensitive learning vector quantisation and genetic algorithms. Presented at: IEEE World Congress on Computational Intelligence; Beijing, China; 1994. 18. Zhiping H, Chen G, Chen C, Xu H, Zhang J. A new EEG feature selection method for self-paced brain-computer interface. Presented at: 10th International Conference on Intelligent Systems Design and Applications; Cairo, Egypt; 2010. 19. Li Y, Fan FY. Classification of schizophrenia and depres sion by EEG with ANNs. Conf Proc IEEE Eng Med Biol Soc. 2005;3:2679-2682. 20. Ito S, Mitsukura Y, Fukumi M, et al. An EEG feature detection system using the neural networks based on genetic algorithms. Presented at: IEEE International Symposium on Computational Intelligence in Robotics and Automation; Kobe, Japan; 2003. 21. Leuchter AF, Cook IA, Lufkin RB, et al. Cordance: a new method for assessment of cerebral perfusion and metabolism using quantitative electroencephalography. Neuroimage. 1994;1(3): 208-219.
Downloaded from eeg.sagepub.com at GEORGIAN COURT UNIV on December 13, 2014
6
Clinical EEG and Neuroscience
22. Leuchter AF, Uijtdehaage SH, Cook IA, O’Hara R, Mandelkern M. Relationship between brain electrical activity and cortical perfusion in normal subjects. Psychiatry Res. 1999;90(2):125-140. 23. Khodayari A, Reilly J, Hasey G, DeBruin H, MacCrimmon D. Using pre-treatment electroencephalography data to predict response to transcranial magnetic stimulation therapy for major depression. Presented at: 33rd Annual International Conference of the IEEE EMBS; Boston, MA; August 2011. 24. Micheli-Tzanakou E. Neural networks in biomedical signal processing. In: Bronzino JD, ed. The biomedical engineering handbook. Boca Raton, FL: CRC Press; 2000:426-430. 25. Fausett L. Fundamentals of neural networks: architectures, algorithms, and applications. Englewood Cliffs, NJ: Prentice Hall; 1994. 26. Haykin S. Neural networks: a comprehensive foundation. New York: Macmillan; 1994. 27. Hong Z. Recognition of epileptic EEG signals based on BP neural networks. Sci Technol Inform. 2009;35:18. 28. Lek S, Guegan J. Artificial neural networks as a tool in ecological modelling: an introduction. Ecol Model. 1999;120(2-3):65-73. 29. Yom-Tov E, Inbar GF. Feature selection for the classification of movements from single movement related potentials. IEEE Trans Neural Syst Rehabil Eng. 2002;10(3):170-177. 30. Blum AL, Langley P. Selection of relevant features and examples in machine learning. Artif Intell. 1997;97(1-2):245-271. 31. Yao X. Evolving artificial neural networks. Proc IEEE. 1999;87(9):694-713. 32. Khodayari-Rostamabad A, Hasey GM, Maccrimmon DJ, Reilly JP, de Bruin H. A pilot study to determine whether machine learn-
ing methodologies using pre-treatment electroencephalography can predict the symptomatic response to clozapine therapy. Clin Neurophysiol. 2010;121(12):1998-2006. 33. Landsness EC, Goldstein MR, Peterson MJ, Tononi G, Benca RM. Antidepressant effects of selective slow wave sleep deprivation in major depression: a high-density EEG investigation. J Psychiatr Res. 2011;45(8):1019-1026. 34. Leuchter A, Cook IA, Marangell LB, et al. Comparative effectiveness of biomarkers and clinical indicators for predicting outcomes of SSRI treatment in major depressive disorder: results of the BRITE-MD study. Psychiatry Res. 2009;169(2):124-131. 35. Cook IA. Midline and right frontal brain function as a physiologic biomarker of remission in major depression. Psychiatry Res. 2009;174(2):152-157. 36. Noda Y, Nakamura M, Saeki T, Inoue M, Iwanari H, Kasai K. Potentiation of quantitative electroencephalograms following prefrontal repetitive transcranial magnetic stimulation in patients with major depression. Neurosci Res. 2013;77(1-2):70-77. 37. Arns M, Cerquera A, Gutiérrez RM, Hasselman F, Freund JA. Non-linear EEG analyses predict non-response to rTMS treatment in major depressive disorder. Clin Neurophysiol. In press. 38. Hosseinifard B, Moradi MH, Rostami R. Classifying depression patients and normal subjects using machine learning techniques. Presented at: 19th Iranian Conference on Electrical Engineering; Tehran, Iran; 2011. 39. Lee TW, Wu YT, Yu YW, Chen MC, Chen TJ. The implication of functional connectivity strength in predicting treatment response of major depressive disorder: a resting EEG study. Psychiat Res. 2011;194(3);372-377.
Downloaded from eeg.sagepub.com at GEORGIAN COURT UNIV on December 13, 2014
