Classification of Targets and Distractors Present in Visual Hemifields Using Time-Frequency Domain EEG Features

Sweeti; Deepak Joshi; B. K. Panigrahi; Sneh Anand; Jayasree Santhosh

doi:10.1155/2018/9213707

Classification of Targets and Distractors Present in Visual Hemifields Using Time-Frequency Domain EEG Features

Sweeti, ;Joshi, Deepak;Panigrahi, B. K.;Anand, Sneh;Santhosh, Jayasree 2018-04-01 00:00:00 Hindawi Journal of Healthcare Engineering Volume 2018, Article ID 9213707, 10 pages https://doi.org/10.1155/2018/9213707 Research Article Classification of Targets and Distractors Present in Visual Hemifields Using Time-Frequency Domain EEG Features 1 1 2 1 3 Sweeti , Deepak Joshi, B. K. Panigrahi , Sneh Anand, and Jayasree Santhosh Centre for Biomedical Engineering, IIT Delhi, New Delhi, India Department of Electrical Engineering, IIT Delhi, New Delhi, India Department of Computer Engineering & Computer Science, Manipal International University, Putra Nilai, Malaysia Correspondence should be addressed to Sweeti; sweeti.bme@gmail.com Received 29 August 2017; Revised 17 January 2018; Accepted 1 February 2018; Published 1 April 2018 Academic Editor: Feng-Huei Lin Copyright © 2018 Sweeti et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. This paper presents a classiﬁcation system to classify the cognitive load corresponding to targets and distractors present in opposite visual hemiﬁelds. The approach includes the study of EEG (electroencephalogram) signal features acquired in a spatial attention task. The process comprises of EEG feature selection based on the feature distribution, followed by the stepwise discriminant analysis- (SDA-) based channel selection. Repeated measure analysis of variance (rANOVA) is applied to test the statistical signiﬁcance of the selected features. Classiﬁers are developed and compared using the selected features to classify the target and distractor present in visual hemiﬁelds. The results provide a maximum classiﬁcation accuracy of 87.2% and 86.1% and an average classiﬁcation accuracy of 76.5 ± 4% and 76.2 ± 5.3% over the thirteen subjects corresponding to the two task conditions. These correlates present a step towards building a feature-based neurofeedback system for visual attention. 1. Introduction [5]. A normal human visual system takes about 150 ms to process the visuals [6]. Primarily two known approaches to study visual attention exist: stimulus-driven (bottom-up) In the past decade, motor brain-computer interface (BCI) has witnessed signiﬁcant progress in the area of rehabilitation. and self-driven (top-down) [5] that follow the ventral and With the promising results in motor BCI, several new direc- dorsal pathways, respectively [7]. The attention capacity of tions in this ﬁeld are emerging. Cognitive BCI (cBCI) is one a subject is limited, and the subject can attend two tasks in of them that has drawn the attention of researchers working parallel using its ability of attention shift. It is supposed to work like a spotlight phenomenon and processes only those in the area of BCI. Examples of cognitive signals used for cBCI are the subject’s motivation [1], attention orientation, objects that come under this spotlight [8]. Characteristics of and mental calculation [2]. The fundamental long-term goal human attention mechanism can be understood from the of cBCI is to develop therapeutic tools for the treatment of various theories given in the past like capacity limit theory, cognitive disorders. attention shift theory, and early and late selection theory Attention is an important entity for cBCI which works [9]. Due to the potential application of selective visual atten- with diﬀerent sensory inputs, like visual, auditory, and tactile. tion in cBCI, it is apparent to explore further the various Cognitive attention oﬀers a person the ability to select the other aspects of attention. object of interest by ignoring distractors [3]. The information In the current scenario, researchers are using advanced gained with this selection goes to the higher processing modalities like fMRI, PET, and fNIRS [5, 10] in cognitive mechanism that works like a metacognitive system and neuroscience research. These techniques are noninvasive results in perception, recognition, and memory formation and oﬀer a good spatial resolution. However, these tech- [4]. The attention paid by a subject to a visual object present niques suﬀer lower temporal resolution and higher cost. at a speciﬁc spatial location is called visual spatial attention The present study uses the EEG data to study the selective 2 Journal of Healthcare Engineering visual attention due to its good temporal resolution and low 2. Materials and Methods cost. The aim of this paper is to develop a classiﬁer that can 2.1. Task Description. This study uses the publically available classify the target and distractor present in opposite visual EEG dataset acquired by Jeanne Townsend in the laboratory hemiﬁelds using EEG features. These EEG features vary with of Eric Courchesne at UCSD. Figure 1(a) gives an illustration the task conditions and cognitive load provided. Delta of one sequence of the task trial presented to the subjects. The subband provides the best separation between low and high task presented ﬁve spatial locations represented by square cognitive loads [11]. Previous works suggest that theta band boxes placed from left to right and located 0.8 cm above the oscillations in the hippocampocortical feedback loop reﬂect central midline of the computer screen along with a center the encoding of new information. Upper alpha oscillations ﬁxation cross present on the computer screen. During each in the thalmo-cortical feedback loop reﬂect the search and 76-second task block, subjects focused covertly at the retrieval process in long-term memory [12]. Alpha power is attended stimulus location which is presented with a diﬀer- higher for target intake during the attention task while beta ently shaded box. The attended location is shown in band power increases preceding the correct response and Figure 1(a) by a dashed square box for the illustration pur- does not change in case of the erroneous response [13, 14]. pose. In each task block, 100 disk stimuli appeared randomly Gamma band power increases in the contralateral hemiﬁeld at diﬀerent spatial locations bounded in the square box. when the subject attends to a stimulus [15, 16]. Temporal Stimuli appeared for a duration of 117 ms at a diﬀerent evaluation of alpha and beta during spatial visual task spatial location with an interstimulus duration of 225– indicates a decrease in alpha band power in a time window 1000 ms [23]. Subjects responded via a thumb switch, as of 375–500 ms and an increase in beta band power of 500– soon as stimuli appeared at the attended location. The next 875 ms [17]. Also, the modulation of EEG rhythms exists dur- sequence begins after this. Data was recorded for such ing preparatory attention interval. Preparatory attention is thirty task blocks. This paper analyzes the activity with associated with the alpha power decrease in the left and right respect to the two conditions where the subject attends to temporal and occipital areas and the beta power decrease in the target present in the left and right visual hemiﬁelds the bilateral occipital, left frontal, and middle frontal occipital while ignoring the distractor present in right and left visual [16, 18]. Some other EEG features used for the cognitive load hemiﬁelds, respectively. measurement include log variance, Hjorth parameters [19], spectral entropy, spectral edge frequency, intensity weighted mean frequency, and intensity weighted bandwidth [11]. 2.2. EEG Recording and Signal Processing. This study uses the Despite previous signiﬁcant research in cognition-related publically available EEG dataset acquired by Jeanne Town- studies, it is of utmost importance to extrapolate EEG-based send in the laboratory of Eric Courchesne at UCSD. EEG research in attentional orientation for classiﬁcation purposes data collected from 13 (two female, eleven male; ages 22–40 in neurofeedback applications. In most of the target versus years) healthy right-handed subjects performing a visual nontarget classiﬁcation problems, the task protocols involve spatial attention task was recorded from 29 scalp electrode the targets and distractors appearing at the central visual ﬁeld locations using an EEG standard electrode cap (Electro-Cap or center of the screen. Such tasks imitate a very diﬀerent sce- International Inc.) and two EOG (electrooculogram) elec- nario than the real-life situations where target and distractors trodes with a sampling frequency of 512 Hz. Figure 1(b) may appear at diﬀerent spatial locations or visual ﬁelds. The shows the location of EEG scalp electrodes placed for record- present work involves the classiﬁcation of the target and dis- ing data. The data were collected with reference to the right tractor present in diﬀerent visual hemiﬁelds while the subject mastoid electrode position within the analog passband of attends the objects using peripheral visual attention. This is 0.01–50 Hz. Additional digital ﬁltering was done using more like a real-life situation, for example, looking at the 4th-order Butterworth band-pass ﬁlter in the range of road while attending to or ignoring the objects appearing 0.1–45 Hz, and the signal was average rereferenced. The on both sides of the road. This paper explores the two task Automatic Artifact Removal (AAR) v1.3 toolbox based on conditions: the ﬁrst task condition compares the activity the blind source separation principle was used to remove while the subject attends to the target object present in the the eye blinks and muscle artifacts [24]. EEG analysis left visual hemiﬁeld, ignoring the distractor present in the involved an epoch length of 800 ms, starting 200 ms before right visual hemiﬁeld. The second task condition compares the stimulus onset to 600 ms after the stimulus oﬀset, over the activity while the subject attends to the target object 90 trials from each subject. Figure 1(b) shows the scalp present in the right visual hemiﬁeld, ignoring the distractor electrode locations. present in the left visual hemiﬁeld. The selected data is divided into four datasets, namely A, In this work, we will explore diﬀerent EEG features and B, C, and D; the summary is given in Table 1. Dataset A cor- classiﬁers are developed further using most relevant features. responds to the activity while the subject attends to the left The results would facilitate the development of a neurofeed- visual hemiﬁeld and the target appears at the same location back system for selective visual attention [20–22]. Section 2 in the left visual hemiﬁeld. Dataset B corresponds to activity describes the materials and methods used in the study: task while the subject attends to the left visual ﬁeld and the description, data preprocessing method, and methodology distractor appears in the right visual hemiﬁeld. Dataset C followed. Section 3 gives the details of channel selection, sta- represents the activity while the subject attends to the right tistical analysis, and classiﬁcation performed. Section 4 dis- visual hemiﬁeld and the distractor appears in the left visual cusses the results. ﬁeld. Dataset D represents the activity while the subject Journal of Healthcare Engineering 3 Channel locations F3 F4 FZ 117 ms FC5 FC6 FC1 FC2 Attended location in T7 C4 C3 CZ T8 shaded square CP1 CP2 CP5 CP6 Target at attended PZ location P3 P4 P7 P8 PO3 POZ PO4 117 ms PO7 PO8 O1 OZ O2 Attended location in shaded square (a) (b) Figure 1: (a) A sequence of the task trial. (b) EEG scalp electrode placement locations. Table 1: Summary of four datasets selected for two task conditions. Task condition 1 Task condition 2 Task condition Dataset A Dataset B Dataset C Dataset D Attending target object Ignoring distractor objects Ignoring distractor objects Attending target object Patient state in left hemiﬁeld in right hemiﬁeld in left hemiﬁeld in right hemiﬁeld Eye ﬁxation Center Center Center Center Electrode type Surface Surface Surface Surface International 10-20 International 10-20 International 10-20 International 10-20 Electrode placement placement system placement system placement system placement system Number of subjects 13 13 13 13 Number of electrodes 31 31 31 31 Number of trials from 90 90 90 90 each subject Epoch duration 800 ms 800 ms 800 ms 800 ms attends to the right visual ﬁeld and the target appears at the (1) Hjorth complexity (H ): it is a measure of the complexity same location in the right visual hemiﬁeld. Task condition spread of the spectrum and represents the change in 1 compares the dataset A and dataset B. The task condition frequency [25]. 2 compares the dataset C and dataset D. In other words, 2 2 the ﬁrst task condition compares the activity while the std d X/dt std X subject attends to the target object present in the left visual H = , 1 complexity std dX/dt hemiﬁeld ignoring the distractor present in the right visual hemiﬁeld. The second task condition compares the activity while the subject attends to the target object present in the where std is standard deviation function. right visual hemiﬁeld ignoring the distractor present in the left visual hemiﬁeld. Figure 2 demonstrates the ﬂow chart (2) Hjorth mobility (H ): it is a measure of mean mobility of the complete signal analysis process. frequency. 2.3. Feature Extraction. EEG features studied are discussed std dX/dt brieﬂy in this section. These features were selected based on H = , 2 mobility std Xt their applications in biomedical signal processing as dis- cussed in Introduction. In this work, we are trying to explore the utility of these features in attentional studies: where std is standard deviation function. Button Next ISI First sequence sequence press 225−1000 ms 4 Journal of Healthcare Engineering (6) Median power frequency (mpf): it is the frequency Multichannel EEG data below which 50% of the total power of the signal is present, calculated over the half of the total area of the power spectrum. Preprocessing (filtering and artifact removal) (7) Spectral edge frequency (sef): it is the frequency below which 95% of the total power of the signal is Single trial analysis present, calculated over 95% of the total area of the power spectrum. Feature calculation 2.4. Channel Selection. Channel selection is an important step as it helps in selecting the channels that can distinguish two or more datasets. This procedure can help in reducing the computational burden by minimizing the number of Channel selection channels. A number of channel selection methods including ﬁlters, wrappers, and embedded methods exist. Among these methods, wrapper methods are good at providing a Statistical analysis reliable set of features. The stepwise discriminant analysis (SDA) used in this study is a wrapper method that gener- ates a reliable set of features with multivariate analysis of variance (MANOVA). Classification and conclusion A data matrix (C ) is created for each subject corre- sponding to each EEG feature. The matrix C is composed Figure 2: Flowchart for feature-based EEG data analysis. of 29 channels and a total of 180 observations with respect to targets and nontargets for both the task conditions. Rows of the matrix represent the observations, and columns repre- (3) Average frequency: it deﬁnes the number of times the sent the channels as shown: signal crosses the zero value. a … a i,j i,j+n Total zero crossing points Average frequency = C = ⋮⋮ ⋮ , 6 Epoch duration a … a 3 i+m,j i+m,j+n where i and j represent the rows and column, respectively, m (4) Lempel-Ziv complexity (lz-complexity): it is a represents the number of observations, and n represents the harmonic variability metric that shows the distinct number of channels. pattern contained in the sequence as an algorithm The matrix C for each EEG feature, for a single subject, scans the data sequence from left to right [25]. To is pooled to perform SDA to select the three best performing compute the lz-complexity, the signal needs to be channels. MANOVA compares the sample means based decoded ﬁrst with respect to some threshold value on the variance-covariance between variables to test the which could be mean or median of the signal. In this signiﬁcance diﬀerence. MANOVA gives the signiﬁcant dif- way, the signal greater than the threshold which ference value represented by lambda (λ). SDA present the maps to 1 else to 0 to obtain the symbolic sequence channels in descending order of discriminating power λ, is further parsed to obtain the encoded sequence and the number of channels can further be selected as [26]. For an encoded sequence, s(n) of length n, the desired based on the lambda power. In this work, we are lz-complexity can be obtained as below: selecting three channels. 2.5. Statistical Analysis. Repeated measure analysis of sn Lempel − Ziv complexity = 4 variance (rANOVA) was performed to analyze the statistical signiﬁcance of selected EEG features diﬀerentiating the two datasets within the two task conditions. rANOVA investi- (5) Band power: band power can be calculated from the gates the EEG features and task condition interaction over power spectrum of the signal. Power spectrum repeated measurements for the two task conditions. It is used represents the energy of a signal over the frequency to test null hypotheses about the mean. If the mean of the two components that it possesses. For a signal x, power classes is diﬀerent, then the null hypothesis rejects. The spectrum can be calculated using results of rANOVA are presented in the following form: F(df ,df )= F value, p = p value, fc error Power spectrum = X ∗ conjugate X , 5 where df = degree of freedom of feature and task fft fft fc condition interaction, where X represents the Fourier transform of the signal x. df degree of freedom of error, fft error = Journal of Healthcare Engineering 5 F = critical value, Table 2: Selected channels corresponding to the eight EEG features. p = signiﬁcance value. Task condition 1 Task condition 2 This test oﬀers Greenhouse–Geisser (pGG), Huynh– Feature Selected channels Selected channels Feldt (pHF), and lower bound (pLB) corrections for multiple (using SDA method) (using SDA method) comparisons to avoid false rejection of the null hypothesis. Average frequency C3, CP2, PZ C3, CP2, P3 lz-complexity CP1, PZ, P4 CP2, P3, P4 2.6. Classiﬁcation and Cross-Validation. In the present work, we need to classify the two classes of the attentional load Complexity CP1, PZ, P4 CP5, CP1, PZ from targets and distractors. To solve this two-class classi- Mobility CP1, PZ, P4 CP1, P3, PZ ﬁcation problem, this study compares the three classiﬁca- Median power CP2, PZ, P4 CP1, P3, PZ tion approaches, namely, artiﬁcial neural network (ANN), frequency K-nearest neighbor (KNN), and support vector machine Spectral edge CP1, PZ, P4 CP1, PZ, P4 (SVM). The input and target data matrices were constructed frequency for the subject by taking 90 observations from each channel Delta power CP2, PZ, P4 CP1, P3, PZ with respect to each dataset. Features were normalized to Beta power CP1, CP2, PZ CP1, CP2, P3 zero mean and unit variance before classiﬁcation. Artiﬁcial neural network (ANN): a pattern recognition network architecture with three network layers; namely, the perform input layer, the hidden layer, and the output layer were used. ed to avoid the eﬀect of any time-related change There were 29 nodes in the input layer, ten nodes in the in data that may occur during EEG data recording in a hidden layer, and one node in the output layer. The tangent block design. sigmoid activation function was used. The network was Receiver operating characteristic (ROC) curves were used trained with the Levenberg–Marquardt back propagation to compare the performance of the diﬀerent classiﬁers. ROC method of training using a preset amount of data for the is a plot of the two operating characteristics known as a true training, testing, and validation. The model was validated positive rate (TPR) and the false positive rate (FPR), where using cross-validation methods. TPR is the probability of detection while FPR gives the K-Nearest neighbor (KNN): the developed KNN model probability of false alarm. was ﬁtted by means of the Euclidean distance metric between the two nearest neighbors. This model selects the neighbors 3. Results with a known class from the training dataset and assigns weights to it according to the distance to the space variable The preprocessed datasets corresponding to the two task using the exhaustive searcher. It made the decision with the conditions were analyzed with the suggested methodology. majority vote given by selected nearest neighbors. For a space First, to reduce the system complexity, the channel selection variable X, the classiﬁer looks for the nearest neighbor among is performed. Feature selection results are given in Table 2. diﬀerent classes, say A and B, and assigns the class label Selected channels are presented in the descending order having the smallest distance to the space variable X. of their discriminating powers. These results suggest that Support vector machine (SVM): it is a supervised the parietal and central parietal region electrodes are classiﬁcation model that classiﬁes the data by ﬁnding the among the best performing channels to distinguish the targets best hyperplane oﬀering the largest margin between two and distractors. classes. These hyperplanes are the decision planes that Further, the distribution of the diﬀerent EEG features can separate the objects having diﬀerent class member- corresponding to the two task conditions with selected chan- ships. The developed SVM model mapped the predictor nels is studied. This distribution is shown in Figure 3. data using radial basis kernel function. Sequential minimal Figures 3(a) and 3(i) show that the average frequency value optimization (SMO) approach was used to solve this is lower for the targets in both task conditions. lz-complexity, binary classiﬁcation problem. which gives the harmonic variability metric, oﬀers greater value for nontargets, suggesting distinct patterns in the signal 2.7. Cross-Validation. The developed classiﬁer models were while ignoring the distractors as shown in Figures 3(b) and cross-validated using the k-fold cross-validation approach. 3(j). Hjorth complexity feature’s value, which represents the In this approach, the data is divided into k subsamples change in frequency, is higher for targets as given in randomly. At each fold, (k− 1) subsamples are used for the Figures 3(c) and 3(k). It shows that frequency spread is more training, remaining one for the testing. The process is while attending to targets than nontargets. Figures 3(d) and repeated for each fold, and mean accuracy is calculated from 3(l) show that Hjorth mobility, which represents the mean the average of the results of diﬀerent folds. Also, another frequency, is higher for nontargets. Median power frequency cross-validation approach is suggested. This approach works embodying 50% of the total signal power is higher for non- similar to k-fold cross-validation except that the criteria for targets as shown in Figures 3(e) and 3(m). Figures 3(f) and the selection of training and testing datasets is diﬀerent. 3(n) show that the spectral edge frequency that represents Unlike the k-fold approach, the datasets are not selected ran- 95% of the total power is also high while ignoring the nontar- domly but are selected in a way that there is a maximum time gets. Delta power values are higher for targets for both task separation in the data points. This cross-validation was conditions as given in Figures 3(g) and 3(o). Figures 3(h) 6 Journal of Healthcare Engineering Average frequency Lempel-Ziv complexity Average frequency Lempel-Ziv complexity (datasets A versus B) (datasets A versus B) (datasets C versus D) (datasets C versus D) 12 20 12 16 10 10 15 12 8 8 6 10 6 8 4 4 5 4 2 2 0 0 0 0 C3 CP2 PZ CP1 P4 P4 C3 CP2 P3 CP2 P3 P4 Channel number Channel number Channel number Channel number (a) (b) (i) (j) Hjorth complexity Hjorth mobility Hjorth complexity Hjorth mobility (datasets A versus B) (datasets A versus B) (datasets C versus D) (datasets C versus D) 3.5 0.2 3.5 0.2 0.15 0.15 2.5 2.5 2 0.1 0.1 1.5 1.5 0.05 0.05 0.5 0.5 0 0 0 0 CP1 PZ P4 CP1 PZ P4 CP5 CP1 PZ CP1 P3 PZ Channel number Channel number Channel number Channel number (c) (d) (k) (l) Median power frequency Spectral edge frequency Median power frequency Spectral edge frequency (datasets A versus B) (datasets A versus B) (datasets C versus D) (datasets C versus D) 10 35 10 35 30 30 8 8 25 25 6 6 20 20 15 15 4 4 10 10 2 2 5 5 0 0 0 0 CP2 PZ P4 CP1 PZ P4 CP1 P3 PZ CP1 PZ P4 Channel number Channel number Channel number Channel number (e) (f ) (m) (n) Delta power Beta power Delta power Beta power (datasets A versus B) (datasets A versus B) (datasets C versus D) (datasets C versus D) 80 30 80 35 60 60 40 15 40 20 20 0 0 0 0 CP2 PZ P4 CP1 CP2 PZ CP1 CP3 PZ CP1 CP2 P3 Channel number Channel number Channel number Channel number (g) (h) (o) (p) Target Distractor Distractor Target (I) (II) Figure 3: Feature distribution over selected channels for (I) task condition 1 and (II) task condition 2: (a), (i) average frequency; (b), (j) lz- complexity; (c), (k) Hjorth complexity; (d), (l) Hjorth mobility; (e), (m) median power frequency; (f), (n) spectral edge frequency; (g), (o) delta power; (h), (p) beta power. and 3(p) give higher beta power for distractors that advise Next, the repeated measure analysis of variance more activation of the attentional network in inhibiting the (rANOVA) was performed for all the EEG features extracted nontargets. It can be concluded from Figure 3 that Hjorth from the two task conditions. It was performed using complexity and delta power values are higher for targets cor- three selected channel features, two datasets, and 90 obser- responding to the task conditions 1 and 2 while other features vations from each dataset. The degree of freedom (DF) show lower values. was (3− 1) = 2 for channel features, (3− 1) ∗ (2− 1) = 2 Complexity Delta power Median power frequency Average frequency Mobility Spectral edge frequency Iz-complexity Beta power Median power frequency Complexity Delta power Average frequency Mobility Iz-complexity Beta power Spectral edge frequency Journal of Healthcare Engineering 7 using Hjorth complexity, Hjorth mobility, delta power, for channel feature-condition interaction, and (180 – 2) ∗ (3− 1) = 356 for error. The rANOVA results show that and beta power, respectively. four among the extracted EEG features exhibit a signiﬁ- cant diﬀerence for channel feature and condition interac- tion with a signiﬁcance level of p <0 1. The result 4. Discussion indicates a signiﬁcant channel feature and task condition interaction with lower bound correction for Hjorth com- Selective visual attention is the ability to select the visual plexity [F(2,356) = 6.0145, p =0 0151], Hjorth mobility information of interest present in the visual ﬁeld. It is a key [F(2,356) = 5.62, p =0 0188], delta power [F(2,356) = to various other skills like perception and recognition and 3.8754, p =0 0505], and beta power [F(2,356) = 3.7965, memory as well; it can also aﬀect these skills if there is a prob- p =0 0529] over the task condition 1. On the other hand, other lem with it. This paper studies the EEG correlates of visual features including average frequency [F(2,356) = 0.42023, attention in a spatial attention task. It is important to study p =0 51766], lz-complexity [F(2,356) = 0.68079, p =0 41042], features of the EEG as these are very crucial and provide mpf [F(2,356) = 1.1707, p =0 28073], and sef [F(2,356) = more information than the raw data. These features are the 2.3163, p =0 1298] show no statistically signiﬁcant diﬀer- potential candidates that can be used in neurofeedback ence in the target and distractor. For task condition 2, systems to give feedback about their performance to the sub- the statistical results display a signiﬁcant channel feature jects. The present study attempts to ﬁnd the EEG correlates and task condition interaction with lower bound correc- of attention for the task when the subject attends to target tion for Hjorth complexity [F(2,356) = 3.2531, p =0 0729], objects present in one visual hemiﬁeld while ignoring distrac- Hjorth mobility [F(2,356) = 5.0276, p =0 0261], delta power tor objects present in another visual hemiﬁeld. To reduce the [F(2,356) = 3.5058, p =0 0627], and beta power [F(2,356) = system complexity and increase classiﬁcation accuracy, chan- 2.7321, p =0 0900]. There is no statistically signiﬁcant channel nel selection is performed [27]. Channel selection performed feature and task condition interaction over the average over EEG features suggest that the channels with the most frequency [F(2,356) = 1.012, p =0 31579], lz-complexity discriminating power lie in the central-parietal and parietal [F(2,356) = 0.83486, p =0 36211], mpf [F(2,356) = 1.937, regions, which are involved in the visual-spatial processing [28]. The rANOVA-based statistical analysis found that p =0 16573], and sef [F(2,356) = 2.4043, p =0 12278]. The statistical analysis concludes that the Hjorth amongst the features studied, Hjorth complexity, Hjorth complexity, Hjorth mobility, delta power, and beta power mobility, delta power, and beta power can signiﬁcantly diﬀer- can signiﬁcantly diﬀerentiate the activity while the subject entiate the datasets corresponding to the two task conditions. attends to stationary targets and inhibits the distractors This selection suggests higher beta and lower delta for non- present in another visual hemiﬁeld. These selected EEG fea- targets, representing higher cognitive demand or working tures are further explored to develop a classiﬁer to classify memory load for inhibition which agrees with earlier studies the target and distractor present in diﬀerent visual hemiﬁelds. involving targets and nontargets [29]. Hjorth features have Feature matrices with three selected channel features and been used earlier for cognitive load measurement [19]. The 180 observations were prepared for each subject correspond- present study explores these features and shows higher ing to all the EEG features for the target and distractor clas- Hjorth mobility and lower Hjorth complexity for nontargets siﬁcations. A comparison of the three diﬀerent classiﬁers which correspond to the mean frequency and change in using the four EEG features for a subject is presented in this frequency, respectively. section. The ROC curves of Figure 4 illustrate that the artiﬁ- The classiﬁcation system is further developed using cial neural network (ANN) classiﬁes the target and distractor selected features to distinguish the activities corresponding better than the other two methods. So the details of the to targets and nontargets. The importance of such classiﬁ- classiﬁcation performance parameters, namely, sensitivity, cation lies in applications like cognitive brain-computer speciﬁcity, and accuracy corresponding to the ANN classiﬁer interface or neurofeedback system for training where the only, are given in Tables 3 and 4. cognitive control measures are used to control the BCI Tables 3 and 4 show the mean and maximum values of and train the subjects. Classiﬁcation accuracy in cBCI is the classiﬁcation results obtained from thirteen subjects for limited by various internal and external factors, like sen- the two task conditions using four diﬀerent EEG features. sory and cognitive, comparative to reasonable accuracy in The classiﬁcation was performed for each and every subject, motor BCI [30, 31]. Due to this limitation, accuracy and then the mean was taken over the classiﬁcation results reported in previous research was restricted to only 75% obtained from a population of 13 subjects. These mean values and 79% in spatial attention tasks [32, 33] using noninva- and standard deviation for sensitivity, speciﬁcity, and sive techniques. We could reach a maximum accuracy of accuracy, corresponding to diﬀerent features, are given in 87.2% and 86.1% and a mean accuracy of 76.5% and Tables 3 and 4. These tables also give the maximum value 76.2% over thirteen subjects for the two task conditions, respectively, by using the EEG-based noninvasive method. of the classiﬁcation results obtained. Table 3 spectates a maximum classiﬁcation accuracy of 80.6%, 87.2%, 82.2%, In this way, a classiﬁer is developed that can classify the and 80% for the task condition 1 using Hjorth complexity, peripheral attention paid to targets and distractors present Hjorth mobility, delta power, and beta power, respectively. in diﬀerent visual hemiﬁelds. Such a classiﬁer can facilitate Table 4 shows a maximum classiﬁcation accuracy of the development of an EEG feature-based neurofeedback system for attention [34]. 84.4%, 86.1%, 83.3%, and 86.1% for the task condition 2 8 Journal of Healthcare Engineering ROC curve for task condition 1 ROC curve for task condition 2 ROC curve for task condition 1 (Hjorth complexity) (Hjorth complexity) (Hjorth mobility) 1 1 1 0.9 0.9 0.9 0.8 0.8 0.8 0.7 0.7 0.7 0.6 0.6 0.6 0.5 0.5 0.5 0.4 0.4 0.4 0.3 0.3 0.3 0.2 0.2 0.2 0.1 0.1 0.1 0 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 False positive rate False positive rate False positive rate ANN ANN ANN KNN KNN KNN SVM SVM SVM (a) (b) (c) ROC curve for task condition 2 ROC curve for task condition 1 ROC curve for task condition 2 (Hjorth mobility) (Delta power) (Delta power) 1 1 1 0.9 0.9 0.9 0.8 0.8 0.8 0.7 0.7 0.7 0.6 0.6 0.6 0.5 0.5 0.5 0.4 0.4 0.4 0.3 0.3 0.3 0.2 0.2 0.2 0.1 0.1 0.1 0 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 False positive rate False positive rate False positive rate ANN ANN ANN KNN KNN KNN SVM SVM SVM (d) (e) (f) ROC curve for task condition 1 ROC curve for task condition 2 (Beta power) (Beta power) 1 1 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 False positive rate False positive rate ANN ANN KNN KNN SVM SVM (g) (h) Figure 4: ROC curve for (a) and (b) Hjorth complexity, (c) and (d) Hjorth mobility, (e) and (f) delta power, and (g) and (h) beta power, using artiﬁcial neural network (ANN), K-nearest neighbor (KNN), and support vector machine (SVM) classiﬁers during the testing corresponding to task conditions 1 and 2, respectively. True positive rate True positive rate True positive rate True positive rate True positive rate True positive rate True positive rate True positive rate Journal of Healthcare Engineering 9 Table 3: Classiﬁcation results displaying % sensitivity (SN), % speciﬁcity (SP), and % accuracy (AC) obtained using ANN for task condition 1 during testing. Task condition 1 Features Mean Maximum Sensitivity Speciﬁcity Accuracy Sensitivity Speciﬁcity Accuracy Hjorth complexity 77.1 ± 7.37 73.6 ± 3.55 75.3 ± 3.72 87.8 78.9 80.6 Hjorth mobility 76.4 ± 5.13 76.6 ± 4.80 76.5 ± 4 88.9 85.6 87.2 Delta power 74.1 ± 5.67 73.0 ± 4.97 73.6 ± 3.93 86.7 80 82.2 Beta power 73.4 ± 5.24 75.8 ± 5.40 74.9 ± 2.85 84.4 83.3 80 Table 4: Classiﬁcation results displaying % sensitivity (SN), % speciﬁcity (SP), and % accuracy (AC) obtained using ANN for task condition 2 during testing. Task condition 2 Features Mean Maximum Sensitivity Speciﬁcity Accuracy Sensitivity Speciﬁcity Accuracy Hjorth complexity 73.5 ± 5.12 77.5 ± 6.65 75.5 ± 4.24 81.1 87.8 84.4 Hjorth mobility 77.5 ± 5.36 74.9 ± 6.41 76.2 ± 5.30 86.7 87.8 86.1 Delta power 72.7 ± 5.44 76.0 ± 5.55 74.3 ± 4.76 82.2 85.6 83.3 Beta power 78.6 ± 7.65 72.4 ± 6.38 75.5 ± 5.22 88.9 83.3 86.1 5. Conclusion Supplementary Materials The present study explores the EEG features that can distin- This manuscript includes a supplementary result ﬁle, which guish the targets and nontargets present in the diﬀerent contains the classiﬁcation results obtained. (Supplementary visual hemiﬁelds. A classiﬁcation system to classify the Materials) targets and distractors present in opposite visual hemiﬁelds is proposed in this paper. The analysis is done to optimize References the performance of the system. Results provide EEG correlates of selective visual attention that can classify the [1] R. A. A. S. Musallam, B. D. Corneil, B. Greger, and activities while attending targets and ignoring distractors. H. Scherberger, “Cognitive control signals for neural prosthet- Other EEG features can be explored to further increase the ics,” Science, vol. 305, no. 5681, pp. 258–262, 2004. classiﬁcation accuracy and make the possibility of feature- [2] K. Jerbi, S. Freyermuth, L. Minotti, P. Kahane, A. Berthoz, and based neurofeedback system feasible. J. P. Lachaux, “Chapter 12 watching brain TV and playing brain ball: exploring novel BCI strategies using real‐time analysis of human intracranial data,” International Review of Data Availability Neurobiology, vol. 86, pp. 159–168, 2009. [3] M. I. Posner and S. E. Petersen, “The attention system of the Data supporting this research article are available from the human brain,” Annual Review Neuroscience, vol. 13, no. 1, online Archive of Swartz Center, University of California in pp. 25–42, 1990. San Diego (UCSD), available from http://headit.ucsd.edu/ [4] P. Cavanagh, “Attention routines and the architecture of studies/2747adbc-385a-11e3-b29d-0050563f2612. selection,” in Cognitive Neuroscience of Attention,M. Posner, Ed., pp. 13–28, Guilford Press, New York, NY, Conflicts of Interest USA, 2004. [5] G. R. Mangun, “Neural mechanism of attention,” in The The authors have no conﬂict of interest to declare. Cognitive Electrophysiology of Mind and Brain, pp. 247–259, Elsevier Science (USA), 2002, http://citeseerx.ist.psu.edu/ viewdoc/download?doi=10.1.1.470.5863&rep=rep1&type=pdf. Acknowledgments [6] S. Thorpe, D. Fize, and C. Marlot, “Speed of processing in the This work was supported by The Ministry of Human human visual system,” Nature, vol. 381, no. 6582, pp. 520–522, Resource Development, Government of India. The authors would like to thank the Indian Institute of Technology [7] S. Kastner and L. G. Ungerleider, “Mechanisms of visual atten- (IIT), Delhi, for providing the facilities to carry out the tion in the human cortex,” Annual Review of Neuroscience, vol. 23, pp. 315–341, 2000. research. The authors also thank the Archive of Swartz Center, University of California in San Diego (UCSD), for [8] P. Stanford, “Chapter three: attention,” 2007, http://www- sharing their data online. psych.stanford.edu/~ashas/Cognition Textbook/chapter3.pdf. 10 Journal of Healthcare Engineering [25] A. Kumar and S. Anand, “EEG signal processing for monitor- [9] R. Desimone and J. Duncan, “Neural mechanisms of selective visual attention,” Annual Review of Neuroscience, vol. 18, no. 1, ing depth of anesthesia,” IETE Technical Review, vol. 23, no. 3, pp. 193–222, 1995. pp. 179–186, 2006. [10] S. Yantis, “The neural basis of selective attention cortical [26] J. Hu, J. Gao, and J. C. Principe, “Analysis of biomedical signals sources and targets of attentional modulation,” Current Direc- by the Lempel-Ziv complexity: the eﬀect of ﬁnite data size,” IEEE Transactions on Biomedical Engineering, vol. 53, no. 12, tions in Psychological Science, vol. 17, no. 2, pp. 86–90, 2008. pp. 2606–2609, 2006. [11] P. Zarjam, J. Epps, and F. Chen, “Spectral EEG features for evaluating cognitive load,” Annual International Conference [27] Y. Li and Y. Koike, “A real-time BCI with a small number of channels based on CSP,” Neural Computing and Applications, of the IEEE Engineering in Medicine and Biology Society, vol. 2011, pp. 3841–3844, 2011. vol. 20, no. 8, pp. 1187–1192, 2011. [12] W. Klimesch, “EEG alpha and theta oscillations reﬂect cogni- [28] L. Shen, X. Hu, E. Yacoub, and K. Ugurbil, “Neural correlates tive and memory performance: a review and analysis,” Brain of visual form and visual spatial processing,” Human Brain Research Reviews, vol. 29, no. 2-3, pp. 169–195, 1999. Mapping, vol. 8, no. 1, pp. 60–71, 1999. [13] D. A. Valentino and R. L. Dufresne, “Attention tasks and EEG [29] M. C. Hout and S. D. Goldinger, “Learning in repeated visual power spectra,” International Journal of Psychophysiology, search,” Attention, Perception & Psychophysics, vol. 72, no. 5, vol. 11, no. 3, pp. 299–301, 1991. pp. 1267–1282, 2010. [14] M. Gola, M. Magnuski, I. Szumska, and A. Wróbel, “EEG beta [30] T. Nguyen, A. Khosravi, D. Creighton, and S. Nahavandi, band activity is related to attention and attentional deﬁcits in “EEG data classiﬁcation using wavelet features selected by Wil- the visual performance of elderly subjects,” International coxon statistics,” Neural Computing and Applications, vol. 26, Journal of Psychophysiology, vol. 89, no. 3, pp. 334–341, 2013. no. 5, pp. 1193–1202, 2015. [15] M. M. Muller, T. Gruber, and A. Keil, “Modulation of induced [31] S. Chen, Z. Luo, and H. Gan, “An entropy fusion method for gamma band activity in the human EEG by attention and feature extraction of EEG,” Neural Computing and Applica- visual information processing,” International Journal of tions, 2016. Psychophysiology, vol. 38, no. 3, pp. 283–299, 2000. [32] M. S. Treder, A. Bahramisharif, N. M. Schmidt, M. A. J. van [16] S. Meyberg, M. Werkle-bergner, W. Sommer, and O. Dimigen, Gerven, and B. Blankertz, “Brain-computer interfacing using “Microsaccade-related brain potentials signal the focus of modulations of alpha activity induced by covert shifts of atten- visuospatial attention,” NeuroImage, vol. 104, pp. 79–88, 2015. tion,” Journal of Neuroengineering and Rehabilitation, vol. 8, no. 1, pp. 24–29, 2011. [17] M. V. Marrufo, E. Vaquero, M. J. Cardoso, and C. M. Gómez, “Temporal evolution of α and β bands during visual spatial [33] H. Morioka, A. Kanemura, S. Morimoto et al., “Decoding attention,” Cognitive Brain Research, vol. 12, no. 2, pp. 315– spatial attention by using cortical currents estimated from 320, 2001. electroencephalography with near-infrared spectroscopy prior information,” NeuroImage, vol. 90, pp. 128–139, 2014. [18] C. M. Gómez, J. Marco-Pallarés, and C. Grau, “Location of brain rhythms and their modulation by preparatory attention [34] G. A. Alvarez and P. Cavanagh, “Independent resources for estimated attentional tracking in the left and right visual hemiﬁelds,” by current density,” Brain Research, vol. 1107, no. 1, pp. 151–160, 2006. Psychological Science, vol. 16, no. 8, pp. 637–643, 2005. [19] D. Das, D. Chatterjee, and A. Sinha, “Unsupervised approach for measurement of cognitive load using EEG signals,” in 2013 IEEE 13th International Conference on Bioinformatics and Bioengineering (BIBE), pp. 1–6, Chania, Greece, 2013. [20] N. F. Ramsey, M. P. V. D. Heuvel, K. H. Kho, and F. S. S. Leijten, “Towards human BCI applications based on cogni- tive brain systems: an investigation of neural signals recorded from the dorsolateral prefrontal cortex,” IEEE Transactions on Neural Systems and Rehabilitation Engi- neering, vol. 14, no. 2, pp. 214–217, 2006. [21] M. J. Vansteensel, D. Hermes, E. J. Aarnoutse et al., “Brain- computer interfacing based on cognitive control,” Annals of Neurology, vol. 67, no. 6, pp. 809–816, 2010. [22] E. Astrand, C. Wardak, and S. Ben Hamed, “Selective visual attention to drive cognitive brain-machine interfaces: from concepts to neurofeedback and rehabilitation applications,” Frontiers in Systems Neuroscience, vol. 8, pp. 1–16, 2014. [23] S. Makeig, M. Westerﬁeld, J. Townsend, T. P. Jung, E. Courchesne, and T. J. Sejnowski, “Functionally independent components of early event-related potentials in a visual spatial attention task,” Philosophical Transactions of the Royal Society B: Biological Sciences, vol. 354, no. 1387, pp. 1135–1144, 1999. [24] G. Gómez-Herrero, “Automatic Artifact Removal (AAR) Toolbox v1.3 (Release 09.12.2007) for MATLAB,” 2007, http://germangh.github.io/pubs/aardoc07.pdf. International Journal of Advances in Rotating Machinery Multimedia Journal of The Scientific Journal of Engineering World Journal Sensors Hindawi Hindawi Publishing Corporation Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 http://www www.hindawi.com .hindawi.com V Volume 2018 olume 2013 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Journal of Control Science and Engineering Advances in Civil Engineering Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Submit your manuscripts at www.hindawi.com Journal of Journal of Electrical and Computer Robotics Engineering Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 VLSI Design Advances in OptoElectronics International Journal of Modelling & Aerospace International Journal of Simulation Navigation and in Engineering Engineering Observation Hindawi Hindawi Hindawi Hindawi Volume 2018 Volume 2018 Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com www.hindawi.com www.hindawi.com Volume 2018 International Journal of Active and Passive International Journal of Antennas and Advances in Chemical Engineering Propagation Electronic Components Shock and Vibration Acoustics and Vibration Hindawi Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Healthcare Engineering Hindawi Publishing Corporation http://www.deepdyve.com/lp/hindawi-publishing-corporation/classification-of-targets-and-distractors-present-in-visual-hemifields-X7adMrsjpf

Loading next page...

References

References for this paper are not available at this time. We will be adding them shortly, thank you for your patience.

Publisher: Hindawi Publishing Corporation
Copyright: Copyright © 2018 Sweeti et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
ISSN: 2040-2295
eISSN: 2040-2309
DOI: 10.1155/2018/9213707
Publisher site: See Article on Publisher Site

Abstract

Hindawi Journal of Healthcare Engineering Volume 2018, Article ID 9213707, 10 pages https://doi.org/10.1155/2018/9213707 Research Article Classification of Targets and Distractors Present in Visual Hemifields Using Time-Frequency Domain EEG Features 1 1 2 1 3 Sweeti , Deepak Joshi, B. K. Panigrahi , Sneh Anand, and Jayasree Santhosh Centre for Biomedical Engineering, IIT Delhi, New Delhi, India Department of Electrical Engineering, IIT Delhi, New Delhi, India Department of Computer Engineering & Computer Science, Manipal International University, Putra Nilai, Malaysia Correspondence should be addressed to Sweeti; sweeti.bme@gmail.com Received 29 August 2017; Revised 17 January 2018; Accepted 1 February 2018; Published 1 April 2018 Academic Editor: Feng-Huei Lin Copyright © 2018 Sweeti et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. This paper presents a classiﬁcation system to classify the cognitive load corresponding to targets and distractors present in opposite visual hemiﬁelds. The approach includes the study of EEG (electroencephalogram) signal features acquired in a spatial attention task. The process comprises of EEG feature selection based on the feature distribution, followed by the stepwise discriminant analysis- (SDA-) based channel selection. Repeated measure analysis of variance (rANOVA) is applied to test the statistical signiﬁcance of the selected features. Classiﬁers are developed and compared using the selected features to classify the target and distractor present in visual hemiﬁelds. The results provide a maximum classiﬁcation accuracy of 87.2% and 86.1% and an average classiﬁcation accuracy of 76.5 ± 4% and 76.2 ± 5.3% over the thirteen subjects corresponding to the two task conditions. These correlates present a step towards building a feature-based neurofeedback system for visual attention. 1. Introduction [5]. A normal human visual system takes about 150 ms to process the visuals [6]. Primarily two known approaches to study visual attention exist: stimulus-driven (bottom-up) In the past decade, motor brain-computer interface (BCI) has witnessed signiﬁcant progress in the area of rehabilitation. and self-driven (top-down) [5] that follow the ventral and With the promising results in motor BCI, several new direc- dorsal pathways, respectively [7]. The attention capacity of tions in this ﬁeld are emerging. Cognitive BCI (cBCI) is one a subject is limited, and the subject can attend two tasks in of them that has drawn the attention of researchers working parallel using its ability of attention shift. It is supposed to work like a spotlight phenomenon and processes only those in the area of BCI. Examples of cognitive signals used for cBCI are the subject’s motivation [1], attention orientation, objects that come under this spotlight [8]. Characteristics of and mental calculation [2]. The fundamental long-term goal human attention mechanism can be understood from the of cBCI is to develop therapeutic tools for the treatment of various theories given in the past like capacity limit theory, cognitive disorders. attention shift theory, and early and late selection theory Attention is an important entity for cBCI which works [9]. Due to the potential application of selective visual atten- with diﬀerent sensory inputs, like visual, auditory, and tactile. tion in cBCI, it is apparent to explore further the various Cognitive attention oﬀers a person the ability to select the other aspects of attention. object of interest by ignoring distractors [3]. The information In the current scenario, researchers are using advanced gained with this selection goes to the higher processing modalities like fMRI, PET, and fNIRS [5, 10] in cognitive mechanism that works like a metacognitive system and neuroscience research. These techniques are noninvasive results in perception, recognition, and memory formation and oﬀer a good spatial resolution. However, these tech- [4]. The attention paid by a subject to a visual object present niques suﬀer lower temporal resolution and higher cost. at a speciﬁc spatial location is called visual spatial attention The present study uses the EEG data to study the selective 2 Journal of Healthcare Engineering visual attention due to its good temporal resolution and low 2. Materials and Methods cost. The aim of this paper is to develop a classiﬁer that can 2.1. Task Description. This study uses the publically available classify the target and distractor present in opposite visual EEG dataset acquired by Jeanne Townsend in the laboratory hemiﬁelds using EEG features. These EEG features vary with of Eric Courchesne at UCSD. Figure 1(a) gives an illustration the task conditions and cognitive load provided. Delta of one sequence of the task trial presented to the subjects. The subband provides the best separation between low and high task presented ﬁve spatial locations represented by square cognitive loads [11]. Previous works suggest that theta band boxes placed from left to right and located 0.8 cm above the oscillations in the hippocampocortical feedback loop reﬂect central midline of the computer screen along with a center the encoding of new information. Upper alpha oscillations ﬁxation cross present on the computer screen. During each in the thalmo-cortical feedback loop reﬂect the search and 76-second task block, subjects focused covertly at the retrieval process in long-term memory [12]. Alpha power is attended stimulus location which is presented with a diﬀer- higher for target intake during the attention task while beta ently shaded box. The attended location is shown in band power increases preceding the correct response and Figure 1(a) by a dashed square box for the illustration pur- does not change in case of the erroneous response [13, 14]. pose. In each task block, 100 disk stimuli appeared randomly Gamma band power increases in the contralateral hemiﬁeld at diﬀerent spatial locations bounded in the square box. when the subject attends to a stimulus [15, 16]. Temporal Stimuli appeared for a duration of 117 ms at a diﬀerent evaluation of alpha and beta during spatial visual task spatial location with an interstimulus duration of 225– indicates a decrease in alpha band power in a time window 1000 ms [23]. Subjects responded via a thumb switch, as of 375–500 ms and an increase in beta band power of 500– soon as stimuli appeared at the attended location. The next 875 ms [17]. Also, the modulation of EEG rhythms exists dur- sequence begins after this. Data was recorded for such ing preparatory attention interval. Preparatory attention is thirty task blocks. This paper analyzes the activity with associated with the alpha power decrease in the left and right respect to the two conditions where the subject attends to temporal and occipital areas and the beta power decrease in the target present in the left and right visual hemiﬁelds the bilateral occipital, left frontal, and middle frontal occipital while ignoring the distractor present in right and left visual [16, 18]. Some other EEG features used for the cognitive load hemiﬁelds, respectively. measurement include log variance, Hjorth parameters [19], spectral entropy, spectral edge frequency, intensity weighted mean frequency, and intensity weighted bandwidth [11]. 2.2. EEG Recording and Signal Processing. This study uses the Despite previous signiﬁcant research in cognition-related publically available EEG dataset acquired by Jeanne Town- studies, it is of utmost importance to extrapolate EEG-based send in the laboratory of Eric Courchesne at UCSD. EEG research in attentional orientation for classiﬁcation purposes data collected from 13 (two female, eleven male; ages 22–40 in neurofeedback applications. In most of the target versus years) healthy right-handed subjects performing a visual nontarget classiﬁcation problems, the task protocols involve spatial attention task was recorded from 29 scalp electrode the targets and distractors appearing at the central visual ﬁeld locations using an EEG standard electrode cap (Electro-Cap or center of the screen. Such tasks imitate a very diﬀerent sce- International Inc.) and two EOG (electrooculogram) elec- nario than the real-life situations where target and distractors trodes with a sampling frequency of 512 Hz. Figure 1(b) may appear at diﬀerent spatial locations or visual ﬁelds. The shows the location of EEG scalp electrodes placed for record- present work involves the classiﬁcation of the target and dis- ing data. The data were collected with reference to the right tractor present in diﬀerent visual hemiﬁelds while the subject mastoid electrode position within the analog passband of attends the objects using peripheral visual attention. This is 0.01–50 Hz. Additional digital ﬁltering was done using more like a real-life situation, for example, looking at the 4th-order Butterworth band-pass ﬁlter in the range of road while attending to or ignoring the objects appearing 0.1–45 Hz, and the signal was average rereferenced. The on both sides of the road. This paper explores the two task Automatic Artifact Removal (AAR) v1.3 toolbox based on conditions: the ﬁrst task condition compares the activity the blind source separation principle was used to remove while the subject attends to the target object present in the the eye blinks and muscle artifacts [24]. EEG analysis left visual hemiﬁeld, ignoring the distractor present in the involved an epoch length of 800 ms, starting 200 ms before right visual hemiﬁeld. The second task condition compares the stimulus onset to 600 ms after the stimulus oﬀset, over the activity while the subject attends to the target object 90 trials from each subject. Figure 1(b) shows the scalp present in the right visual hemiﬁeld, ignoring the distractor electrode locations. present in the left visual hemiﬁeld. The selected data is divided into four datasets, namely A, In this work, we will explore diﬀerent EEG features and B, C, and D; the summary is given in Table 1. Dataset A cor- classiﬁers are developed further using most relevant features. responds to the activity while the subject attends to the left The results would facilitate the development of a neurofeed- visual hemiﬁeld and the target appears at the same location back system for selective visual attention [20–22]. Section 2 in the left visual hemiﬁeld. Dataset B corresponds to activity describes the materials and methods used in the study: task while the subject attends to the left visual ﬁeld and the description, data preprocessing method, and methodology distractor appears in the right visual hemiﬁeld. Dataset C followed. Section 3 gives the details of channel selection, sta- represents the activity while the subject attends to the right tistical analysis, and classiﬁcation performed. Section 4 dis- visual hemiﬁeld and the distractor appears in the left visual cusses the results. ﬁeld. Dataset D represents the activity while the subject Journal of Healthcare Engineering 3 Channel locations F3 F4 FZ 117 ms FC5 FC6 FC1 FC2 Attended location in T7 C4 C3 CZ T8 shaded square CP1 CP2 CP5 CP6 Target at attended PZ location P3 P4 P7 P8 PO3 POZ PO4 117 ms PO7 PO8 O1 OZ O2 Attended location in shaded square (a) (b) Figure 1: (a) A sequence of the task trial. (b) EEG scalp electrode placement locations. Table 1: Summary of four datasets selected for two task conditions. Task condition 1 Task condition 2 Task condition Dataset A Dataset B Dataset C Dataset D Attending target object Ignoring distractor objects Ignoring distractor objects Attending target object Patient state in left hemiﬁeld in right hemiﬁeld in left hemiﬁeld in right hemiﬁeld Eye ﬁxation Center Center Center Center Electrode type Surface Surface Surface Surface International 10-20 International 10-20 International 10-20 International 10-20 Electrode placement placement system placement system placement system placement system Number of subjects 13 13 13 13 Number of electrodes 31 31 31 31 Number of trials from 90 90 90 90 each subject Epoch duration 800 ms 800 ms 800 ms 800 ms attends to the right visual ﬁeld and the target appears at the (1) Hjorth complexity (H ): it is a measure of the complexity same location in the right visual hemiﬁeld. Task condition spread of the spectrum and represents the change in 1 compares the dataset A and dataset B. The task condition frequency [25]. 2 compares the dataset C and dataset D. In other words, 2 2 the ﬁrst task condition compares the activity while the std d X/dt std X subject attends to the target object present in the left visual H = , 1 complexity std dX/dt hemiﬁeld ignoring the distractor present in the right visual hemiﬁeld. The second task condition compares the activity while the subject attends to the target object present in the where std is standard deviation function. right visual hemiﬁeld ignoring the distractor present in the left visual hemiﬁeld. Figure 2 demonstrates the ﬂow chart (2) Hjorth mobility (H ): it is a measure of mean mobility of the complete signal analysis process. frequency. 2.3. Feature Extraction. EEG features studied are discussed std dX/dt brieﬂy in this section. These features were selected based on H = , 2 mobility std Xt their applications in biomedical signal processing as dis- cussed in Introduction. In this work, we are trying to explore the utility of these features in attentional studies: where std is standard deviation function. Button Next ISI First sequence sequence press 225−1000 ms 4 Journal of Healthcare Engineering (6) Median power frequency (mpf): it is the frequency Multichannel EEG data below which 50% of the total power of the signal is present, calculated over the half of the total area of the power spectrum. Preprocessing (filtering and artifact removal) (7) Spectral edge frequency (sef): it is the frequency below which 95% of the total power of the signal is Single trial analysis present, calculated over 95% of the total area of the power spectrum. Feature calculation 2.4. Channel Selection. Channel selection is an important step as it helps in selecting the channels that can distinguish two or more datasets. This procedure can help in reducing the computational burden by minimizing the number of Channel selection channels. A number of channel selection methods including ﬁlters, wrappers, and embedded methods exist. Among these methods, wrapper methods are good at providing a Statistical analysis reliable set of features. The stepwise discriminant analysis (SDA) used in this study is a wrapper method that gener- ates a reliable set of features with multivariate analysis of variance (MANOVA). Classification and conclusion A data matrix (C ) is created for each subject corre- sponding to each EEG feature. The matrix C is composed Figure 2: Flowchart for feature-based EEG data analysis. of 29 channels and a total of 180 observations with respect to targets and nontargets for both the task conditions. Rows of the matrix represent the observations, and columns repre- (3) Average frequency: it deﬁnes the number of times the sent the channels as shown: signal crosses the zero value. a … a i,j i,j+n Total zero crossing points Average frequency = C = ⋮⋮ ⋮ , 6 Epoch duration a … a 3 i+m,j i+m,j+n where i and j represent the rows and column, respectively, m (4) Lempel-Ziv complexity (lz-complexity): it is a represents the number of observations, and n represents the harmonic variability metric that shows the distinct number of channels. pattern contained in the sequence as an algorithm The matrix C for each EEG feature, for a single subject, scans the data sequence from left to right [25]. To is pooled to perform SDA to select the three best performing compute the lz-complexity, the signal needs to be channels. MANOVA compares the sample means based decoded ﬁrst with respect to some threshold value on the variance-covariance between variables to test the which could be mean or median of the signal. In this signiﬁcance diﬀerence. MANOVA gives the signiﬁcant dif- way, the signal greater than the threshold which ference value represented by lambda (λ). SDA present the maps to 1 else to 0 to obtain the symbolic sequence channels in descending order of discriminating power λ, is further parsed to obtain the encoded sequence and the number of channels can further be selected as [26]. For an encoded sequence, s(n) of length n, the desired based on the lambda power. In this work, we are lz-complexity can be obtained as below: selecting three channels. 2.5. Statistical Analysis. Repeated measure analysis of sn Lempel − Ziv complexity = 4 variance (rANOVA) was performed to analyze the statistical signiﬁcance of selected EEG features diﬀerentiating the two datasets within the two task conditions. rANOVA investi- (5) Band power: band power can be calculated from the gates the EEG features and task condition interaction over power spectrum of the signal. Power spectrum repeated measurements for the two task conditions. It is used represents the energy of a signal over the frequency to test null hypotheses about the mean. If the mean of the two components that it possesses. For a signal x, power classes is diﬀerent, then the null hypothesis rejects. The spectrum can be calculated using results of rANOVA are presented in the following form: F(df ,df )= F value, p = p value, fc error Power spectrum = X ∗ conjugate X , 5 where df = degree of freedom of feature and task fft fft fc condition interaction, where X represents the Fourier transform of the signal x. df degree of freedom of error, fft error = Journal of Healthcare Engineering 5 F = critical value, Table 2: Selected channels corresponding to the eight EEG features. p = signiﬁcance value. Task condition 1 Task condition 2 This test oﬀers Greenhouse–Geisser (pGG), Huynh– Feature Selected channels Selected channels Feldt (pHF), and lower bound (pLB) corrections for multiple (using SDA method) (using SDA method) comparisons to avoid false rejection of the null hypothesis. Average frequency C3, CP2, PZ C3, CP2, P3 lz-complexity CP1, PZ, P4 CP2, P3, P4 2.6. Classiﬁcation and Cross-Validation. In the present work, we need to classify the two classes of the attentional load Complexity CP1, PZ, P4 CP5, CP1, PZ from targets and distractors. To solve this two-class classi- Mobility CP1, PZ, P4 CP1, P3, PZ ﬁcation problem, this study compares the three classiﬁca- Median power CP2, PZ, P4 CP1, P3, PZ tion approaches, namely, artiﬁcial neural network (ANN), frequency K-nearest neighbor (KNN), and support vector machine Spectral edge CP1, PZ, P4 CP1, PZ, P4 (SVM). The input and target data matrices were constructed frequency for the subject by taking 90 observations from each channel Delta power CP2, PZ, P4 CP1, P3, PZ with respect to each dataset. Features were normalized to Beta power CP1, CP2, PZ CP1, CP2, P3 zero mean and unit variance before classiﬁcation. Artiﬁcial neural network (ANN): a pattern recognition network architecture with three network layers; namely, the perform input layer, the hidden layer, and the output layer were used. ed to avoid the eﬀect of any time-related change There were 29 nodes in the input layer, ten nodes in the in data that may occur during EEG data recording in a hidden layer, and one node in the output layer. The tangent block design. sigmoid activation function was used. The network was Receiver operating characteristic (ROC) curves were used trained with the Levenberg–Marquardt back propagation to compare the performance of the diﬀerent classiﬁers. ROC method of training using a preset amount of data for the is a plot of the two operating characteristics known as a true training, testing, and validation. The model was validated positive rate (TPR) and the false positive rate (FPR), where using cross-validation methods. TPR is the probability of detection while FPR gives the K-Nearest neighbor (KNN): the developed KNN model probability of false alarm. was ﬁtted by means of the Euclidean distance metric between the two nearest neighbors. This model selects the neighbors 3. Results with a known class from the training dataset and assigns weights to it according to the distance to the space variable The preprocessed datasets corresponding to the two task using the exhaustive searcher. It made the decision with the conditions were analyzed with the suggested methodology. majority vote given by selected nearest neighbors. For a space First, to reduce the system complexity, the channel selection variable X, the classiﬁer looks for the nearest neighbor among is performed. Feature selection results are given in Table 2. diﬀerent classes, say A and B, and assigns the class label Selected channels are presented in the descending order having the smallest distance to the space variable X. of their discriminating powers. These results suggest that Support vector machine (SVM): it is a supervised the parietal and central parietal region electrodes are classiﬁcation model that classiﬁes the data by ﬁnding the among the best performing channels to distinguish the targets best hyperplane oﬀering the largest margin between two and distractors. classes. These hyperplanes are the decision planes that Further, the distribution of the diﬀerent EEG features can separate the objects having diﬀerent class member- corresponding to the two task conditions with selected chan- ships. The developed SVM model mapped the predictor nels is studied. This distribution is shown in Figure 3. data using radial basis kernel function. Sequential minimal Figures 3(a) and 3(i) show that the average frequency value optimization (SMO) approach was used to solve this is lower for the targets in both task conditions. lz-complexity, binary classiﬁcation problem. which gives the harmonic variability metric, oﬀers greater value for nontargets, suggesting distinct patterns in the signal 2.7. Cross-Validation. The developed classiﬁer models were while ignoring the distractors as shown in Figures 3(b) and cross-validated using the k-fold cross-validation approach. 3(j). Hjorth complexity feature’s value, which represents the In this approach, the data is divided into k subsamples change in frequency, is higher for targets as given in randomly. At each fold, (k− 1) subsamples are used for the Figures 3(c) and 3(k). It shows that frequency spread is more training, remaining one for the testing. The process is while attending to targets than nontargets. Figures 3(d) and repeated for each fold, and mean accuracy is calculated from 3(l) show that Hjorth mobility, which represents the mean the average of the results of diﬀerent folds. Also, another frequency, is higher for nontargets. Median power frequency cross-validation approach is suggested. This approach works embodying 50% of the total signal power is higher for non- similar to k-fold cross-validation except that the criteria for targets as shown in Figures 3(e) and 3(m). Figures 3(f) and the selection of training and testing datasets is diﬀerent. 3(n) show that the spectral edge frequency that represents Unlike the k-fold approach, the datasets are not selected ran- 95% of the total power is also high while ignoring the nontar- domly but are selected in a way that there is a maximum time gets. Delta power values are higher for targets for both task separation in the data points. This cross-validation was conditions as given in Figures 3(g) and 3(o). Figures 3(h) 6 Journal of Healthcare Engineering Average frequency Lempel-Ziv complexity Average frequency Lempel-Ziv complexity (datasets A versus B) (datasets A versus B) (datasets C versus D) (datasets C versus D) 12 20 12 16 10 10 15 12 8 8 6 10 6 8 4 4 5 4 2 2 0 0 0 0 C3 CP2 PZ CP1 P4 P4 C3 CP2 P3 CP2 P3 P4 Channel number Channel number Channel number Channel number (a) (b) (i) (j) Hjorth complexity Hjorth mobility Hjorth complexity Hjorth mobility (datasets A versus B) (datasets A versus B) (datasets C versus D) (datasets C versus D) 3.5 0.2 3.5 0.2 0.15 0.15 2.5 2.5 2 0.1 0.1 1.5 1.5 0.05 0.05 0.5 0.5 0 0 0 0 CP1 PZ P4 CP1 PZ P4 CP5 CP1 PZ CP1 P3 PZ Channel number Channel number Channel number Channel number (c) (d) (k) (l) Median power frequency Spectral edge frequency Median power frequency Spectral edge frequency (datasets A versus B) (datasets A versus B) (datasets C versus D) (datasets C versus D) 10 35 10 35 30 30 8 8 25 25 6 6 20 20 15 15 4 4 10 10 2 2 5 5 0 0 0 0 CP2 PZ P4 CP1 PZ P4 CP1 P3 PZ CP1 PZ P4 Channel number Channel number Channel number Channel number (e) (f ) (m) (n) Delta power Beta power Delta power Beta power (datasets A versus B) (datasets A versus B) (datasets C versus D) (datasets C versus D) 80 30 80 35 60 60 40 15 40 20 20 0 0 0 0 CP2 PZ P4 CP1 CP2 PZ CP1 CP3 PZ CP1 CP2 P3 Channel number Channel number Channel number Channel number (g) (h) (o) (p) Target Distractor Distractor Target (I) (II) Figure 3: Feature distribution over selected channels for (I) task condition 1 and (II) task condition 2: (a), (i) average frequency; (b), (j) lz- complexity; (c), (k) Hjorth complexity; (d), (l) Hjorth mobility; (e), (m) median power frequency; (f), (n) spectral edge frequency; (g), (o) delta power; (h), (p) beta power. and 3(p) give higher beta power for distractors that advise Next, the repeated measure analysis of variance more activation of the attentional network in inhibiting the (rANOVA) was performed for all the EEG features extracted nontargets. It can be concluded from Figure 3 that Hjorth from the two task conditions. It was performed using complexity and delta power values are higher for targets cor- three selected channel features, two datasets, and 90 obser- responding to the task conditions 1 and 2 while other features vations from each dataset. The degree of freedom (DF) show lower values. was (3− 1) = 2 for channel features, (3− 1) ∗ (2− 1) = 2 Complexity Delta power Median power frequency Average frequency Mobility Spectral edge frequency Iz-complexity Beta power Median power frequency Complexity Delta power Average frequency Mobility Iz-complexity Beta power Spectral edge frequency Journal of Healthcare Engineering 7 using Hjorth complexity, Hjorth mobility, delta power, for channel feature-condition interaction, and (180 – 2) ∗ (3− 1) = 356 for error. The rANOVA results show that and beta power, respectively. four among the extracted EEG features exhibit a signiﬁ- cant diﬀerence for channel feature and condition interac- tion with a signiﬁcance level of p <0 1. The result 4. Discussion indicates a signiﬁcant channel feature and task condition interaction with lower bound correction for Hjorth com- Selective visual attention is the ability to select the visual plexity [F(2,356) = 6.0145, p =0 0151], Hjorth mobility information of interest present in the visual ﬁeld. It is a key [F(2,356) = 5.62, p =0 0188], delta power [F(2,356) = to various other skills like perception and recognition and 3.8754, p =0 0505], and beta power [F(2,356) = 3.7965, memory as well; it can also aﬀect these skills if there is a prob- p =0 0529] over the task condition 1. On the other hand, other lem with it. This paper studies the EEG correlates of visual features including average frequency [F(2,356) = 0.42023, attention in a spatial attention task. It is important to study p =0 51766], lz-complexity [F(2,356) = 0.68079, p =0 41042], features of the EEG as these are very crucial and provide mpf [F(2,356) = 1.1707, p =0 28073], and sef [F(2,356) = more information than the raw data. These features are the 2.3163, p =0 1298] show no statistically signiﬁcant diﬀer- potential candidates that can be used in neurofeedback ence in the target and distractor. For task condition 2, systems to give feedback about their performance to the sub- the statistical results display a signiﬁcant channel feature jects. The present study attempts to ﬁnd the EEG correlates and task condition interaction with lower bound correc- of attention for the task when the subject attends to target tion for Hjorth complexity [F(2,356) = 3.2531, p =0 0729], objects present in one visual hemiﬁeld while ignoring distrac- Hjorth mobility [F(2,356) = 5.0276, p =0 0261], delta power tor objects present in another visual hemiﬁeld. To reduce the [F(2,356) = 3.5058, p =0 0627], and beta power [F(2,356) = system complexity and increase classiﬁcation accuracy, chan- 2.7321, p =0 0900]. There is no statistically signiﬁcant channel nel selection is performed [27]. Channel selection performed feature and task condition interaction over the average over EEG features suggest that the channels with the most frequency [F(2,356) = 1.012, p =0 31579], lz-complexity discriminating power lie in the central-parietal and parietal [F(2,356) = 0.83486, p =0 36211], mpf [F(2,356) = 1.937, regions, which are involved in the visual-spatial processing [28]. The rANOVA-based statistical analysis found that p =0 16573], and sef [F(2,356) = 2.4043, p =0 12278]. The statistical analysis concludes that the Hjorth amongst the features studied, Hjorth complexity, Hjorth complexity, Hjorth mobility, delta power, and beta power mobility, delta power, and beta power can signiﬁcantly diﬀer- can signiﬁcantly diﬀerentiate the activity while the subject entiate the datasets corresponding to the two task conditions. attends to stationary targets and inhibits the distractors This selection suggests higher beta and lower delta for non- present in another visual hemiﬁeld. These selected EEG fea- targets, representing higher cognitive demand or working tures are further explored to develop a classiﬁer to classify memory load for inhibition which agrees with earlier studies the target and distractor present in diﬀerent visual hemiﬁelds. involving targets and nontargets [29]. Hjorth features have Feature matrices with three selected channel features and been used earlier for cognitive load measurement [19]. The 180 observations were prepared for each subject correspond- present study explores these features and shows higher ing to all the EEG features for the target and distractor clas- Hjorth mobility and lower Hjorth complexity for nontargets siﬁcations. A comparison of the three diﬀerent classiﬁers which correspond to the mean frequency and change in using the four EEG features for a subject is presented in this frequency, respectively. section. The ROC curves of Figure 4 illustrate that the artiﬁ- The classiﬁcation system is further developed using cial neural network (ANN) classiﬁes the target and distractor selected features to distinguish the activities corresponding better than the other two methods. So the details of the to targets and nontargets. The importance of such classiﬁ- classiﬁcation performance parameters, namely, sensitivity, cation lies in applications like cognitive brain-computer speciﬁcity, and accuracy corresponding to the ANN classiﬁer interface or neurofeedback system for training where the only, are given in Tables 3 and 4. cognitive control measures are used to control the BCI Tables 3 and 4 show the mean and maximum values of and train the subjects. Classiﬁcation accuracy in cBCI is the classiﬁcation results obtained from thirteen subjects for limited by various internal and external factors, like sen- the two task conditions using four diﬀerent EEG features. sory and cognitive, comparative to reasonable accuracy in The classiﬁcation was performed for each and every subject, motor BCI [30, 31]. Due to this limitation, accuracy and then the mean was taken over the classiﬁcation results reported in previous research was restricted to only 75% obtained from a population of 13 subjects. These mean values and 79% in spatial attention tasks [32, 33] using noninva- and standard deviation for sensitivity, speciﬁcity, and sive techniques. We could reach a maximum accuracy of accuracy, corresponding to diﬀerent features, are given in 87.2% and 86.1% and a mean accuracy of 76.5% and Tables 3 and 4. These tables also give the maximum value 76.2% over thirteen subjects for the two task conditions, respectively, by using the EEG-based noninvasive method. of the classiﬁcation results obtained. Table 3 spectates a maximum classiﬁcation accuracy of 80.6%, 87.2%, 82.2%, In this way, a classiﬁer is developed that can classify the and 80% for the task condition 1 using Hjorth complexity, peripheral attention paid to targets and distractors present Hjorth mobility, delta power, and beta power, respectively. in diﬀerent visual hemiﬁelds. Such a classiﬁer can facilitate Table 4 shows a maximum classiﬁcation accuracy of the development of an EEG feature-based neurofeedback system for attention [34]. 84.4%, 86.1%, 83.3%, and 86.1% for the task condition 2 8 Journal of Healthcare Engineering ROC curve for task condition 1 ROC curve for task condition 2 ROC curve for task condition 1 (Hjorth complexity) (Hjorth complexity) (Hjorth mobility) 1 1 1 0.9 0.9 0.9 0.8 0.8 0.8 0.7 0.7 0.7 0.6 0.6 0.6 0.5 0.5 0.5 0.4 0.4 0.4 0.3 0.3 0.3 0.2 0.2 0.2 0.1 0.1 0.1 0 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 False positive rate False positive rate False positive rate ANN ANN ANN KNN KNN KNN SVM SVM SVM (a) (b) (c) ROC curve for task condition 2 ROC curve for task condition 1 ROC curve for task condition 2 (Hjorth mobility) (Delta power) (Delta power) 1 1 1 0.9 0.9 0.9 0.8 0.8 0.8 0.7 0.7 0.7 0.6 0.6 0.6 0.5 0.5 0.5 0.4 0.4 0.4 0.3 0.3 0.3 0.2 0.2 0.2 0.1 0.1 0.1 0 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 False positive rate False positive rate False positive rate ANN ANN ANN KNN KNN KNN SVM SVM SVM (d) (e) (f) ROC curve for task condition 1 ROC curve for task condition 2 (Beta power) (Beta power) 1 1 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 False positive rate False positive rate ANN ANN KNN KNN SVM SVM (g) (h) Figure 4: ROC curve for (a) and (b) Hjorth complexity, (c) and (d) Hjorth mobility, (e) and (f) delta power, and (g) and (h) beta power, using artiﬁcial neural network (ANN), K-nearest neighbor (KNN), and support vector machine (SVM) classiﬁers during the testing corresponding to task conditions 1 and 2, respectively. True positive rate True positive rate True positive rate True positive rate True positive rate True positive rate True positive rate True positive rate Journal of Healthcare Engineering 9 Table 3: Classiﬁcation results displaying % sensitivity (SN), % speciﬁcity (SP), and % accuracy (AC) obtained using ANN for task condition 1 during testing. Task condition 1 Features Mean Maximum Sensitivity Speciﬁcity Accuracy Sensitivity Speciﬁcity Accuracy Hjorth complexity 77.1 ± 7.37 73.6 ± 3.55 75.3 ± 3.72 87.8 78.9 80.6 Hjorth mobility 76.4 ± 5.13 76.6 ± 4.80 76.5 ± 4 88.9 85.6 87.2 Delta power 74.1 ± 5.67 73.0 ± 4.97 73.6 ± 3.93 86.7 80 82.2 Beta power 73.4 ± 5.24 75.8 ± 5.40 74.9 ± 2.85 84.4 83.3 80 Table 4: Classiﬁcation results displaying % sensitivity (SN), % speciﬁcity (SP), and % accuracy (AC) obtained using ANN for task condition 2 during testing. Task condition 2 Features Mean Maximum Sensitivity Speciﬁcity Accuracy Sensitivity Speciﬁcity Accuracy Hjorth complexity 73.5 ± 5.12 77.5 ± 6.65 75.5 ± 4.24 81.1 87.8 84.4 Hjorth mobility 77.5 ± 5.36 74.9 ± 6.41 76.2 ± 5.30 86.7 87.8 86.1 Delta power 72.7 ± 5.44 76.0 ± 5.55 74.3 ± 4.76 82.2 85.6 83.3 Beta power 78.6 ± 7.65 72.4 ± 6.38 75.5 ± 5.22 88.9 83.3 86.1 5. Conclusion Supplementary Materials The present study explores the EEG features that can distin- This manuscript includes a supplementary result ﬁle, which guish the targets and nontargets present in the diﬀerent contains the classiﬁcation results obtained. (Supplementary visual hemiﬁelds. A classiﬁcation system to classify the Materials) targets and distractors present in opposite visual hemiﬁelds is proposed in this paper. The analysis is done to optimize References the performance of the system. Results provide EEG correlates of selective visual attention that can classify the [1] R. A. A. S. Musallam, B. D. Corneil, B. Greger, and activities while attending targets and ignoring distractors. H. Scherberger, “Cognitive control signals for neural prosthet- Other EEG features can be explored to further increase the ics,” Science, vol. 305, no. 5681, pp. 258–262, 2004. classiﬁcation accuracy and make the possibility of feature- [2] K. Jerbi, S. Freyermuth, L. Minotti, P. Kahane, A. Berthoz, and based neurofeedback system feasible. J. P. Lachaux, “Chapter 12 watching brain TV and playing brain ball: exploring novel BCI strategies using real‐time analysis of human intracranial data,” International Review of Data Availability Neurobiology, vol. 86, pp. 159–168, 2009. [3] M. I. Posner and S. E. Petersen, “The attention system of the Data supporting this research article are available from the human brain,” Annual Review Neuroscience, vol. 13, no. 1, online Archive of Swartz Center, University of California in pp. 25–42, 1990. San Diego (UCSD), available from http://headit.ucsd.edu/ [4] P. Cavanagh, “Attention routines and the architecture of studies/2747adbc-385a-11e3-b29d-0050563f2612. selection,” in Cognitive Neuroscience of Attention,M. Posner, Ed., pp. 13–28, Guilford Press, New York, NY, Conflicts of Interest USA, 2004. [5] G. R. Mangun, “Neural mechanism of attention,” in The The authors have no conﬂict of interest to declare. Cognitive Electrophysiology of Mind and Brain, pp. 247–259, Elsevier Science (USA), 2002, http://citeseerx.ist.psu.edu/ viewdoc/download?doi=10.1.1.470.5863&rep=rep1&type=pdf. Acknowledgments [6] S. Thorpe, D. Fize, and C. Marlot, “Speed of processing in the This work was supported by The Ministry of Human human visual system,” Nature, vol. 381, no. 6582, pp. 520–522, Resource Development, Government of India. The authors would like to thank the Indian Institute of Technology [7] S. Kastner and L. G. Ungerleider, “Mechanisms of visual atten- (IIT), Delhi, for providing the facilities to carry out the tion in the human cortex,” Annual Review of Neuroscience, vol. 23, pp. 315–341, 2000. research. The authors also thank the Archive of Swartz Center, University of California in San Diego (UCSD), for [8] P. Stanford, “Chapter three: attention,” 2007, http://www- sharing their data online. psych.stanford.edu/~ashas/Cognition Textbook/chapter3.pdf. 10 Journal of Healthcare Engineering [25] A. Kumar and S. Anand, “EEG signal processing for monitor- [9] R. Desimone and J. Duncan, “Neural mechanisms of selective visual attention,” Annual Review of Neuroscience, vol. 18, no. 1, ing depth of anesthesia,” IETE Technical Review, vol. 23, no. 3, pp. 193–222, 1995. pp. 179–186, 2006. [10] S. Yantis, “The neural basis of selective attention cortical [26] J. Hu, J. Gao, and J. C. Principe, “Analysis of biomedical signals sources and targets of attentional modulation,” Current Direc- by the Lempel-Ziv complexity: the eﬀect of ﬁnite data size,” IEEE Transactions on Biomedical Engineering, vol. 53, no. 12, tions in Psychological Science, vol. 17, no. 2, pp. 86–90, 2008. pp. 2606–2609, 2006. [11] P. Zarjam, J. Epps, and F. Chen, “Spectral EEG features for evaluating cognitive load,” Annual International Conference [27] Y. Li and Y. Koike, “A real-time BCI with a small number of channels based on CSP,” Neural Computing and Applications, of the IEEE Engineering in Medicine and Biology Society, vol. 2011, pp. 3841–3844, 2011. vol. 20, no. 8, pp. 1187–1192, 2011. [12] W. Klimesch, “EEG alpha and theta oscillations reﬂect cogni- [28] L. Shen, X. Hu, E. Yacoub, and K. Ugurbil, “Neural correlates tive and memory performance: a review and analysis,” Brain of visual form and visual spatial processing,” Human Brain Research Reviews, vol. 29, no. 2-3, pp. 169–195, 1999. Mapping, vol. 8, no. 1, pp. 60–71, 1999. [13] D. A. Valentino and R. L. Dufresne, “Attention tasks and EEG [29] M. C. Hout and S. D. Goldinger, “Learning in repeated visual power spectra,” International Journal of Psychophysiology, search,” Attention, Perception & Psychophysics, vol. 72, no. 5, vol. 11, no. 3, pp. 299–301, 1991. pp. 1267–1282, 2010. [14] M. Gola, M. Magnuski, I. Szumska, and A. Wróbel, “EEG beta [30] T. Nguyen, A. Khosravi, D. Creighton, and S. Nahavandi, band activity is related to attention and attentional deﬁcits in “EEG data classiﬁcation using wavelet features selected by Wil- the visual performance of elderly subjects,” International coxon statistics,” Neural Computing and Applications, vol. 26, Journal of Psychophysiology, vol. 89, no. 3, pp. 334–341, 2013. no. 5, pp. 1193–1202, 2015. [15] M. M. Muller, T. Gruber, and A. Keil, “Modulation of induced [31] S. Chen, Z. Luo, and H. Gan, “An entropy fusion method for gamma band activity in the human EEG by attention and feature extraction of EEG,” Neural Computing and Applica- visual information processing,” International Journal of tions, 2016. Psychophysiology, vol. 38, no. 3, pp. 283–299, 2000. [32] M. S. Treder, A. Bahramisharif, N. M. Schmidt, M. A. J. van [16] S. Meyberg, M. Werkle-bergner, W. Sommer, and O. Dimigen, Gerven, and B. Blankertz, “Brain-computer interfacing using “Microsaccade-related brain potentials signal the focus of modulations of alpha activity induced by covert shifts of atten- visuospatial attention,” NeuroImage, vol. 104, pp. 79–88, 2015. tion,” Journal of Neuroengineering and Rehabilitation, vol. 8, no. 1, pp. 24–29, 2011. [17] M. V. Marrufo, E. Vaquero, M. J. Cardoso, and C. M. Gómez, “Temporal evolution of α and β bands during visual spatial [33] H. Morioka, A. Kanemura, S. Morimoto et al., “Decoding attention,” Cognitive Brain Research, vol. 12, no. 2, pp. 315– spatial attention by using cortical currents estimated from 320, 2001. electroencephalography with near-infrared spectroscopy prior information,” NeuroImage, vol. 90, pp. 128–139, 2014. [18] C. M. Gómez, J. Marco-Pallarés, and C. Grau, “Location of brain rhythms and their modulation by preparatory attention [34] G. A. Alvarez and P. Cavanagh, “Independent resources for estimated attentional tracking in the left and right visual hemiﬁelds,” by current density,” Brain Research, vol. 1107, no. 1, pp. 151–160, 2006. Psychological Science, vol. 16, no. 8, pp. 637–643, 2005. [19] D. Das, D. Chatterjee, and A. Sinha, “Unsupervised approach for measurement of cognitive load using EEG signals,” in 2013 IEEE 13th International Conference on Bioinformatics and Bioengineering (BIBE), pp. 1–6, Chania, Greece, 2013. [20] N. F. Ramsey, M. P. V. D. Heuvel, K. H. Kho, and F. S. S. Leijten, “Towards human BCI applications based on cogni- tive brain systems: an investigation of neural signals recorded from the dorsolateral prefrontal cortex,” IEEE Transactions on Neural Systems and Rehabilitation Engi- neering, vol. 14, no. 2, pp. 214–217, 2006. [21] M. J. Vansteensel, D. Hermes, E. J. Aarnoutse et al., “Brain- computer interfacing based on cognitive control,” Annals of Neurology, vol. 67, no. 6, pp. 809–816, 2010. [22] E. Astrand, C. Wardak, and S. Ben Hamed, “Selective visual attention to drive cognitive brain-machine interfaces: from concepts to neurofeedback and rehabilitation applications,” Frontiers in Systems Neuroscience, vol. 8, pp. 1–16, 2014. [23] S. Makeig, M. Westerﬁeld, J. Townsend, T. P. Jung, E. Courchesne, and T. J. Sejnowski, “Functionally independent components of early event-related potentials in a visual spatial attention task,” Philosophical Transactions of the Royal Society B: Biological Sciences, vol. 354, no. 1387, pp. 1135–1144, 1999. [24] G. Gómez-Herrero, “Automatic Artifact Removal (AAR) Toolbox v1.3 (Release 09.12.2007) for MATLAB,” 2007, http://germangh.github.io/pubs/aardoc07.pdf. International Journal of Advances in Rotating Machinery Multimedia Journal of The Scientific Journal of Engineering World Journal Sensors Hindawi Hindawi Publishing Corporation Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 http://www www.hindawi.com .hindawi.com V Volume 2018 olume 2013 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Journal of Control Science and Engineering Advances in Civil Engineering Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Submit your manuscripts at www.hindawi.com Journal of Journal of Electrical and Computer Robotics Engineering Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 VLSI Design Advances in OptoElectronics International Journal of Modelling & Aerospace International Journal of Simulation Navigation and in Engineering Engineering Observation Hindawi Hindawi Hindawi Hindawi Volume 2018 Volume 2018 Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com www.hindawi.com www.hindawi.com Volume 2018 International Journal of Active and Passive International Journal of Antennas and Advances in Chemical Engineering Propagation Electronic Components Shock and Vibration Acoustics and Vibration Hindawi Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Journal

Journal of Healthcare Engineering – Hindawi Publishing Corporation

Published: Apr 1, 2018

References

Cognitive control signals for neural prosthetics,

R. A. A. S. Musallam, B. D. Corneil, B. Greger, and H. Scherberger
Chapter 12 watching brain TV and playing brain ball: exploring novel BCI strategies using real‐time analysis of human intracranial data,

K. Jerbi, S. Freyermuth, L. Minotti, P. Kahane, A. Berthoz, and J. P. Lachaux
The attention system of the human brain,

M. I. Posner and S. E. Petersen
Attention routines and the architecture of selection,

P. Cavanagh
Neural mechanism of attention,

G. R. Mangun
Speed of processing in the human visual system,

S. Thorpe, D. Fize, and C. Marlot
Mechanisms of visual attention in the human cortex,

S. Kastner and L. G. Ungerleider
Chapter three: attention,

P. Stanford
Neural mechanisms of selective visual attention,

R. Desimone and J. Duncan
The neural basis of selective attention cortical sources and targets of attentional modulation,

S. Yantis
Spectral EEG features for evaluating cognitive load,

P. Zarjam, J. Epps, and F. Chen
EEG alpha and theta oscillations reflect cognitive and memory performance: a review and analysis,

W. Klimesch
Attention tasks and EEG power spectra,

D. A. Valentino and R. L. Dufresne
EEG beta band activity is related to attention and attentional deficits in the visual performance of elderly subjects,

M. Gola, M. Magnuski, I. Szumska, and A. Wróbel
Modulation of induced gamma band activity in the human EEG by attention and visual information processing,

M. M. Muller, T. Gruber, and A. Keil
Microsaccade-related brain potentials signal the focus of visuospatial attention,

S. Meyberg, M. Werkle-bergner, W. Sommer, and O. Dimigen
Temporal evolution of α and β bands during visual spatial attention,

M. V. Marrufo, E. Vaquero, M. J. Cardoso, and C. M. Gómez
Location of brain rhythms and their modulation by preparatory attention estimated by current density,

C. M. Gómez, J. Marco-Pallarés, and C. Grau
Unsupervised approach for measurement of cognitive load using EEG signals,

D. Das, D. Chatterjee, and A. Sinha
Towards human BCI applications based on cognitive brain systems: an investigation of neural signals recorded from the dorsolateral prefrontal cortex,

N. F. Ramsey, M. P. V. D. Heuvel, K. H. Kho, and F. S. S. Leijten
Brain-computer interfacing based on cognitive control,

M. J. Vansteensel, D. Hermes, E. J. Aarnoutse et al.
Selective visual attention to drive cognitive brain-machine interfaces: from concepts to neurofeedback and rehabilitation applications,

E. Astrand, C. Wardak, and S. Ben Hamed
Functionally independent components of early event-related potentials in a visual spatial attention task,

S. Makeig, M. Westerfield, J. Townsend, T. P. Jung, E. Courchesne, and T. J. Sejnowski
Automatic Artifact Removal (AAR) Toolbox v1.3 (Release 09.12.2007) for MATLAB,

G. Gómez-Herrero
EEG signal processing for monitoring depth of anesthesia,

A. Kumar and S. Anand
Analysis of biomedical signals by the Lempel-Ziv complexity: the effect of finite data size,

J. Hu, J. Gao, and J. C. Principe
A real-time BCI with a small number of channels based on CSP,

Y. Li and Y. Koike
Neural correlates of visual form and visual spatial processing,

L. Shen, X. Hu, E. Yacoub, and K. Ugurbil
Learning in repeated visual search,

M. C. Hout and S. D. Goldinger
EEG data classification using wavelet features selected by Wilcoxon statistics,

T. Nguyen, A. Khosravi, D. Creighton, and S. Nahavandi
An entropy fusion method for feature extraction of EEG,

S. Chen, Z. Luo, and H. Gan
Brain-computer interfacing using modulations of alpha activity induced by covert shifts of attention,

M. S. Treder, A. Bahramisharif, N. M. Schmidt, M. A. J. van Gerven, and B. Blankertz
Decoding spatial attention by using cortical currents estimated from electroencephalography with near-infrared spectroscopy prior information,

H. Morioka, A. Kanemura, S. Morimoto et al.
Independent resources for attentional tracking in the left and right visual hemifields,

G. A. Alvarez and P. Cavanagh

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Classification of Targets and Distractors Present in Visual Hemifields Using Time-Frequency Domain EEG Features

Classification of Targets and Distractors Present in Visual Hemifields Using Time-Frequency Domain EEG Features

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Classification of Targets and Distractors Present in Visual Hemifields Using Time-Frequency Domain EEG Features

Classification of Targets and Distractors Present in Visual Hemifields Using Time-Frequency Domain EEG Features

References

Abstract

Journal

Recommended Articles

References

Our policy towards the use of cookies