PropertyValue
is nif:broaderContext of
nif:broaderContext
is schema:hasPart of
schema:isPartOf
nif:isString
  • Ten volunteers participated in the study. The participants were aged between 21 and 34 (mean 26.8 ±3.7, 7 female). All were recruited from the local student and staff population and either paid for their expenditure of time or granted course credit. All participants had no known prior or current pathological neurological condition (based on self report) and normal or corrected-to-normal vision. The experimental procedure and written consent form for this study were approved by the ethics committee at Bielefeld University, and adhered to the ethical standards of the sixth revision of the Declaration of Helsinki. All participants gave their informed written consent to participate in the study. We used the EyeFollower (LC Technology, Clearwater, USA) remote binocular eye-tracking system for the proposed study. It allows for head movements (76 x 51 x 40cm) without the need to wear a headset. The EyeFollower has a sampling rate of 120 Hz and an accuracy of <0.4°over the whole range of head movements (see Fig 1, left). We used the g.USBamp 16-channel EEG amplifier (Guger Techologies, Graz, Austria) for the study. Twelve EEG channels were recorded at the locations Fz, F3, F4, Cz, C3, C4, Pz, P3, P4, PO7, PO8 and Oz, referenced to the mastoids. Impedances were kept below 5 kΩ. Additionally, we recorded the vertical and horizontal electrooculogram (VEOG and HEOG) to register eye movements together with the EEG channels. Two VEOG electrodes were placed above and below the right eye and two HEOG electrodes beside the left and the right eye, respectively. These electrodes register the corneo-retinal standing potential that exists between the front and the back of the human eye. This data allows to investigate the influence of eye movements on the recorded EEG data. The EOG comprises four additional channels that are recorded alongside the EEG channels leading to a total of sixteen channels that are registered by the EEG amplifier. The images containing every-day objects were taken from the Amsterdam Library of Object Images (ALOI, http://aloi.science.uva.nl). An overall of 112 objects was selected, which equals the number of trials. The originally black background was converted to white. For each stimulus, twelve images were arranged in a grid-like fashion (plus random offset, see Fig 1 (right). One out of the 112 objects was selected as the target for a trial such that each object served only once as target. The remaining eleven objects per stimulus were chosen randomly from the set. The participants were instructed to search for the target object and to press a key on the keyboard when they found it. Each trial started with the presentation of the target object in the center of the screen. The participants had to press a key to start the search task. We implemented a gaze-contingent approach, i.e., not the whole stimulus image was visible but only a circle area, a “keyhole”, centered at the current gaze coordinates on the screen as delivered by the eye tracker with a size of 3.5°of visual angle. The remaining part of the screen was black. The size of the visible area was chosen to cover the foveal (ca. 2°visual angle) plus half of the parafoveal area (ca. 3°). The latter corresponds to the part of the parafoveal area that still provides rather sharp vision (given the decreasing degree of sharpness toward the edges of the parafoveal area). All algorithms for analysis, feature extraction and classification were implemented in MATLAB (Release 2015a). Fixation detection on the raw eye tracking data was done using an threshold-based algorithm proposed by the manufacturer of the eye tracking device (LC Technology International Inc., Clearwater, USA. http://www.eyegaze.com/). The threshold is based on a spatial criterion, that is, consecutive samples are considered to be part of a fixation when they fall inside a circle with a diameter of 2°visual angle. Saccades were not analyzed in this study. The resulting fixations were grouped according to their locations on the stimulus image: on a target object, on a non-target object or on the blank background. The fixation durations in the three groups did not significantly differ: Fixations on target objects had a duration of 270 ms (±38 ms) on average, non-target fixations 281 ms (±37 ms) and background fixations 241 ms (±29 ms). The EEG data was segmented according to fixation onsets and grouped according to their locations. A fixation was considered to be an onto-object fixation, when it fell inside a circular area of 2.5°visual angle (ca. 100 pixel) anchored at the object center. Fig 1 (right) shows an example scanpath for one trial, that is, the sequence of the participant’s fixations and saccades while he/she searches for the target object. Removing eye artifacts with ICA: Independent component analysis (ICA) is a blind source separation method to split a multivariate signal into linearly independent sources [20]. These sources must be linearly mixed in the recorded signal. With respect to EEG data, the sources computed by the ICA can be directly interpreted as cortical sources of neural activity that provide a unique contribution to the overall signal that is picked up. This is of particular interest because of the poor spatial distribution of the EEG method, caused by the spatial low-pass filtering effect of volume conduction. For eye movement related signal portions this effect is particularly severe because of their very high amplitudes compared to genuine cortical signals. ICA is capable of identifying eye movement related signal portions—eye artifacts—as single sources. Hence, these deteriorating artifacts can be removed by ICA. A drawback of the method is that it usually requires a human expert to identify the affected components by visual inspection. The selection is based on a component’s spatial distribution (scalp maps) and its spectral properties. In the present offline study, we used this manual selection technique to identify to-be-removed components. The ICA was computed using the EEGLAB toolbox [29], which implements an Infomax approach. A manual selection of components is of course not possible during online operation of a system. However, the computation of the ICA projection matrix, the inspection process and the selection could be done on training data, which is needed for the classifier anyway. The stored ICA matrix with the affected components removed can subsequently be used during the online run to remove eye artifacts in an automated fashion. Moreover, Winkler and colleagues [30] have recently developed an algorithm for an automated classification of artifactual ICA components. First, the continuous, multi-channel EEG data was segmented according to the temporal onsets of the fixations. Each epoch started at the the first sample of a fixation and lasted for 800 ms, which equals 205 samples per channel. The channel-wise epochs were first bandpass filtered with cut-off frequencies of 0.5 to 10 Hz. Afterwards, the channel-wise epochs were concatenated to form one high-dimensional vector xi∈R2460. Then, we applied Principal Component Analysis (PCA) to reduce the dimensionality [31]. PCA seeks to find a new set of basis vectors that represent the directions of the largest variances in the high-dimensional input space. This is achieved by determining the eigenvectors of the data covariance matrix. The corresponding eigenvalues quantify the amount of variance that is represented by the respective eigenvector. The dimensionality reduction is achieved by projecting the data into a feature space spanned by those k eigenvectors that represent a certain amount of variance. We used a criterion of 99.9% of variance to determine the dimensionality of the feature space. The latter is calculated by ∑i=1kλi∑j=1nλj with k << n given the eigenvalues λ are sorted in descending order. Typically, this resulted in a feature space with a dimensionality around 240. Classification was achieved using Fisher’s Linear Discriminant Analysis (FDA) [32]. The FDA is mostly used as a binary classifier separating two classes. However, in general the FDA can be applied to n classes using the same optimization criteria. Because the FDA is a supervised method, we need to group the feature vectors according to the class labels first resulting in n groups of vectors {xj}Ci with i being the respective class and j = 1, 2, …, NCi where NC−i is the number of vectors in the respective class. Then, we define the between-class scatter matrix as Sb=∑CjNcμCj-μμCj-μT,(1) with μ=1n∑CjμCj. The within-class scatter matrix is given as Sw=∑i∈Cjxi-μCjxi-μCjT In practice, an adequate estimation of the scatter or covariance matrices is often difficult. This difficulty results from a disproportion of available training samples to feature dimensions. As training data is often costly, many applications suffer from too small training set sizes. This is particularly true for EEG-based BMI systems, where in addition the feature vectors extracted from the raw data are usually rather high-dimensional. To tackle this issue, Ledoit and Wolf [33] have developed a method known as the Ledoit-Wolf theorem which helps to estimate well-conditioned covariance matrices despite the described disproportion. The authors in [34] give a more practical formulation of the theorem. We have followed this approach for all covariance matrices computed during our analysis and used a shrinkage estimate instead of the pure sample covariance matrix: Σ*=λΣ˜+1-λΣ,(2) with Σ being the sample covariance matrix, Σ˜ the sample covariance of a sub-model, and λ ∈ [0, 1] denotes the shrinkage intensity. As proposed in Schaefer and Strimmer [34], we use a diagonal matrix where all diagonal elements are equal, that is, all variances are equal and all covariances are zero. Using Eqs (1) and (2), the FDA projection matrix can now be defined as the following optimization problem: Pfda=argmaxP∈Rd×(n-1)trPTSbPPTSwP. (3) Note that this formulation is equivalent to the Rayleigh coefficient. The solution can be expressed as a generalized eigenvalue problem of the form SbV=λSwV. (4) Obviously, in case of a binary classification task, Pfda reduces to a weight vector w and all projections are simple scalar products yi=x·wT+b and the class labels y^∈{-1,1} can be simply determined by the sign function. The performance of the classifier was assessed using the area under curve (AUC) measure. The AUC is defined as the area under the so called “Receiver Operator Characteristics” (ROC) [35]. The ROC is obtained by mapping all real-valued classifier outputs yi = x ⋅ w onto the interval [0, 1] and testing y^i=yi+bk for b1 = 0 and bK = 1 with k = 100 and the bk monotonically increasing and equally spaced. Then, the number of true positives is expressed as a function of the number of false positives. The AUC measure is chosen, because it is not effected by the prior class probability as is the simple measure of accuracy, which simply gives the ratio of correctly classified samples and is only meaningful when both classes are perfectly balanced in the full testset. For the AUC, the chance level is always AUC = 0.5. Please refer to Fig 2 (bottom left) as an illustrative example of the ROC of dataset 8 in the intra-subject classification scheme. A dataset corresponds to the data of one participant recorded in one continuous session. Figure data removed from full text. Figure identifier and caption: 10.1371/journal.pone.0146848.g002 Top row: Verification of the linear separability for the best (left-) and poorest (right) datasets by inspecting the projections of the feature vectors on a 2d space spanned by the 2 largest eigenvectors of a 3-class FDA. The results illustrate that the 3-class FDA separates the data samples into three clearly distinguishable classes for the best dataset depending if the fixation appeared on targets, non-targets and background. Although the overlapping areas are larger in case of the poorest dataset, the three distinguishable classes are clearly evident. Bottom left: Comparison of the ROCs of participant 8 in the intra-subject classification scheme using only EEG channels and only EOG channels. The ROC values are averaged over the ten cross-validations runs. Bottom right: Classification results for the 3-class FDA. The dotted black line indicates the chance level.
rdf:type