1. Introduction
Attention is the direction and concentration of psychological activities on an object, and is a psychological feature that accompanies mental processes, such as memory, thinking and imagination. The applications of attention-level recognition serve our lives in healthcare [
1,
2], safe driving [
3,
4], and education [
5]. Posner divides attention into endogenous attention and exogenous attention [
6]. Endogenous attention, also known as active attention, refers to the individual’s allocation of attention according to their goals or intentions to dominate behavior; exogenous attention, also knowns as passive attention, refers to the individual’s attention caused by external information, usually from unexpected stimuli. In the experimental paradigm designed by Posner and his colleagues, endogenous attention is aroused by presenting target cues in the fixation area, while exogenous attention is aroused by cues emerging in the vicinity of the target [
7]. In this study, we focus on positive attention due to its importance in our daily life. We give the subjects clear goal intention through specific tasks, and induce the subjects’ attention state to varying degrees by controlling the difficulty of task execution.
In previous studies, attention levels can be identified by external representations, such as eye state and facial expression [
8,
9]. However, relying on external representations to identify attention may not be reliable [
10]. With the development of cognitive psychology, researchers have found that the cerebral cortex is the most advanced area for generating attention. Attention reliably modulates neural activity in primary and secondary cortices, affecting the mean neuronal firing rate as well as its variability and correlation across neurons [
11,
12]. Therefore, attention-level recognition based on electroencephalogram (EEG) signals is gradually emerging.
EEG is a physiological signal produced by nerve activity of the brain, which can be obtained by placing electrodes on the surface of the human scalp. The neural activity of the brain changes with people’s mental state, emotion and cognitive activity [
13,
14]. Troy et al. recorded EEG signals during reading, painting, and other cognitive tasks in eight children with attention deficit disorder and eight normal children. The results showed significant θ band (4–8 Hz) amplitude difference between groups [
15]. Ryota Kobayashi et al. collected EEG data from 61 healthy college students at rest with their eyes closed and concluded that individuals had higher attentional control at lower θ/β levels [
16]. Li et al. acquired the EEG signals under three conditions of attention task, inattentive task and rest task through “tennis test” and “walking test”, and classified them by approximate entropy, sample entropy and multi-scale entropy features. The highest accuracy of 85.24% was obtained using sample entropy [
17]. F Fahimi et al. developed an end-to-end depth Convolutional Neural Network (CNN) to decode attention information from EEG time series. Three different EEG representations were fed into the network, and the final average classification accuracy was 79.26% [
18]. Hu et al. obtained three types of attention data through self-evaluation after online learning and extracted 25 features from 6 EEG channels, respectively. Using the correlation-based feature-selection (CFS) method and the K-nearest neighbor (KNN) classifier, an accuracy of 80.84% was reported in distinguishing the three attention states [
19].
There are relatively few studies on multi-level attention classification based on EEG, and feature screening is rarely considered. An effective feature-selection algorithm can provide insights into the data, improve model generalization performance, as well as identify irrelevant features [
20]. In this study, we choose a limited number of ten channels and filter the number of features, which reduce the computational complexity for the implementation of a miniaturized and intelligent detection device [
21,
22]. At the same time, multi-level attention recognition has more application prospects than single-level attention recognition. The four-level classification of attention avoids simple judgments and provides a transition interval for classifying attention states. This is practical, for example, in the process of detecting the driver’s attention and providing feedback when a downward trend in attention is detected (from high to medium), rather than until a low attention span occurs [
23].
In this paper, four different experimental situations are designed to enable the subjects to achieve four states of high, medium, low and non-externally directed attention. A total of 10 features are extracted from EEG signals, including time-domain measurements, sample entropy, and frequency band energy ratios. Based on these features, an average recognition accuracy of 88.7% is achieved in classifying the four attention states using a support vector machine (SVM) classifier [
24]. To further improve the classification performance and reduce the dimension of the feature space, feature selection is performed to identify the most informative features from the original feature set. In this work, we use the sequential-forward-selection (SFS) method [
25] to generate the candidate feature subsets. Based on the optimal feature set, an improved classification accuracy of 94.1% was achieved, which demonstrates the effectiveness of the proposed feature selection and classification scheme in multi-level attention recognition.
The rest of this paper is organized as follows. In
Section 2, we provide a detailed description of the experimental design.
Section 3 explains the data processing procedures, including EEG preprocessing, data segmentation, and feature extraction. In
Section 3, we show the classification results with different feature-selection methods.
Section 4 and
Section 5 are the discussions and conclusions, respectively.
2. Experimental Design
2.1. Channel Selection
EEG can be divided into five rhythms (frequency bands) of δ, θ, α, β and γ in the frequency domain, and different rhythms have different characteristics [
26]. Among them, the frequency bands related to attention mainly include θ, α, and β waves, which have the following characteristics:
θ wave, with frequency ranging from 4 Hz to 8 Hz and amplitude ranging from 20 uV to 40 uV, usually occurs when people are relaxed or tired, and is mainly distributed in the central area of the brain. θ wave in awake state is related to attention alertness.
A wave, with frequency ranging from 8 Hz to 13 Hz and amplitude ranging from 10 uV to 80 uV, usually appears when people are calm, and is mainly distributed in the occipital and the parietal lobes.
B wave, with frequency ranging from 13 Hz to 30 Hz and amplitude ranging from 3 uV to 50 uV, usually appears when people are excited, and is mainly distributed in the frontal and the central areas.
According to previous research, compared with the non-attention state, EEG signals in the attention state have more β waves, but less θ waves and α waves [
27]. Therefore, when selecting channels, we choose the ones in the frontal lobe, the central area (β wave), the occipital lobe, the parietal lobe (α wave) and the central area (θ wave). The final selected channels are Fp_1, Fp_2, F_3, F_4, C_3, C_4, P_3, P_4, O_1 and O_2.
2.2. Data Collection
14 subjects aged between 20 and 24 years old participated in this study, including 6 female students and 8 male students. All subjects were undergraduates or postgraduates, with right handedness and normal or corrected vision. Sufficient sleep was guaranteed before the experiment. The experiment was conducted in the Laboratory of Geography and Biology at Nanjing University of Posts and Telecommunications.
The laboratory has sufficient light and suitable temperature, which can make the subjects in a relaxed and comfortable atmosphere. The sound insulation effect is good, preventing uncontrollable factors outside the laboratory from interfering with the data acquisition process. During the experiment, electronic devices such as mobile phones are turned off to avoid electromagnetic interference generated by devices in the environment. The device used in the experiment is a multi-channel wet electrode EEG acquisition instrument produced by Nanjing Weisi Medical Institution. The experimental instrument can complete multi-channel EEG signal acquisition, amplification, sampling, filtering, etc. The electrode distribution conforms to the international 10-20 system standard electrode placement method.
2.3. Experimental Scheme
In contrast to the conventional two-level attention experiment, this experiment induced four different levels of attention states by controlling the difficulty of the tasks. The original data which met the experimental requirements based on self-evaluation scale were kept for further analysis. The four types of attention tasks are shown in
Table 1.
2.4. Experimental Process
Before performing Task 1, the subjects were asked to do a set of numerical exercises. The experimenter named a number within 100. If the number was a prime number, the subjects were asked to answer “Yes”. If the number was not a prime number, the subjects were asked to say a factor of the number. For example, if the experimenter said “35”, the subjects could answer “5” or “7”. If the experimenter said “17”, the subjects should answer “Yes”. The purpose of setting the number exercise is to awaken the subjects’ sensitivity to numbers before starting the formal experiment.
The four types of tasks were carried out sequentially. The task time of Task 1 started when the subjects browsed the first number and ended when the last number was judged. The task time of Task 2 to Task 4 is the same as that of Task 1. After each task, the subjects rested for 30 s. During rest, they were asked to fill out corresponding questionnaires to self-evaluate their attention state during the experiment [
28], so that we could screen the samples with subjective evaluations.
To ensure that the subjects were successfully induced to an appropriate attention state during the experiment, a subjective questionnaire was designed. At the end of each task, subjects were asked to fill out a corresponding questionnaire to assess the state they had just experienced during the experiment. The questionnaire for each task consisted of 3 questions, each with 5 options A–E. Each option received an increasing score from A, with 1 point for A and 5 points for E.
Two attention scales are shown in
Table 2 and
Table 3, respectively. For
Table 2, a total score greater than 12 is considered to meet the expectations of Task 1, and 9–12 is considered to meet the expectations of Task 2. For
Table 3, a total score of more than 12 is considered to meet the expectations of Task 3, and 9–12 is considered to meet the expectations of Task 4.
The purpose of the first question in
Table 2 is to make sure that the subjects do not reject text tasks and digital tasks. During the experiment, all subjects were able to complete both the text tasks and the digital tasks (Score 5). The flow chart of each group of experiments is shown in
Figure 1, where t1 is the time taken by the subjects to complete Task 1. At the end of the experiment, the experimental content was compared with the subjective scale scores, and the EEG signals that matched the purpose of the experiment were reserved for analysis, while those failed to match were invalidated. Each subject repeated the experiment in two groups with a 30-min interval between the two groups.
5. Discussion
In this study, we classified the subjects’ attention state into four layers. After collecting the corresponding EEG signals, we extracted ten features including six time-domain features, three frequency domain features, and a nonlinear feature. These features are screened and optimized, and finally, the five best feature combinations are selected. Based on the optimal feature set, the classification accuracy is improved significantly compared with using the original feature set without feature selection.
Experimental results show that feature F9 (E_α/E_all) was not included in the optimal feature set. In
Figure 10a, we compare the F9 feature of Task 2 to Task 3 of the same subject. Significantly lower α power of medium attention task was observed compared with that of the low attention task, which is consistent with the previous findings [
44,
45,
46]. However, for tasks 3 and 4, the α-wave energy ratio is indistinguishable, as shown in
Figure 10b. This may have led to the exclusion of the α wave energy ratio from the optimal feature set.
The feature filtering algorithm in this paper can improve the classification performance while reducing the dimension of the feature set. The reduction of computation can also serve real-time EEG attention and fatigue detection system, which is promising [
21,
22]. In the current study, all 10 channels are included for feature extraction. Some studies have proposed a channel-based feature-selection method that takes into account the performance of a single-channel model and its physical location for studying groups of channels related to attention detection. This can be combined with feature screening to better improve classification performance [
47].
In the feature-selection process, we tried three methods, and the results show that the combination of features obtained using the wrapper method achieves the highest accuracy. However, this does not prove that the wrapper method is superior to the other feature-selection methods. This is because the wrapper method enumerates all the different feature subsets and chooses the one that makes the model work best. This is suitable for this study when the number of features is small, but for applications with very large number of features, the time complexity of wrapper methods may be too high.
This study also has several limitations: (1) A time window of four seconds was chosen for data processing. Time windows of different lengths may have an impact on classification accuracy. (2) The number of subjects recruited and the range of their age was limited. Future studies should cover different age groups and expand the sample size. (3) The number of EEG channels could also be further screened to reduce the complexity of data processing [
48].
In addition, the subjects in this experiment were all university students in normal physical and mental conditions. There are several diseases that can affect people’s attention levels, such as attention deficit hyperactivity disorder. Future studies will also be directed towards comparing people with attention deficits with healthy subjects [
49].
6. Conclusions
In this paper, we designed four experimental scenarios to induce different levels of attention. Combined with the self-assessment questionnaire, the EEG signals of four states of high attention, middle attention, low attention and non-externally directed attention were collected.
After pretreatment, a total of 3403 samples from 14 subjects were obtained. Ten features are extracted from each of the ten EEG channels, which results in a 100-dimensional feature vector to classify the four categories of EEG signals. An 88.7% classification was achieved using a support vector machine classifier.
To identify the optimal subset of discriminating features from the original feature set, the sequence-forward-selection method is employed. After feature selection, sample entropy, standard deviation, root mean square, rectified mean value and margin factor are retained, based on which the classification accuracy was improved to 94.1%.
At the subject level, when using the first group of data for training and the second group of data for testing, the average classification accuracy was improved by 1.97% after feature selection. These promising results indicate the effectiveness of feature selection in attention-level recognition.