EEG Feature Analysis Related to Situation Awareness Assessment and Discrimination

Abstract

In order to discriminate situation awareness (SA) levels on the basis of SA-sensitive electroencephalography (EEG) features, the high-SA (HSA) group and low-SA (LSA) groups, which are representative of two SA levels, were classified according to the situation awareness global assessment technology (SAGAT) scores measured in the multi-attribute task battery (MATB) II tasks. Furthermore, three types of EEG features, namely, absolute power, relative power, and slow-wave/fast-wave (SW/FW), were explored using spectral analysis. In addition, repeated analysis of variance (ANOVA) was conducted in three brain regions (frontal, central, and parietal) × three brain lateralities (left, middle, and right) × two SA groups (LSA and HSA) to explore SA-sensitive EEG features. The statistical results indicate a significant difference between the two SA groups according to SAGAT scores; moreover, no significant difference was found for the absolute power of four waves (delta (δ), theta (θ), alpha (α), and beta (β)). In addition, the LSA group had a significantly lower β relative power than the HSA group in central and partial regions. Lastly, compared with the HSA group, the LSA group had higher θ/β and (θ + α)/(α + β) in all analyzed brain regions, higher α/β in the parietal region, and higher (θ + α)/β in all analyzed regions except for the left and right laterality in the frontal region. The above SA-sensitive EEG features were fed into principal component analysis (PCA) and the Bayes method to discriminate different SA groups, and the accuracies were 83.3% for the original validation and 70.8% for the cross-validation. The results provide a basis for real-time assessment and discrimination of SA by investigating EEG features, thus contributing to monitoring SA decrement that might lead to threats to flight safety.

1. Introduction

Human error was shown to contribute to more than 70% of safety accidents and incidents in safety-critical industrial areas, while 88% of human error was found to be due to problems with situation awareness (SA) [1,2]. The broadest definition of SA is “the perception of elements in the environment within a volume of time and space, the comprehension of their meaning, and the projection of their status in the near future” [3]. In consideration of its importance in human factor/ergonomics areas, the related work to measure and discriminate SA continuously and effectively is still insufficient [3,4].

According to the related literature, SA measurement can be mainly categorized into subjective measures, freezing or real-time probe measures, performance measures, and physiological measures [5,6]. Each of these measures has its advantages and disadvantages. For example, the three-dimensional situation awareness rating technology (3D-SART) is a typical subjective method to measure SA [7]. The primary advantages of subjective measures are that they are inexpensive, are easy to administer, and have high face validity [8]. However, the sensitivity and subjectivity of this method have been criticized considering the post-trial collection of the operator’s scoring. As for freezing or real-time probe techniques, the situational awareness global assessment technology (SAGAT) method and the situation present assessment method (SPAM) are two excellent representatives [9,10]. SAGAT is a freezing probe SA measurement which has received widespread acceptance with large amounts of empirical validations [11,12,13,14]. However, the SAGAT method is not conducted in real time and introduces interference to the primary task (requires the testing scenario to be frozen randomly), which makes it hard to be accepted in a real environment. SPAM is a real-time probe measure with the advantage of in field application and intrusion decrements imposed by task freezes in the freezing probe techniques [5]. However, the limitations of the intrusion to the primary task and its validity have been debated [15]. Another method is represented by performance measures. Although it is continuous, objective, and nonintrusive, its serious drawback is its unclear relationships with SA, whereby efficient performance does not necessarily lead to an operator having good SA during the task, and vice versa [5].

Physiological measures, which use sensors to collect human biological signals, have the advantages of being continuous, objective, mobile, and real-time [16,17]. Compared with the previous measures, physiological measures show great potential considering the above benefits. Electrocardiographic (ECG), electrodermal (EDA), respiration, and eye movement signals have been explored in the physiological measurement of SA. Sun et al. (2017) comprehensively used EDA, ECG, and respiration in a flight simulation environment, and the result revealed the correlations between physiological indicators (standard deviation of normal to normal (SDNN), respiratory rate, and coefficient variation of RR interval (RRCV)) and SA. Van de Merwe et al. (2012) demonstrated the potential of eye movement indicators (fixation rates, dwell times, and entropy) to assess SA in a flight simulator experiment. Similarly, de Winter et al. (2019) outlined the limitations of the SAGAT and indicated the potential of using eye-tracking metrics in eye-based SA [11]. Among these physiological measures, eye-tracking technology was the most common modality adopted in the identified SA studies, and it was often correlated with direct SA measures [11].

Compared with the above physiological methods, electroencephalography (EEG) possesses high time resolution (accurately records quick changes in electrical activity on a millisecond timescale), direct measurement of operator’s neural activity at group level, and correlations with SA-related cognitive constructs, e.g., perception, comprehension, and working memory, which makes it suitable for use in SA studies [18,19,20,21,22]. Moreover, mental states that are closely related to SA can also be reflected in EEG, such as mental workload [23,24], fatigue [25], and alertness [26]. Although a large number of EEG measures have been reported for the above mental states, few published SA-related studies have explored EEG measures. For example, Catherwood et al. (2014), found that the early coactivity in high-order and visual perception regions may provide a basis for top-down influence on perception during early intervals in the loss of SA in target and threat identification tasks [18]. Yeo et al. (2017) found that EEG is a promising approach for measuring the SA of air traffic controllers by investigating machine learning methods combined with EEG indicators (power, power ratio, and entropy) [19]. Kaur et al. (2020) found that the pre-task resting alpha frontal asymmetry, pre-task resting absolute alpha, and associated neural mechanisms significantly influence the outcome of SA tasks via the Stroop effect [27]. Kästle et al. (2021) identified the EEG signatures associated with SA in various regions of the brain under the operator’s high and low SA [28].

Another approach is to discriminate different SA levels. For instance, Li et al. (2020) discriminated two fatigue-related SA levels (high and low SA), yielding various accuracies of the spectral–spatial (S–S) model (70.6%), logistic regression (67.9%), support vector machine (63.8%), EEGNet-4.2 (59.0%), and EEGNet-8.2 (60.3%) [29]. However, the 2020 model of Li et al. uses the reaction time to quantitatively define the SA, leading to concerns regarding its validity because of the unclear relationship between performance measures and SA. Kästle et al. (2021) used random forest and boosted trees combined with all EEG data to discriminate SA levels, achieving 67% accuracy [28]. In the 2021study of Kästle et al., the power spectral density (PSD) of five EEG bands was used to perform the discrimination, lacking connection with the results of SA-sensitive EEG signals. Despite the above efforts, several limitations can be highlighted. First, the EEG features used for discrimination were limited, and most accuracies were less than 70%. Second, traditional discrimination methods, such as regression analysis [30] and Bayes method [31], which have been widely used for the discrimination of several typical human states, were rarely used in the discrimination of SA. Considering their good interpretability, they could benefit from clarification of the underlying mechanisms of SA. Third, most of the current studies used all EEG features rather than SA-sensitive EEG features, especially those in the frequency domain, which probably introduced interruption and led to nonoptimal discriminate results.

On the basis of the above limitations, an exploration of EEG experiments and feature extraction for different SA groups was conducted in this manuscript. At first, the high-SA (HSA) group and low-SA (LSA) group, which are representative of two SA levels, were classified according to SAGAT scores measured in the multi-attribute task battery (MATB) II tasks. Moreover, three types of EEG features, namely, absolute power, relative power, and slow-wave/fast-wave (SW/FW), were explored using spectral analysis. In addition, repeated analysis of variance (ANOVA) was conducted in three brain regions (frontal, central, and parietal) × three brain lateralities (left, middle, and right) × two SA groups (LSA and HSA) to explore SA-sensitive EEG features. Lastly, the above SA-sensitive EEG features were fed into principal component analysis (PCA) and the Bayes method to discriminate between the two SA groups. Correspondingly, five hypotheses were also proposed in the current study:

Hypothesis 1 (H1).

Groups with low and high SA levels according to SAGAT scores can be distinguished when operating the MATB II tasks.

Hypothesis 2 (H2).

There are absolute power differences between HSA and LSA groups.

Hypothesis 3 (H3).

There are relative power differences between HSA and LSA groups.

Hypothesis 4 (H4).

There are SW/FW differences between HSA and LSA groups.

Hypothesis 5 (H5).

The combination of SA-sensitive EEG features and a linear discrimination method can achieve effective SA classification results.

2. Methods

2.1. Experimental Design

This research aimed to investigate the sensitive EEG features for SA assessment and discrimination. In our experimental design, the SA level was the independent variable, while the measured EEG features were the dependent variables. The SA levels were achieved by high- and low-SA groups, which were distinguished according to SAGAT scores in the MATB II tasks. The EEG features of absolute power, relative power, and SW/FW were analyzed, as shown in Figure 1.

2.2. Subjects

Forty-eight graduate students with an average age of 23.15 years (standard deviation (SD) = 0.72, ranging from 22 to 24 years) were selected as subjects (39 males and 9 females) from the School of Aeronautic Science and Engineering, Beihang University. All participants were informed of the instructions and procedures of the experiment, and they signed written informed consent forms before the experiment. All subjects were in good health, with right-hand dominance, normal vision or corrected-to-normal vision, and adequate sleep (at least 8 h) during the 24 h before the experiment. This research was approved by the Beihang University Biological and Medical Ethics Committee.

2.3. Experimental Platform and Tasks

MATB II developed by the National Aeronautics and Space Administration (NASA) is widely used when building simulation scenarios in a laboratory environment. MATB II was originally designed as a classic computer-based platform to evaluate operator performance and workload [32]. After several updates, it was validated in numerous studies and extended to several related areas [33,34,35]. In consideration of its acceptance in testing the human cognitive state, the MATB II platform was used in this study for evaluating SA. MATB II includes four typical subtasks: system monitoring, target tracking, scheduling display, and resource management, which originated from typical flight task characteristics [32]. The flying joystick, keyboard, and mouse can be used to interact with the platform.

2.4. EEG Equipment and Indicators

In the current study, the Neuroscan Neuamps amplifier was used to record the EEG data, including 32 channels: F7, FT7, T3, TP7, T5, A1, FP1, F3, FC3, C3, CP3, P3, O1, FZ, FCZ, CZ, CPZ, PZ, OZ, FP2, F4, FC4, C4, CP4, P4, O2, F8, FT8, T4, TP8, T6, and A2. All electrodes were Ag/AgCl. The recorded wave was 0.1–200 Hz, and the sampling rate was 1000 Hz. The impedance of the electrodes was maintained under 5 kΩ. GND served as the ground, and the left mastoid A1 was referenced online. The vertical and horizontal channels of the electrooculogram (EOG) were recorded simultaneously. In addition, the EEG signal was re-referenced to the average of the bilateral mastoid electrodes offline. The analysis was performed using the toolbox MATLAB R2014a eeglab14_1_1b (available from https://sccn.ucsd.edu/eeglab/download/daily/, accessed on 16 September 2022), and the bandpass filtering of 1–30 Hz was adopted. The EEG signals were analyzed with independent component analysis (ICA), and the Adjust 1.1.1 toolbox (available from https://www.nitrc.org/frs/?group_id=739&release_id=3054#, accessed on 16 September 2022) was used to eliminate the artifacts. All EEG data were segmented, and fast Fourier transform (FFT) was performed. Then, the segmented data were averaged and divided into four waves: 1–4 Hz for the delta (δ) wave, 4–8 Hz for the theta (θ) wave, 8–12 Hz for the alpha (α) wave, and 13–30 Hz for the beta (β) wave. The overall band power was determined from the integration of the power across the designated spectrum. The unit of power analysis was microvolts squared per octave (μV²). The EEG signals were collected in real time using the EEG acquisition device throughout the experiment.

2.5. EEG Equipment and Indicators

Before the formal experiment, all subjects received enough training and practice to make sure that they were skilled in completing the designed subtasks (with an accuracy higher than 90%). Meanwhile, each subject was required to wear EEG sensors and complete the resting test. After the training session, the SAGAT measures were performed to test the SA levels of subjects in MATB II tasks [28]. According to previous studies, SAGAT is a highly sensitive, reliable, and predictive measure of SA shown to be useful across a wide variety of domains and experimental settings [13]. Therefore, the SAGAT method based on the computerized freezing technique was used in our study to measure SA. Specifically, the MATB II interface was frozen and completely covered by the presented question interface at a random time during the experiment. The presentation interval of freezing was about 5 min, and six questions were raised each time. The average presentation time of questions was less than 1 min, and the interval of each freezing was more than 2 min [10]. It was guaranteed that no question was asked within the first 3 min so that the subject could obtain information in a stable scenario [10]. The popup question selected from the SAGAT database was answered by the subject using mouse clicks.

3. Results and Analysis

3.1. Grouping of SA Levels

In our study, two SA groups (i.e., HSA group and LSA group) were classified according to SAGAT scores. Referring to previous studies, subjects with the top 25% SAGAT scores (12 subjects) were categorized as the HSA group, while subjects with the bottom 25% SAGAT scores (12 subjects) were categorized as the LSA group [28,36]. The mean age of the HSA group was 23.08 years (SD = 0.79), while that of the LSA group was 23.17 years (SD = 0.58). An independent t-test was conducted using SPSS 25.0 on SAGAT scores of different SA groups. It was found that SAGAT scores showed an increasing trend from the LSA group (70.83 ± 5.86) to the HSA group (91.20 ± 2.86). Statistical results indicated that the SAGAT score of the LSA group was significantly lower than that of the HSA group (t(22) = −10.817, p < 0.001). The above results demonstrate that the grouping effect of SA among the two groups was validated.

3.2. Quantitative EEG Analysis of Different SA Groups

3.2.1. Data Analysis Methods

Electrode placements in three brain regions (frontal: F3, Fz, F4; central: C3, Cz, C4; parietal: P3, Pz, P4), which have close connections with the construct of SA, were considered for analysis in our study [37,38,39]. Moreover, the differences among left, right, and central brain literalities have often been investigated [24,39]; hence, three lateralities (left: F3, C3, P3; central: Fz, Cz, Pz; right: F4, C4, P4) were also studied. On the basis of the SA grouping, the sensitive EEG features were further explored for SA. One of the most standard methods of EEG analysis is frequency-domain analysis [21]. Considering its representativeness in the distribution of full-head electrode sites and widespread use in EEG research, nine electrode sites (F3, Fz, F4, C3, Cz, C4, P3, Pz, and P4) including three brain regions (frontal, central, and parietal) and three lateralities (left, middle, and right) were selected for the analysis of EEG spectral features [23,24,25]. The absolute power (μV²) and relative power (%) of the four bands, as well as four SW/FW ratios, were selected as the EEG spectral features. The relative power of a certain wave is the ratio of the absolute power of a certain wave to the absolute power of the whole wave (1–30 Hz). The four typically used SW/FW ratios of θ/β, α/β, (θ + α)/β, and (θ + α)/(α + β) were adopted for analysis in this manuscript [40].

Therefore, three types of EEG spectral features under the LSA group and HSA group were selected for repeated-measures ANOVA of three brain regions (frontal, central, and parietal) × three brain lateralities (left, middle, and right) × 2 SA groups (LSA and HSA). The Bonferroni method was used for post hoc analysis [41]. Mauchly’s method was used for the test of sphericity, and, if the assumption was rejected, the Greenhouse–Geisser method was used for correction [42,43]. The topographical maps in Figure 2 provide a descriptive demonstration of the three types of EEG spectral features. Higher activations in relevant electrode sites are demonstrated by brighter colors (in this case, red) in the figures; otherwise, blue indicates nonactivated electrode sites.

3.2.2. Absolute Power of Different SA Groups

Descriptive results of the absolute power in four waves for the different SA groups are shown in Figure 3. ANOVA results showed that no significant interaction effect (p > 0.05) of the absolute power was found. Moreover, the main effect of the SA group for the absolute power in four waves was also not significant (p > 0.05), as shown in

Table 1. ANOVA results of absolute power for the two SA groups.

Waves (μV2)	Region × Laterality × SA Group			Region × SA Group			Laterality × SA Group			SA Group
	F	p	η2	F	p	η2	F	p	η2	F	p	η2
δ	0.879	0.437	0.038	0.155	0.751	0.007	0.974	0.378	0.042	3.020	0.096	0.121
θ	0.653	0.556	0.029	0.003	0.967	<0.001	1.450	0.247	0.062	3.272	0.084	0.129
α	0.945	0.429	0.041	0.455	0.560	0.020	0.266	0.733	0.012	0.911	0.350	0.040
β	0.453	0.681	0.020	0.639	0.452	0.028	0.392	0.667	0.018	0.490	0.491	0.022

.

3.2.3. Relative Power of Different SA Groups

Descriptive results of the relative power in four waves for the different SA groups are shown in Figure 4. The ANOVA results demonstrated a significant interaction effect of the beta relative power for the region × SA group (F(1.506, 33.130) = 4.467, p = 0.028, η² = 0.169), and further simple effect analysis indicated that the beta relative power under the LSA group was significantly lower than that of the HSA group in the central and parietal brain regions (p = 0.016, p = 0.006). No significant interaction effects and main effects were found in the other waves (p > 0.05), as shown in

Table 2. ANOVA results of relative power for the two SA groups.

Waves (%)	Region × Laterality × SA Group			Region × SA Group			Laterality × SA Group			SA Group
	F	p	η2	F	p	η2	F	p	η2	F	p	η2
δ	1.460	0.234	0.062	0.239	0.694	0.011	1.854	0.171	0.078	0.871	0.361	0.038
θ	0.535	0.652	0.024	0.246	0.673	0.011	0.134	0.859	0.006	1.439	0.243	0.061
α	0.663	0.575	0.029	0.421	0.623	0.019	1.448	0.246	0.062	0.276	0.064	0.012
β	2.204	0.097	0.091	4.467	0.028 *	0.169	1.918	0.160	0.080	6.911	0.015 *	0.239

* p < 0.05.

.

3.2.4. SW/FW Ratios of Different SA Groups

Descriptive results of the four SW/FW ratios for the different SA groups are shown in Figure 5. The statistical results of the four SW/FW ratios for the different SA groups are shown in

Table 3. ANOVA results of SW/FW ratios for the two SA groups.

SW/FW Ratios	Region × Laterality × SA Group			Region × SA Group			Laterality × SA Group			SA Group
	F	p	η2	F	p	η2	F	p	η2	F	p	η2
θ/β	2.228	0.105	0.092	1.185	0.303	0.051	1.572	0.221	0.067	9.610	0.005 *	0.304
α/β	3.404	0.020 *	0.134	5.920	0.007 *	0.212	1.057	0.343	0.046	4.040	0.057	0.155
(θ + α)/β	3.103	0.033 *	0.124	2.778	0.092	0.112	1.770	0.184	0.074	9.072	0.006 *	0.292
(θ + α)/(α + β)	1.180	0.321	0.051	0.299	0.661	0.013	0.708	0.485	0.031	5.761	0.025 *	0.208

* p < 0.05.

. For θ/β, the main effect of the SA group was significant, with the LSA group having a significantly higher θ/β than the HSA group (p = 0.005). For α/β and (θ + α)/β, significant interaction effects were found in the region × laterality × SA group (F(3.175, 69.858) = 3.404, p = 0.020, η² = 0.134; F(2.953, 64.975) = 3.103, p = 0.033, η² = 0.124). Simple effect analysis showed that, in the parietal brain region, the α/β of the LSA group was significantly higher than that of the HSA group in left (p = 0.018), midline (p = 0.013), and right laterality (p = 0.007). For (θ + α)/β, except for the left and right laterality in the frontal region, (θ + α)/β of the LSA group was significantly higher than that of the HSA group in the other electrode sites (p < 0.005), as shown in Figure 5. The main effect of the SA group was significant, with the LSA group having a significantly higher (θ + α)/(α + β) than the HSA group (p = 0.025). No other significant interaction effects and main effects were found (p > 0.05).

3.3. Bayesian Discrimination Model of SA Based on Sensitive EEG Features

According to the above analysis, three aspects of EEG features of the high- and low-SA groups were analyzed: (1) the absolute power showed no significant differences; (2) the beta relative power showed significant differences in the central and parietal regions; (3) the SW/FW ratios of θ/β and (θ + α)/(α + β) showed significant differences in the frontal, central, and parietal regions, the α/β showed significant differences in the parietal brain region, and the (θ + α)/β showed significant differences in the other electrode sites except for the left and right laterality in the frontal region.

Thirty-four SA-sensitive EEG features at relevant electrode sites were used for the classification of high and low SA levels, as shown in

Table 4. Component matrix for SA-sensitive EEG features.

	Component	F1	F2	F3	F1	F2	F3	F1	F2	F3	F1	F2	F3	F1	F2	F3
Electrode Sites		β Relative Power			θ/β			α/β			(θ + α)/β			(θ + α)/(α + β)
F3					0.788	−0.065	0.492							0.762	0.419	0.441
FZ					0.841	0.171	0.449				0.863	−0.074	0.446	0.687	0.558	0.338
F4					0.790	−0.120	0.518							0.721	0.406	0.402
C3		−0.910	0.138	0.074	0.939	0.002	0.083				0.890	−0.315	0.109	0.843	0.485	−0.004
CZ		−0.932	−0.006	0.060	0.938	0.209	0.027				0.964	−0.105	0.094	0.729	0.598	−0.088
C4		−0.899	0.145	0.073	0.970	0.009	−0.001				0.899	−0.357	0.051	0.835	0.484	−0.064
P3		−0.918	0.096	0.164	0.947	−0.030	−0.122	0.629	−0.743	0.049	0.903	−0.363	−0.056	0.859	0.387	−0.214
PZ		−0.881	0.053	0.267	0.916	0.120	−0.338	0.673	−0.692	0.017	0.923	−0.264	−0.210	0.738	0.497	−0.416
P4		−0.916	0.116	0.244	0.883	0.008	−0.303	0.683	−0.690	0.046	0.890	−0.305	−0.180	0.767	0.351	−0.414

. Two steps were taken. Initially, the EEG features were downscaled through a PCA method, and three principal components of F₁, F₂, and F₃ were selected, which explained 91.37% of the total square deviation (72.2%, 12.4%, and 6.7%, respectively), as shown in Figure 6.

PCA is a method of data dimensionality reduction, which characterizes multiple sets of related data as several unrelated principal components [44,45]. If O_i1 is the original dataset, where O_i1 = [o₁, o₂, …, o_i]^T, and F_j1 represents the new principal components, where F_j1 = [f₁, f₂, …, f_j]^T, then the relationship between the two variables can be expressed as

F_j1 = C_ji × O_i1,

(1)

where C is the component matrix, and $C_{ji}=\left[\begin{array}{cccc}{a}_{11}& a_{12}& \cdots & a_{1i}\\ a_{21}& a_{22}& \cdots & a_{2i}\\ \cdots & \cdots & \cdots & \cdots \\ a_{j1}& a_{j2}& \cdots & a_{ji}\end{array}\right]$. The component matrix of F₁–F₃ is shown in Table 4.

Then, F₁, F₂, and F₃ were fed into a classification model based on the Bayesian method.

There are $c$ categories ${\omega }_1,{\omega }_2,\dots ,{\omega }_c$; the prior probability of category ${\omega }_k$ is $P({\omega }_k)$, and ${\displaystyle \sum _{k=1}^{c}P({\omega }_k)}=1$. Given sample $x$, the class-conditional probability density of ${\omega }_k$ on $x$ is $p(x|{\omega }_k)$. Therefore, its posterior probability of belonging to ${\omega }_k$, $P({\omega }_k|x)$, can be described by the Bayes equation.

$$P({\omega }_k|x)=\frac{p(x|{\omega }_k)P({\omega }_k)}{p(x)}=\frac{p(x|{\omega }_k)P({\omega }_k)}{{\displaystyle \sum _{l=1}^{c}p(x|{\omega }_l)P({\omega }_l)}}.$$

(2)

The discriminant function $g_k(x)$ is a description of the classifier, which can be described as $\arg \underset{k}{\max }{g}_k(x)$. For the Bayes classifier, the discriminant function is $g_k(x)=P({\omega }_k|x)$, which can also be simplified to $g_k(x)=p(x|{\omega }_k)P({\omega }_k)$ according to the Bayes equation, or expressed as $g_k(x)=\ln p(x|{\omega }_k)+\ln P({\omega }_k)$ in logarithmic form.

Suppose that the class-conditional probability density of sample $x\in R^d$ follows the multivariate Gaussian density $p(x)\sim N(\mu ,\sum )$, i.e.,

$$p(x)=\frac{1}{{(2\pi )}^{\frac{d}{2}}|\sum {|}^{\frac{1}{2}}}\exp (-\frac{1}{2}{(x-\mu )}^T{\sum }^{-1}(x-\mu )).$$

(3)

Substituting Equation (3) into $g_k(x)=\ln p(x|{\omega }_k)+\ln P({\omega }_k)$, the quadratic discriminant function can be obtained as follows:

$$g_k(x)=-\frac{1}{2}{(x-{\mu }_k)}^T{\sum }_k{}^{-1}(x-{\mu }_k)-\frac{1}{2}\ln \left|{\sum }_k\right|+\ln P({\omega }_k).$$

(4)

If the covariance matrices of all categories are considered equal, ${\sum }_k=\sum$, Equation (4) can be expanded and simplified to obtain the corresponding linear discriminant function.

$$g_k(x)=({\mu }_k{}^T{\sum }^{-1})x-\frac{1}{2}{\mu }_k{}^T{\sum }^{-1}{\mu }_k+\ln P({\omega }_k).$$

(5)

On the basis of the above equations, the linear discrimination equation in our manuscript was built as follows:

Z₁ = −0.961F₁ + 0.167F₂ + 0.419F₃ − 1.032.

(6)

Z₂ = 0.961F₁ − 0.167F₂ − 0.419F₃ + 1.032.

(7)

The values of F₁–F₃ were fed into Equations (6) and (7), and the maximum output of Z₁–Z₂ was set as the grouping result. The classification results were 83.3% accuracy in the original validation and 70.8% accuracy in the cross-validation (leaving 1 out), as shown in

Table 5. Component matrix for SA-sensitive EEG features.

Validation Methods	SA Group	Predicted Accuracy (%)		Average Accuracy (%)
		HSA	LSA
Original validation	HSA	83.3	16.7	83.3
	LSA	16.7	83.3
Cross-validation	HSA	58.3	41.7	70.8
	LSA	16.7	83.3

.

4. Discussion

To investigate the EEG changes related to different SA levels, 48 subjects were recruited to complete MATB II tasks. In our study, subjects with the highest 25% and the lowest 25% SAGAT scores were defined as the HSA group and the LSA group, respectively. The independent-sample t-test showed that SAGAT scores of the LSA group were significantly lower than those of the HSA group, thus confirming H1. In addition, the EEG features related to SA levels were analyzed on the basis of the SA grouping. The experiment demonstrated significant differences in the beta relative power and four SW/FW ratios. Choosing the SA-sensitive EEG features as input, discrimination accuracies of 83.3% in the original validation and 70.8% in the cross-validation were obtained for different SA groups using the PCA and Bayes methods.

PSD represents the signal power distributed over a range of frequencies [20]. δ oscillations are frequently observed in states of diminished consciousness, including slow-wave sleep and disorders of consciousness [46]. θ oscillations are involved in active maintenance and recall of working memory (WM) representations, which increase with added WM load or higher attentional demands [29,47]. α waves occur during relaxed conditions at decreased attention levels, often in a drowsy but wakeful state [23,29,48]. β oscillations are associated with the preservation of the current brain state, typically alertness and arousal [23,49,50]. Although the above literature indicates a possible connection between EEG waves and SA-related construct, no obvious EEG absolute power differences were demonstrated in our study, thus rejecting H2. The results may indicate that absolute power as an EEG feature has limited contribution to detecting different SA levels.

In addition, the analysis of relative power in four waves was carried out. The results showed a significantly lower β relative power for the LSA group than the HSA group in central and parietal brain regions, thus partly confirming H3. The findings of β relative power differences in the central and parietal regions are consistent with the findings of Kästle et al. (2021) and Shayesteh and Jebelli (2022) [28,51]. The β wave activity is usually observed during the state of normal consciousness, alertness, or active concentration. According to previous studies, the β wave is normally considered to be associated with excitation, and is expected to be more active when the SA is higher and significantly reduced under fatigue [20,21,25]. In the present study, the LSA group may have had lower alertness or mental arousal, which was not conducive to the perception, comprehension, and prediction of the situation, thus causing poor SA. The central brain region is considered to be associated with sensory and motor functions [38], while the parietal brain regions control the sensory and motor systems, awareness, and perception [39,52]. The findings in our research provide support that the two brain regions are both closely correlated to the cognitive processes of SA, which could provide evidence for exploration of the underlying neurophysiological mechanisms of SA.

In addition, analysis of the four SW/FW ratios (θ/β, α/β, (θ + α)/β, and (θ + α)/(α + β)) in the high- and low-SA groups was also conducted. The results showed that the four SW/FW ratios were sensitive to different SA groups, thus confirming H4, which could provide sensitive EEG features for SA classification in similar situations. The SW/FW ratios of EEG are typical indicators used by researchers to reflect the cortical control function over subcortical [20]. Studies have observed an increase in the four SW/FW ratios under fatigue, which can be applied in human alertness assessment [26,53]. The θ/β ratio was initially introduced in attention-deficit/hyperactivity disorder diagnosis, and recent studies explored its connection to various aspects of executive cognitive performance in healthy individuals, such as working memory, attention (especially executive attentional control), stress, and sleepiness [20,40,54,55]. The adoption of θ/β ratio in the current study was based on the assumption that the LSA group may have had decreased alertness and task engagement, which would have resulted in a decrease in β power and an increase in θ power, i.e., a higher θ/β [40]. The α/β ratio is often used to investigate an individual’s cognitive state such as mental attentiveness, mental fatigue, and stress [20]. Furthermore, (θ + α)/β has been used to study alertness and task engagement, mental attentional investment, mental effort, vigilance, and attention, while a high value of this index may be associated with a state of high vigilance [40]. Our findings are consistent with the above assumption, as higher SW/FW ratios were found in the LSA group, providing new and important SA-sensitive EEG features for SA discrimination.

According to the repeated ANOVA results, sensitive SA-related EEG features were found in our research, which were fed into the Bayes model for discrimination after PCA. According to calculation, the discrimination result of the Bayes method based on sensitive EEG features was 83.3% for the original validation and 70.8% for the cross-validation, thus H5 was confirmed. Up to now, a limited number of studies have been conducted on SA discrimination. For example, the 2020 model of Li et al. discriminated two fatigue-related SA levels (high and low SA), with accuracies recorded for the S–S model (70.6%), logistic regression (67.9%), support vector machine (63.8%), EEGNet-4.2 (59.0%), and EEGNet-8.2 (60.3%) [29]; Kästle et al. (2021) used Random Forest and Boosted Trees combined with all EEG data to discriminate SA levels, and achieved 67% accuracy [28]. In our study, the selection of sufficient SA-sensitive EEG features and adequate consideration of their contribution is conductive to the improvement of discrimination accuracy, as conflicting information is excluded to reduce the possible confusion. Therefore, the combination of SA-sensitive EEG features and a linear discriminant model can also achieve effective discrimination results, which can be a feasible direction in more realistic situations.

This study performed an EEG feature analysis related to different SA levels. It is possible to provide an in-depth analysis of the mechanisms underlying the decrease in SA through the findings, propose achievable preventive measures, and ensure safe completion of the operator’s task [19]. The research characteristics of our study include three aspects. First, performance measures were commonly used in most previous studies related to SA assessment and classification, however, the relationship between SA and performance was unclear [56], which reduces the reliability of research results. In our study, the SAGAT method was used to classify the high and low SA levels. As a highly sensitive, reliable, and predictive measure of SA that was useful across a wide variety of domains and experimental settings [13], the SAGAT method can improve the objectivity of SA grouping effectively [29]. Second, it is well accepted that physiological measurements have promising potential for online monitoring and identification of SA; However, most of the current studies have focused on eye movement measurements, and the correlation between EEG and SA remains unclear. For online discrimination of SA, EEG measurement has the advantages of high temporal resolution and direct measurement of population-level neural activity in the operator. According to our experiment, EEG features of beta relative power and SW/FW are found to be sensitive to different SA levels, thus providing useful indicators for the discriminative analysis of SA. Most studies on the discrimination of operator cognitive states focus on mental workload and fatigue, while the discriminative analysis of SA based on objective EEG features is rare and has limited predictive accuracy. Bayesian methods are applied on the basis of sensitive EEG features in our study, yielding an enhanced discriminative accuracy. In addition, compared with the machine learning methods, the results of classical Bayes discriminant model are more interpretable for SA discrimination mechanisms [19].

5. Conclusions

This study distinguished HSA and LSA groups according to SAGAT scores in MATB II tasks. EEG spectrum analysis of high- and low-SA groups was performed to explore SA-sensitive features, including the absolute power and relative power of four frequency bands, and the SW/FW ratios were selected. In addition, the Bayes method was used to further discriminate different SA groups, combined with SA-sensitive EEG features and PCA. According to the experimental results, several conclusions can be drawn. First, the LSA group had significantly lower SAGAT scores than the HSA group. Second, the absolute power of the four waves showed no significant difference between LSA and HSA groups. Third, the LSA group had a significantly lower beta relative power than the HSA group in the central and parietal brain regions. Fourth, compared with the HSA group, the LSA group had higher θ/β and (θ + α)/(α + β) in the frontal, central, and parietal regions, higher α/β in the parietal region, and higher (θ + α)/β in all regions except for the left and right laterality in the frontal region. Lastly, in combination with the above SA-sensitive EEG features, through the PCA and Bayes method, discrimination could be achieved between the two SA groups with 83.3% accuracy for the original validation and 70.8% accuracy for the cross-validation. These results provide evidence for online assessment and discrimination of SA using EEG features, thus contributing to dealing with SA decrement that might lead to threats to flight safety.

However, this study had some limitations. First, this was a pilot study in a laboratory environment; hence, the results need to be further validated in actual task situations, especially for operators in the field and on real operation platforms (typically professional pilots in high-fidelity flight simulators). Second, three or more SA levels can be defined to uncover the mechanisms between SA and EEG, and a more flexible objective and reliable SA measure is needed to set/collect more labels for machine learning and pattern recognition. Our research explored and validated the applicability of EEG features to assess and discriminate SA levels. Combinations with other physiological signals, such as eye tracking and heart rate variability (HRV), can also be helpful in future research on the discrimination of SA. Furthermore, future work can be carried out to explore various SA characteristics under different subject groups, such as the male/female groups, as well as professional pilots/flight cadets/student groups with an aviation background.