EEG-Based Measurement for Detecting Distraction in Coal Mine Workers

Abstract

In the high-attention-demanding environment of underground coal mines, distraction is a major cause of unsafe behavior and decreased safety performance. Research on the cognitive neural mechanisms and monitoring of distraction in miners is limited. This study used an electroencephalogram (EEG) to examine the correlation between distraction and brain activity in coal miners, aiming to provide an objective method for monitoring distraction in coal miners. Thirty participants completed a simulated hazard recognition task, using the Sustained Attention to Response Task (SART) and noise to induce distraction. Brain activity was recorded and labeled as focused or distracted based on the correctness of the hazard recognition task. EEG features were extracted and selected, and a Random Forest model for distraction identification was constructed based on the selected features. In the focused state, delta power in the temporal region and theta power in the frontal region increased significantly. In the distracted state, alpha power in the temporal and occipital regions and beta power in the occipital and parietal regions increased. The selected EEG features could be used to identify distraction with 84% accuracy. This method can objectively identify distraction in coal miners, highlighting the potential of using EEG for real-time distraction monitoring.

1. Introduction

Coal mining has garnered widespread attention due to its high accident rates. This persistent issue remains unresolved, even amid frequent and severe incidents [1]. Research shows that over 90% of these accidents are attributable to unsafe human behaviors [2]. With advancements in safety management and technological innovations, the focus on the pivotal role of coal miners in mitigating unsafe behaviors has intensified. Miners’ unsafe behaviors are influenced by the complex interplay between human factors, machinery, and the environment [3]. Critical individual factors, such as fatigue, stress, and decreased alertness, significantly contribute to these unsafe behaviors [4,5]. These adverse physiological and psychological states impair miners’ ability to filter irrelevant information, thereby affecting their attention to crucial details, leading to distractions [6].

Workplace distraction refers to the shift of attention from a primary task to irrelevant stimuli during task execution [7]. This shift can be induced by external factors, such as noise or changes in lighting, or internal factors, including fatigue and emotional states [8,9]. While distraction is a common issue across various industries, coal miners face distinct challenges. Underground coal mines are characterized by confined spaces with low visibility and constant, high-intensity noise generated by equipment, such as ventilation fans and conveyor belts. In addition to these auditory distractions, miners are also exposed to a range of environmental stressors, including airborne dust particles and toxic gases. According to the Coal Mine Safety Regulations issued by the State Administration of Work Safety [10], coal miners must meticulously follow operational instructions, monitor instrument readings and machinery conditions, and respond rapidly to anomalies. For example, miners are required to check gas concentrations during work. If the methane concentration in the mine tunnel reaches 2%, all work must stop, machinery must be shut down, and the situation must be reported to the control room.

From a Gestalt psychology perspective, attention and other cognitive functions in such environments may be influenced by a “cooperative” effect, where the combination of environmental and task-specific demands significantly heightens the risk and consequences of distraction in coal mining, underscoring the critical need for sustained attention in this setting.

Behavioral analyses of 200 coal mine incidents have identified distraction and negligence as significant contributors to unsafe behaviors, which in turn are key factors in coal mine accidents [11]. Mining tasks, typically involving low cognitive load and continuous similar work, often induce fatigue and reduced attention, increasing the risk of operational errors. This is evidenced by prolonged reaction times, higher error rates, and physiological indicators, such as elevated heart rates and increased blinking frequency [12]. Additionally, the high noise levels typical of coal mine environments can impair miners’ attention and response capabilities [13]. This auditory interference may prevent miners from perceiving critical warning sounds, such as trolley car bells, equipment alarms, or water seepage signals during emergencies, thereby elevating accident risk [14]. Distraction also contributes to the neglect of safety protocols and the adoption of risky behaviors, often culminating in severe incidents [15]. These findings underscore the urgent need to study the effects of distraction on coal miners to improve safety performance and mitigate accident risks.

Psychologists have traditionally relied on self-reports and task performance to understand and predict human behavior [16,17]. However, these metrics are limited predictors of behavior in specific contexts. Electroencephalography (EEG) offers a complementary approach by revealing the underlying connections between human behavior and brain activity. Due to its non-invasive nature, high temporal resolution, and low cost, EEG has been widely used to analyze and assess cognitive states. For instance, Jebelli et al. collected EEG signals from construction workers during on-site tasks and utilized time-domain and frequency-domain features to identify work-related stress with an accuracy of 80.32% [18]. Similarly, Guo et al. used single power spectral density (PSD) features in the alpha and beta bands to assess driver alertness in a simulated driving environment, achieving an accuracy of 82.8% [19]. Mehmood et al. employed wearable EEG sensors to record the brain activity of construction workers, developing a mental fatigue detection model with an accuracy exceeding 90% [20]. Furthermore, other researchers have extensively used EEG to investigate emotions [21], risk perception [22], workload [23], and situational awareness [24].

Previous studies have demonstrated that frequency-domain features are key EEG indicators for identifying distraction across various fields. For example, frequency-domain features in the prefrontal and temporal regions effectively distinguish between focused and distracted states in construction workers [25]. Theta-band activity in the prefrontal cortex has been used to detect driver distraction [26]. Additionally, alpha-band activity in the temporal region has been suggested to correlate with distraction [27]. The P300 component, a widely studied event-related potential, has also been used to reflect attention levels in response to stimuli [28]. However, few studies have addressed the unique environmental challenges of coal mining and explored miner distraction from the perspective of brain activity.

To address this research gap, we propose a method utilizing EEG devices to monitor the distraction of coal miners. By identifying EEG features that distinguish between focused and distracted states, we developed and evaluated a model for recognizing miner distraction. Using the aforementioned sensitive indicators, we anticipate that the proposed model can objectively and accurately identify distraction among coal miners. The results are expected to provide novel technical solutions and scientific insights for distraction detection, thereby mitigating the risks associated with unsafe behaviors and enhancing workplace safety and health protocols.

2. Experiment

2.1. Participants and Equipment

This study recruited 30 male graduate students from the School of Safety Science and Engineering at Xi’an University of Science and Technology, all with foundational knowledge in coal mine safety management. Participants, averaging 24.95 ± 3.96 years old, had normal or corrected vision and hearing, clean scalps, and were right-handed. Due to the low accuracy rates (below 50%) of Subjects 7 and 14 in the hazard recognition task, their data were excluded, resulting in 28 effective datasets. This study received ethical approval from the Ethics Committee of Xi’an University of Science and Technology and adhered to the latest version of the Declaration of Helsinki. Written informed consent was obtained from each participant before any procedures were conducted.

EEG signals were collected using a Neuroscan EEG system (Compumedics Neuroscan, Charlotte, NC, USA) with a 64-electrode cap placed according to the international 10–20 system (Figure 1). The reference electrode was placed at the Cz location, which is centrally located on the scalp to provide a reliable and stable signal reference. The ground electrode was positioned at the Fpz location to minimize external noise interference and ensure proper grounding of the EEG system. The recording frequency was 1000 Hz, and scalp resistance was kept below 10 kΩ. The participants acclimated to the equipment before the experiment and could stop at any time if uncomfortable.

2.2. Stimuli

Stimuli for the hazard identification task were presented in the form of images against a black background on a computer screen. These images, sourced from coal mining enterprise safety training material, depicted various scenes from coal mines, and were categorized into hazardous and safe types, as shown in Figure 2. Hazardous images depicted potential dangers within the coal mining scenes, such as personnel without safety helmets, loose equipment screws, exposed electrical wires, and obstructions in conveyor belts. Safe images, on the other hand, depicted scenes free of any potential hazards.

Due to the specific requirements of EEG experiments, the coal mine work scene images used for experimental stimuli were standardized to avoid inconsistencies in the difficulty of hazard recognition and disputes over the presence of hazards in the images. A pre-experiment was conducted with 10 participants recruited from the same population as the main study. Participants evaluated 60 hazardous and 60 safe coal mine images sourced from official safety training materials. The evaluation criteria included recognizability and difficulty. Each image was rated on a 5-point scale (1 = very easy to identify, 5 = very difficult to identify).

Based on these ratings, images with an average score between 5 and 6 were selected to ensure a moderate and consistent level of challenge across trials. Images with extreme ratings (e.g., overly obvious or overly ambiguous hazards) were excluded. Ultimately, 30 images depicting various hazardous situations were selected, involving common issues, such as violations by personnel, hidden dangers at the workplace, and abnormal conditions of production facilities and equipment, ensuring these situations were broadly representative. Additionally, 30 safe scene images were matched to these hazardous situations, each representing the rectification of one of the 30 hazardous conditions, to ensure consistency of the scenarios.

Numerous studies provide compelling evidence that noise significantly contributes to workplace distraction [25,29]. To assess distraction, our experiment involved two distinct groups: a control group and a noise group. Participants in the control group performed their tasks in a quiet setting, whereas those in the noise group were exposed to simulated disruptions via noise. According to “Coal Mine Safety Regulations,” miners exposed to noisy environments for more than eight hours daily should not exceed 85 dB [10]. Consequently, in our study, participants in the noise group were exposed to 85 dB, ensuring their auditory health while replicating real-world coal mining conditions. The noise, sourced from typical mining equipment, such as ventilation fans, coal cutters, and scraper conveyors, was delivered to participants via headphones at the specified dB level.

2.3. Procedure

The Sustained Attention to Response Task (SART) is a widely adopted experimental paradigm in psychology and neuroscience, aimed at assessing individuals’ sustained attention and their ability to detect unexpected events [30]. In SART, participants are asked to perform a repetitive response task, where the “target” stimuli require a response, while other “non-target” stimuli do not. This requires special attention and a response, allowing the assessment of participants’ sustained attention and their ability to inhibit responses to non-target stimuli. In our study, we integrated a hazard recognition task into the SART framework, where hazardous images served as target stimuli and safe images as non-target stimuli, mirroring miners’ daily hazard identification challenges. Both hazardous and safe images were evenly distributed with a 1:1 ratio across all tests.

Each trial began with a red “+” fixation point displayed for 1000 ms, followed by a randomly presented safe or hazardous image (Figure 3). Subjects were required to quickly assess potential hazards within 6000 ms of the image’s appearance. Failure to respond within this timeframe was noted as an error in hazard recognition. After each response or timeout, the image disappeared, and the next trial started. To prevent learning biases, the sequence of testing in quiet and noisy environments was randomized across participants. Each test consisted of 30 trials, resulting in a total of 60 trials per participant.

Between test sessions, participants rated their attention levels using a five-question scale, such as “I can maintain focus during tasks,” scored from 1 (strongly disagree) to 5 (strongly agree), with a maximum total score of 25. The experimental setup involved E-Prime software (Psychology Software Tools, Inc., Pittsburgh, PA, USA) and EEG equipment to record responses and brain activity simultaneously. Although self-reported attention scores provided valuable insights into participants’ perceived attention levels, they may have been influenced by biases, such as social desirability or recall limitations. Therefore, combining self-reported attention scores, task performance, and EEG data helps to reduce subjective biases and provides a more comprehensive and objective reflection of participants’ attention states and corresponding brain activity.

3. Methods

This study collected behavioral performance, brain electrical activity, and subjective attention reports to comprehensively and objectively assess participants’ attentional states during the experiment. Behavioral data from E-Prime included trial correctness (binary 0 or 1) and reaction times (milliseconds), along with subjective attention self-assessments. These data were analyzed using SPSS (IBM, Armonk, NY, USA), providing a reference for synchronous brain activity analysis. The EEG data processing framework comprises four stages (Figure 4): preprocessing, feature extraction, feature selection, and model training and validation. Initially, standard EEG preprocessing protocols ensured data quality. Various feature extraction techniques then identified candidate features related to distraction, based on distraction indices previously validated in the literature. These features were refined using filter and embedded methods to effectively distinguish between focused and distracted states. Finally, a model recognizing distraction states was developed and its reliability validated.

3.1. EEG Preprocessing

The raw EEG data were preprocessed using the EEGLAB toolbox in MATLAB (MathWorks, Natick, MA, USA). Initially, a basic Finite Impulse Response (FIR) filter in EEGLAB retained EEG signals within the 1–40 Hz range, suppressing extraneous frequencies. This range was chosen because biologically meaningful EEG information is predominantly found within 1–40 Hz, while frequencies outside this range often represent noise, such as electromyographic activity and electrical interference. A 50 Hz notch filter was also applied to eliminate environmental electrical noise. After filtering, Independent Component Analysis (ICA) was used to remove non-neural artifacts. Using expert knowledge and the ADJUST plugin [31], we reviewed each participant’s ICA results to identify and remove components associated with eye movements, muscle activity, and other non-neural noises. The remaining independent components were then re-projected onto the channel level, resulting in clean EEG data, free from artifacts.

To meet the requirements for Event-Related Potential (ERP) and spectral analyses, we segmented the purified EEG data. For ERP analysis, each trial’s EEG data were divided into 1000 ms segments with a time window of [−200, 800] ms, using the first 200 ms as the baseline correction period. For spectral analysis, 3000-ms segments were selected with a time window of [−1000, 2000] ms, using the first 1000 ms as the baseline. Amplitude thresholds were set at ±100 μV to exclude trials with amplitudes outside this range, ensuring the analysis was not contaminated by extreme artifacts.

After segmenting the EEG data by trial, they were regrouped into “focused” and “distracted” categories. Participants’ attention during tasks was dynamic and varied individually. Noise interference typically increases the likelihood of distraction, but not all participants are equally affected. Some maintain high focus despite noise, while others are more sensitive to it. Based on task performance, EEG data were categorized: in the control group, trials where participants correctly identified hazardous images were marked as “focused”; in the noise group, trials where participants failed to identify hazards correctly were marked as “distracted.” We compiled a total of 840 epochs, with 571 correct trials in the control group and 269 incorrect trials in the noise group. After stringent quality control to exclude non-neuronal artifacts, 532 of the correct trials and 231 of the incorrect trials were deemed valid. Thus, we obtained 532 valid epochs of focused states and 231 valid epochs of distracted states.

3.2. EEG Features Extraction

3.2.1. Time Domain

ERPs are widely used to reveal the brain’s response patterns to specific stimuli, with the P300 component being a sensitive indicator of attention allocation due to its amplitude and latency [28]. When participants focus on or anticipate a stimulus, the P300 waveform exhibits significant amplitude changes. To quantify this attention-related brain activity, we selected the peak amplitude of the P300 component as a key EEG feature. For all trials in the focused and distracted groups, a time window of [−200, 800] ms around stimulus presentation was chosen, and epoch data were averaged across participants to enhance the signal-to-noise ratio (SNR) and clarify ERP components. By averaging ERP waveforms across the two groups, we identified the peak latency of the P300 waveform and calculated the average amplitude within 20 ms before and after this peak, representing the peak amplitude of the P300 waveform for both groups.

3.2.2. Frequency Domain

Spectral analyses are crucial for understanding brain dynamics and neural synchrony, transforming time-domain signals into frequency-domain spectra using mathematical techniques like the Fast Fourier Transform (FFT). This process intuitively displays the energy distribution across frequency components. Given that EEG signals are discrete, the Discrete Fourier Transform (DFT) method is typically used, where the DFT is defined for a signal of length $N$ as Equation (1):

(1)$X\left(k\right)=\sum _{n=0}^{N-1}x\left(n\right)\cdot W_N^{kn},W_N=e^{-j{\displaystyle \frac{2\pi }{N}}}$

where $x\left(n\right)$ represents the input signal, $N$ is the total number of samples, $k$ is the frequency domain index, $X\left(k\right)$ denotes the Fourier component at frequency $k$, and $W_N$ is the kernel function that represents the relationship between discrete frequency points. In this study, FFT was performed on EEG data from two groups, categorizing the resulting spectra into standard EEG frequency bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–12 Hz), and beta (12–30 Hz). We calculated the average power within each band to assess neural activity intensity. Additionally, we incorporated various power ratio indices across different frequency bands, such as $\theta /\beta$, $\theta /\alpha$, $\alpha /\beta$, $(\alpha +\theta )/\beta$, $(\alpha +\beta )/\theta$, and $(\alpha +\theta )/(\alpha +\beta )$, to evaluate cognitive states [32,33,34]. Thus, these indices were also incorporated into our consideration of spectral features.

3.3. EEG Features Selection

Feature selection is essential for identifying the most predictive subset of features, enhancing model performance, simplifying structure, improving interpretability, and reducing overfitting. In this study, we used a strategy combining statistical tests with Random Forests.

Step 1: Preliminary screening via statistical tests

We employed the Wilcoxon rank–sum test to identify features with significant differences between the two groups. This non-parametric test is suitable due to its lack of strict requirements regarding data distribution and sample size, making it ideal for our non-normally distributed data. Using MATLAB, we conducted the Wilcoxon rank–sum test on all EEG features extracted across all channels, setting the significance level at 0.05.

Although multiple statistical tests were conducted in the feature screening process, we did not apply multiple comparison corrections at this stage. This decision was made because the aim of this step was exploratory: to identify potential EEG features associated with distraction. Applying strict corrections might have overly constrained the findings and limited the discovery of meaningful patterns. This procedure identified features and channels with statistically significant differences. The identified features were later validated during the model-building phase, where we carefully interpreted the features in the context of known neurophysiological mechanisms.

Step 2: Feature selection based on feature importance

Random Forest (RF), an ensemble learning method, was used for feature selection due to its ability to assess feature importance through the Mean Decrease Gini (MDG). MDG quantifies the importance of each feature by measuring its impact on classification accuracy. In this study, using the significant EEG features identified in the first step, we employed RF to select the most impactful features for our model’s predictive performance.

Due to the inherent randomness of the RF algorithm, which can lead to variability in accuracy, we trained 30 independent RF models using the filtered feature set to obtain a more stable estimate of feature importance. Averaging the feature importance scores from these models mitigated the randomness of individual models. We then selected the top 40 features based on average importance scores as the candidate subset. Using a stepwise feature addition strategy, we evaluated model performance. Starting with the highest-ranked feature, we sequentially added features to the model, reassessing accuracy after each addition. By monitoring the trend in accuracy, we identified the point at which it stabilized, indicating that adding more features no longer significantly improved performance. This set of features was determined to be the final feature subset.

3.4. Model Training and Validation

RF was also widely used for classification and regression tasks. They enhance accuracy and generalization by aggregating predictions from multiple decision trees. This process involved bootstrap sampling to create multiple independent subsets from the original dataset, with a decision tree constructed for each subset. Individual tree predictions were aggregated via majority voting to produce the final output. The model’s performance is influenced by two key parameters: the number of trees and the minimum sample count per leaf node.

In this study, the identified feature subset served as the model’s input, with subjects’ attention state labels as the output. To address the imbalance between the two groups, we employed the Synthetic Minority Over-sampling Technique (SMOTE) to oversample the minority group, creating a rebalanced dataset to improve the model’s recognition of minority samples [35]. Subsequently, grid search was used to optimize model parameters—namely, the number of trees and the minimum sample size per leaf—combined with 5-fold cross-validation to mitigate overfitting risks.

Given potential inter-individual variability, the trained model was validated separately on each subject’s dataset. To comprehensively assess feature extraction effectiveness and feature set performance in classification, we computed accuracy, precision, recall, and F1 score for models trained on raw extracted features, statistically filtered features, and optimally selected features. Receiver Operating Characteristic (ROC) curves and their Areas Under the Curve (AUC) were used to evaluate the discriminatory power of the different feature sets, providing a robust benchmark for comparative analysis.

4. Results

4.1. Behavioral and Subjective Responses

Given the non-uniform variance and rejection of normality in the behavioral data, along with differences in the number of trials between the focused and distracted groups, we employed the non-parametric Wilcoxon rank–sum test for analysis. The statistical results for accuracy, response time, and subjective attention evaluations in the hazard identification task for both control and noise test groups, as well as between the focused and distracted groups, are presented in Table 1.

The results showed a significant reduction in mean response times for the noise group compared to the control group (1385.95 ms vs. 1646.59 ms, p < 0.001). However, no significant difference in mean response accuracy was observed between the two groups. This aligns with Golmohammadi et al. [36], who suggest that high-intensity noise exposure may expedite reaction times. In other words, intense noise levels appear to amplify psychological stress, prompting individuals to respond quickly as a coping mechanism against auditory distraction.

The self-reported attention scores of the control group were significantly higher than those of the noise group (18.68 vs. 14.82, p < 0.001). This suggests that the noise condition likely contributed to the observed distraction. According to the load theory of attention, noise acts as a distraction whose impact is mediated by the cognitive demands of the primary task [37]. High cognitive load tasks occupy attentional resources, reducing susceptibility to distractions like noise. Conversely, low cognitive load tasks leave surplus attentional resources vulnerable to noise interference [29,38]. For coal miners, hazard identification is a routine task with a relatively low cognitive load, making them more susceptible to the effects of noise.

Following the categorization protocol, which divided all trials into focused and distracted groups, we found no statistically significant difference in mean response times between these groups (1514.80 ms vs. 1538.94 ms, p = 0.782). This indicates the absence of a meaningful statistical difference.

4.2. EEG Features Extraction

4.2.1. Time Domain: P300 Amplitude

To elucidate the neurophysiological differences between the focused and distracted groups, we computed ERP waveforms across all electrodes within the [−200, 800] ms post-stimulus time window. Using the Wilcoxon rank–sum test, we assessed statistical significance for voltage amplitudes at each millisecond for both groups. Figure 5 shows the ERP waveforms at the PZ channel, with corresponding p-values. The P300 component peaks around 360 ms post-stimulus, with significantly greater amplitude in the focused group than in the distracted group. This indicates successful elicitation of the P300 component, suggesting heightened attention and greater task engagement in the focused group. Conversely, the attenuated P300 amplitude in the distracted group suggests diverted attentional resources.

To enhance stability in quantifying the P300 peak voltage, we averaged voltage amplitudes within a 360 ± 20 ms window post-stimulus, centered around the expected P300 peak. This procedure was consistently applied across all channels. Using the Wilcoxon rank–sum test, we compared these representative P300 peak voltages across all electrodes for both groups. Figure 6 summarizes the comparative results of P300 amplitudes and their significance levels across channels. Significant differences in P300 peak voltages were found at specific channels, namely CP3, CPZ, P3, P1, PZ, and P2. Based on these findings, we selected the recorded P300 peak voltages from these significant channels for further investigation.

4.2.2. Frequency Domain: Spectral Amplitude

We examined the average FFT results from the whole-brain electrodes of both groups. Since no significant differences were observed in the overall average power across electrodes, we conducted comparative analyses at the channel level. Figure 7 illustrates the mean power spectra within the 1–40 Hz frequency band at the O1 channel for both groups, along with their significance levels. The power spectral intensity shows that, compared to the focused group, the distracted group exhibited elevated EEG power spectra in the mid-to-high frequency range (8–40 Hz, encompassing α and β bands). This aligns with the findings of Awais et al., suggesting that increased α and β band power may be associated with a decline in alertness and attention levels [39].

To elucidate the differential EEG power distribution across frequency bands between the two groups, we segmented the power spectra into four key bandwidths: δ (1–4 Hz), θ (4–8 Hz), α (8–12 Hz), and β (12–30 Hz). We excluded the γ band (above 30 Hz) due to weak signal strength and susceptibility to noise. Figure 8 shows topographical maps illustrating the average power in these bands and their statistical significance. The focused group exhibited higher mean power in the δ and θ bands compared to the distracted group, with significant differences in the δ band over the right temporal region (T8) and in the θ band across the frontal areas (FP1, FPZ). Conversely, the distracted group showed elevated mean power in the α and β bands, with significant differences in the α band in the left temporal and occipital areas (TP7, T7, P3, P1) and in the β band in the occipital and parietal areas (O1, PO7, C3, C4). We incorporated the average power from these significant electrode frequencies as indicators for subsequent investigations.

4.2.3. Frequency Domain: Band Power Ratios

Figure 9 presents topographical maps illustrating the scalp distribution of various EEG frequency band power ratio indices for both groups. The focused group exhibit significantly higher values in the occipital region for three power ratio indices: θ/β (Index A), α/β (Index C), and (α + θ)/(α + β) (Index F) compared to the distracted group (p < 0.05). Conversely, for the (α + β)/θ (Index E) ratio, the distracted group shows significantly higher values in the occipital area compared to the focused group (p < 0.05). Although spatial variations in θ/α (Index B) and (α + θ)/β (Index D) ratios between the groups are noticeable in the topographic maps, these differences are not statistically significant.

These comparative results suggest significant differences at most occipital electrodes (PO3, PO5, POZ) and some parietal electrodes (P1, P3, PZ) across different attentional states. Therefore, the band power ratio indices from these significantly different electrodes are also incorporated as candidate indices for input into the classifier.

Through the feature extraction and significance testing of the EEG data described above, we successfully constructed a comprehensive feature set consisting of 143 features (Table 2).

4.3. Model Development and Performance Evaluation

4.3.1. Feature Selection

In our pursuit of extracting crucial features and streamlining our feature set, we employed a feature selection method based on feature importance. This approach involved working with a dataset comprising 763 stratified samples (532 focused and 231 distracted samples), with the minority distracted group augmented using SMOTE to address class imbalance [35]. To reduce the variability inherent in individual RF iterations, we trained an ensemble of 30 distinct RF models and aggregated their average feature importance scores. These averaged scores were subsequently ranked, and for clarity in visualization and to simplify chart complexity, only the top 40 features were depicted in the feature importance ranking graph, as shown in Figure 10.

Given the pronounced disparities in feature importance, it is essential to streamline and optimize the input features to avoid redundant information that could undermine model performance. Therefore, we selected the top 40 features based on their average importance rankings and incrementally incorporated them into the training model to examine how model accuracy evolves with increasing feature counts (Figure 11). As illustrated in the figure, the RF model initially exhibits low accuracy levels with a limited feature set, indicating insufficient critical feature support. As the number of features increases, model accuracy significantly improves, reaching 85.1% with 15 features. This suggests the model effectively captures the essential features for classification performance. Beyond this point, accuracy stabilizes with minor fluctuations, indicating marginal or negligible gains from additional features. In some cases, an excess of features even reduces accuracy, suggesting the detrimental effects of superfluous or noisy features. Consequently, we selected the top 15 features by importance as the optimal set for our classification model.

4.3.2. Model Training and Validation

To develop our model, we performed hyperparameter tuning using grid search with five-fold cross-validation to minimize the impact of randomness on experimental outcomes. The grid search explored the number of trees (n) from 50 to 400 in increments of 50 and minimum leaf sizes (l) from 1 to 5 in increments of 1. As shown in Figure 12, the model achieved its highest score of 0.857 with n set to 200 and l set to 1, indicating this as the optimal parameter combination for model performance.

To validate our feature selection methods, we conducted a comparative evaluation of model performance using three distinct feature subsets: all EEG features, features selected through statistical significance testing, and features optimized via algorithms. The results are summarized in Figure 13, where the model using statistically significant features outperformed the baseline model employing all EEG features. This improvement suggests that removing insignificant features enhances the model’s ability to identify crucial ones. Notably, the model trained on optimized features achieved the highest performance across all metrics, reaching an accuracy of 84%. It also demonstrated balanced precision and recall, indicating reduced false positives and negatives without overfitting.

Furthermore, we assessed the discriminative ability of models corresponding to the three feature subsets by plotting ROC curves (Figure 14). When trained using all EEG features, the model achieved an AUC of 0.62, suggesting predictive ability slightly better than random chance and limited reliability for some applications. In contrast, using features selected through statistical significance testing increased the model’s AUC to 0.73, indicating a significant improvement in distinguishing between positive and negative samples compared to using all features. Notably, employing features optimized by the algorithm boosted the AUC to 0.86, underscoring the crucial role of our feature selection strategy in enhancing the model’s discriminative performance.

To account for individual variability, we conducted comparative validation by evaluating the performance of two models on independent datasets from all participants: one model was trained using optimized features, while the other model was trained with all features. Performance evaluation results for both models are summarized in Table 3, demonstrating a notable enhancement in overall classification performance with the optimized feature set. This approach illuminates performance variability across individuals, thereby assessing the model’s flexibility and applicability.

5. Discussion

This study developed an EEG-based computational framework for monitoring distraction in coal miners. Key findings include (1) the extraction and identification of EEG features that reliably distinguish between focused and distracted states, and (2) the construction and evaluation of the distraction recognition model. The following discussion addresses feature implications, model performance, and limitations.

5.1. Feature Implications

The interpretable nature of the RF algorithm highlights the significance of EEG features. The feature importance ranking shows that the EEG characteristics most correlated with attentional states predominantly cluster in the temporal (TP7, T8) and frontal (FP1, FPZ) regions.

The temporal region, crucial for auditory processing and sound integration, may exhibit increased EEG activity due to noise, potentially obscuring task-related neural dynamics [40]. Environmental noise, as a non-targeted auditory input, stimulates EEG activity in the temporal area, diverting attentional resources and impacting task engagement [41]. In this study, the distracted group showed a significant increase in alpha band power in the temporal region (TP7) compared to the focused group, consistent with previous findings [27]. Research has shown that alpha band oscillations may inhibit neural activities unrelated to the current task, improving task-relevant information processing [42]. This suggests the brain enhances noise suppression in the temporal region during distraction to reduce task interference. Additionally, the focused group exhibited a larger P300 amplitude in the temporal region, indicating more cognitive resources allocated to the task and greater attention to task-related stimuli. Since P300 amplitude reflects the amount of attention paid to stimuli, it is commonly used to measure cognitive resource allocation [28]. These findings underscore the critical role of EEG indicators in the temporal lobe for distinguishing attention states.

The prefrontal cortex, particularly the anterior segment, is associated with advanced cognitive functions, decision-making, and attentional control, including the suppression of irrelevant stimuli and the maintenance of task-relevant information [43]. Notably, in the focused group, significant enhancements in theta band activity were observed over prefrontal regions (FP1, FPZ). This is consistent with the findings of Wen et al. [44], who state that this increase in theta band power indicates an amplification of neural processes underlying attentional control, highlighting the critical role of prefrontal theta oscillations in modulating attention.

Our analysis revealed a significant increase in beta power in the occipital (O1, PO7) and parietal (C3, C4) regions in the distracted group, consistent with the findings of Craig et al. [45]. Studies have shown that as mental activity decreases, the brain attempts to maintain alertness, increasing beta power in these regions. Additionally, power ratio features in the parietal and occipital lobes, such as the theta/beta ratio, are closely related to attention [33,46]. The theta/beta ratio serves as a biomarker for attention control and is negatively correlated with distraction [47,48]. Our findings support this, as the focused group exhibited a higher theta/beta ratio in the occipital region (PO3) compared to the distracted group.

Compared to research on distraction in other industries, the EEG features identified in this study reveal distinct characteristics specific to coal miners. Unlike construction workers, coal miners perform cognitively demanding, yet monotonous, tasks requiring sustained attention. This difference may explain the prominence of P300 amplitudes in temporal regions and the increased reliance on prefrontal theta oscillations in coal miners, which are less evident in construction workers [25]. These findings highlight the necessity of designing EEG-based monitoring systems tailored to occupational contexts, incorporating the unique cognitive demands and environmental conditions of specific industries.

5.2. Model Performance

To evaluate the model’s effectiveness and generalizability, we employed a dual-evaluation approach using both pooled and individual datasets. The pooled dataset assessed the model’s ability to generalize across diverse data. Our results demonstrated a significant accuracy improvement from 61% to 84% when trained with optimized features, alongside notable enhancements in precision, recall, and AUC metrics. On the individual datasets, the model exhibited strong adaptability to different subjects, with average accuracy increasing from 70.55% to 95.2% following feature optimization. These findings underscore the importance of identifying features that reliably distinguish between attention states. Moreover, they affirm the positive impact of feature reduction on optimizing computational efficiency and boosting classification performance.

Despite extensive research on EEG-based distraction detection in various fields [25,49,50], studies on distraction recognition among coal miners remain limited. Our research fills this gap, demonstrating the feasibility of using EEG data to monitor the distraction state of coal miners. The robust performance of the distraction recognition model demonstrates the feasibility of distinguishing between focused and distracted states in coal miners and highlights its potential utility in real-world applications. These findings provide a foundation for integrating EEG-based monitoring systems into safety protocols, potentially enabling real-time attention tracking to mitigate risks associated with distraction. Furthermore, embedding this technology into wearable safety equipment, such as helmets, could significantly enhance hazard recognition capabilities in high-risk environments.

5.3. Limitations

While this study demonstrates the feasibility of using EEG to monitor distraction in controlled environments, several challenges must be addressed before applying this method to real-world coal mining operations. First, EEG signals are highly susceptible to noise and artifacts caused by environmental electromagnetic interference and physical movements, which are prevalent in mining settings. Future studies should explore advanced artifact-removal algorithms and robust hardware designs to mitigate these issues. Second, integrating EEG devices with miners’ safety helmets presents significant design challenges. Current EEG systems, while effective in controlled environments, may lack the durability, comfort, and usability required for prolonged use in harsh mining conditions. Developing lightweight, wearable EEG devices specifically tailored for mining scenarios is a crucial step toward practical implementation. Additionally, miners perform diverse tasks with varying cognitive demands, necessitating further validation of the model’s applicability across different scenarios. Future research could employ task-specific EEG feature selection or adaptive machine-learning approaches to address this challenge.

6. Conclusions

Monitoring coal miners’ distraction is a promising approach to reducing unsafe behaviors and enhancing the overall safety of coal mining systems. This study proposed an objective method for recognizing coal miners’ distraction based on an EEG signal. The results demonstrated that the selected EEG features could effectively distinguish between focused and distracted states, achieving an accuracy of 84%. This finding suggests the feasibility of daily attention monitoring in underground coal mines. By embedding EEG devices into safety helmets, it is possible to achieve real-time monitoring of miners’ distraction, providing early warnings of potential unsafe behaviors, thereby improving the safety of coal mining systems.

Author Contributions

Conceptualization, Y.K. and S.T.; methodology, Y.K. and S.T.; software, C.Y.; validation, Y.K., C.Y. and L.C.; formal analysis, Y.K.; investigation, Y.K. and S.T.; resources, Y.K. and S.T.; data curation, L.C.; writing—original draft preparation, Y.K.; writing—review and editing, S.T. and H.L.; visualization, C.Y.; supervision, S.T. and H.L.; project administration, H.L.; funding acquisition, H.L. All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

The study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of Xi’an University of Science and Technology.

Informed Consent Statement

Informed consent was obtained from all subjects involved in the study.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Conflicts of Interest

The authors declare no conflicts of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. EEG equipment and electrode distribution.

Figure 2. Sample images of hazard recognition task. (a) Safe image. (b) Hazardous image.

Figure 3. Experiment procedure.

Figure 4. Overview of proposed data-processing framework.

Figure 5. ERP waveforms and statistical analysis for two groups. (a) ERP waveforms on PZ channel, with the P300 peak labeled; (b) corresponding p-values at each time point. The red line in Figure 5 and Figure 7 represents a significance level of p = 0.05. Data points below this line indicate a statistically significant difference.

Figure 6. P300 peak amplitudes and p-values across channels for two groups.

Figure 7. Spectral amplitude and statistical analysis for two groups. (a) Spectral amplitude on O1 channel; (b) corresponding p-values at each frequency band.

Figure 8. Average power of 4 frequency bands and p-values across channels for two groups.

Figure 9. Frequency band power ratio indices and p-values across channels for two groups.

Figure 10. Ranking of the top 40 features by importance. The naming convention for features is: Indicator_Channel.

Figure 11. Model accuracy by number of features.

Figure 12. Results of the grid search.

Figure 13. Performance metrics of models with different feature combinations.

Figure 14. ROC curves of models with different feature combinations.

Table 1

Statistical results for different environmental test groups and different attention groups.

	Group	M ± SD	Sample Size	U	p
Response accuracy	Control	0.68 ± 0.17	28	−0.42	0.675
	Noise	0.68 ± 0.13	28
Self-reported attention level	Control	18.68 ± 3.07	28	−3.809	<0.001
	Noise	14.82 ± 3.85	28
Response time (ms)	Control	1646.59 ± 1039.01	840	−5.987	<0.001
	Noise	1385.95 ± 889.22	840
	Focused	1514.80 ± 959.34	532	0.276	0.782
	Distracted	1538.94 ± 985.36	231

Table 2

EEG feature indicators and the number of significant features.

	Indicators		Number of Significant Features
Time domain	P300 amplitude		10
Frequency domain	Spectral amplitude	δ-band	1
		θ-band	4
		α-band	19
		β-band	19
	Band power ratios	A	10
		C	33
		E	26
		F	21
Total			143

Table 3

Performance comparison of models with all features and optimized features on different subject datasets.

Subject No.	Trial Size	All Features as Inputs				Selected Features as Inputs
		Accuracy	Precision	Recall	F1	Accuracy	Precision	Recall	F1
1	32	65.63%	81.67%	57.69%	52.11%	93.75%	93.52%	93.52%	93.52%
2	26	65.38%	82.00%	55.00%	48.12%	92.31%	91.67%	93.75%	92.12%
3	30	73.33%	86.67%	50.00%	42.31%	100.00%	100.00%	100.00%	100.00%
4	27	74.07%	87.04%	50.00%	42.55%	96.30%	93.75%	97.50%	95.38%
5	31	87.10%	93.55%	50.00%	46.55%	93.55%	83.33%	96.30%	88.08%
6	30	76.67%	88.33%	50.00%	43.40%	96.67%	97.92%	92.86%	95.09%
7	19	73.68%	86.84%	50.00%	42.42%	94.74%	91.67%	96.43%	93.60%
8	20	65.00%	82.50%	50.00%	39.39%	100.00%	100.00%	100.00%	100.00%
9	29	65.52%	82.76%	50.00%	39.58%	93.10%	91.67%	94.74%	92.68%
10	32	81.25%	90.63%	50.00%	44.83%	100.00%	100.00%	100.00%	100.00%
11	26	69.23%	84.62%	50.00%	40.91%	88.46%	87.59%	84.72%	85.95%
12	25	72.00%	37.50%	47.37%	41.86%	96.00%	92.86%	97.37%	94.80%
13	29	75.86%	87.93%	50.00%	43.14%	96.55%	97.83%	92.86%	95.04%
14	31	74.19%	87.10%	50.00%	42.59%	100.00%	100.00%	100.00%	100.00%
15	27	62.96%	81.48%	50.00%	38.64%	92.59%	94.74%	90.00%	91.67%
16	20	90.00%	95.00%	50.00%	47.37%	90.00%	75.00%	94.44%	80.39%
17	32	59.38%	79.69%	50.00%	37.25%	96.88%	96.43%	97.37%	96.80%
18	32	71.88%	85.94%	50.00%	41.82%	90.63%	87.73%	90.10%	88.77%
19	24	58.33%	77.27%	58.33%	49.58%	95.83%	96.15%	95.83%	95.83%
20	30	66.67%	34.48%	47.62%	40.00%	90.00%	89.20%	86.51%	87.69%
21	36	69.44%	60.29%	52.55%	48.37%	97.22%	98.08%	95.45%	96.64%
22	19	68.42%	83.33%	57.14%	52.50%	89.47%	92.86%	85.71%	87.82%
23	23	47.83%	73.91%	50.00%	32.35%	100.00%	100.00%	100.00%	100.00%
24	25	72.00%	86.00%	50.00%	41.86%	100.00%	100.00%	100.00%	100.00%
25	26	80.77%	90.00%	58.33%	58.73%	96.15%	92.86%	97.50%	94.87%
26	25	68.00%	83.33%	55.56%	50.00%	92.00%	94.44%	88.89%	90.81%
27	30	66.67%	82.76%	54.55%	47.92%	93.33%	92.82%	92.82%	92.82%
28	27	74.07%	87.04%	50.00%	42.55%	100.00%	100.00%	100.00%	100.00%
Overall	763	70.55%	80.70%	51.58%	44.24%	95.20%	94.00%	94.81%	93.94%

References

1. Tian, S.; Mao, J.; Li, H. Disaster-Causing Mechanism of Hidden Disaster-Causing Factors of Major and Extraordinarily Serious Gas Explosion Accidents in Coal Mine Goafs. Sustainability; 2022; 14, 12018. [DOI: https://dx.doi.org/10.3390/su141912018]

2. Yin, W.; Fu, G.; Yang, C.; Jiang, Z.; Zhu, K.; Gao, Y. Fatal gas explosion accidents on Chinese coal mines and the characteristics of unsafe behaviors: 2000–2014. Saf. Sci.; 2017; 92, pp. 173-179. [DOI: https://dx.doi.org/10.1016/j.ssci.2016.09.018]

3. Wang, C.; Wang, J.; Wang, X.; Yu, H.; Bai, L.; Sun, Q. Exploring the impacts of factors contributing to unsafe behavior of coal miners. Saf. Sci.; 2019; 115, pp. 339-348. [DOI: https://dx.doi.org/10.1016/j.ssci.2019.02.003]

4. Yang, L.; Wang, X.; Zhu, J.; Qin, Z. Influencing Factors, Formation Mechanism, and Pre-control Methods of Coal Miners’ Unsafe Behavior: A Systematic Literature Review. Front. Public Health; 2022; 10, 792015. [DOI: https://dx.doi.org/10.3389/fpubh.2022.792015]

5. Kuang, Y.; Tian, S.; Li, G.; Zhao, D.; Qing, T. The relationship between human error and occupational stress of commissioning workers in nuclear power plants: The mediating effects of anxiety. J. Nucl. Sci. Technol.; 2024; 61, pp. 1126-1134. [DOI: https://dx.doi.org/10.1080/00223131.2023.2300313]

6. Faber, L.G.; Maurits, N.M.; Lorist, M.M. Mental fatigue affects visual selective attention. PLoS ONE; 2012; 7, e48073. [DOI: https://dx.doi.org/10.1371/journal.pone.0048073] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23118927]

7. Aagaard, J. Media multitasking, attention, and distraction: A critical discussion. Phenomenol. Cogn. Sci.; 2015; 14, pp. 885-896. [DOI: https://dx.doi.org/10.1007/s11097-014-9375-x]

8. Kim, S.; Kim, J. Influence of light fluctuation on occupant visual perception. Build. Environ.; 2007; 42, pp. 2888-2899. [DOI: https://dx.doi.org/10.1016/j.buildenv.2006.10.033]

9. Chan, M.; Singhal, A. Emotion matters: Implications for distracted driving. Saf. Sci.; 2015; 72, pp. 302-309. [DOI: https://dx.doi.org/10.1016/j.ssci.2014.10.002]

10. Safety, S.A.O.W. Coal Mine Safety Regulations; China Coal Industry Press: Beijing, China, 2016.

11. Tong, R.; Yang, Y.; Ma, X.; Zhang, Y.; Li, S.; Yang, H. Risk Assessment of Miners’ Unsafe Behaviors: A Case Study of Gas Explosion Accidents in Coal Mine, China. Int. J. Environ. Res. Public Health; 2019; 16, 1765. [DOI: https://dx.doi.org/10.3390/ijerph16101765]

12. Sun, L.; Guo, Z.; Yuan, X.; Wang, X.; Su, C.; Jiang, J.; Li, X. An Investigation of the Effects of Brain Fatigue on the Sustained Attention of Intelligent Coal Mine VDT Operators. Int. J. Environ. Res. Public Health; 2022; 19, 11034. [DOI: https://dx.doi.org/10.3390/ijerph191711034]

13. Li, J.; Qin, Y.; Yang, L.; Wang, Z.; Han, K.; Guan, C. A simulation experiment study to examine the effects of noise on miners’ safety behavior in underground coal mines. BMC Public Health; 2021; 21, 324. [DOI: https://dx.doi.org/10.1186/s12889-021-10354-2] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33563245]

14. Moorselaar, D.; Slagter, H.A. Inhibition in selective attention. Ann. Ny. Acad. Sci.; 2020; 1464, pp. 204-221. [DOI: https://dx.doi.org/10.1111/nyas.14304] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31951294]

15. Robinson, P.R. Safety and the Environment. Springer Handbook of Petroleum Technology; Hsu, C.S.; Robinson, P.R. Springer International Publishing: Cham, Switzerland, 2017; pp. 85-147.

16. Paulhus, D.; Aks, D.; Coren, S. Independence of Performance and Self-Report Measures of Distractibility. J. Soc. Psychol.; 1990; 130, pp. 781-787. [DOI: https://dx.doi.org/10.1080/00224545.1990.9924630]

17. Kaida, K.; Takahashi, M.; Åkerstedt, T.; Nakata, A.; Otsuka, Y.; Haratani, T.; Fukasawa, K. Validation of the Karolinska sleepiness scale against performance and EEG variables. Clin. Neurophysiol.; 2006; 117, pp. 1574-1581. [DOI: https://dx.doi.org/10.1016/j.clinph.2006.03.011] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/16679057]

18. Jebelli, H.; Hwang, S.; Lee, S. EEG-based workers’ stress recognition at construction sites. Automat. Constr.; 2018; 93, pp. 315-324. [DOI: https://dx.doi.org/10.1016/j.autcon.2018.05.027]

19. Guo, Z.; Pan, Y.; Zhao, G.; Cao, S.; Zhang, J. Detection of Driver Vigilance Level Using EEG Signals and Driving Contexts. IEEE Trans. Reliab.; 2018; 67, pp. 370-380. [DOI: https://dx.doi.org/10.1109/TR.2017.2778754]

20. Mehmood, I.; Li, H.; Qarout, Y.; Umer, W.; Anwer, S.; Wu, H.; Hussain, M.; Fordjour Antwi-Afari, M. Deep learning-based construction equipment operators’ mental fatigue classification using wearable EEG sensor data. Adv. Eng. Inform.; 2023; 56, 101978. [DOI: https://dx.doi.org/10.1016/j.aei.2023.101978]

21. Rahman, M.M.; Sarkar, A.K.; Hossain, M.A.; Hossain, M.S.; Islam, M.R.; Hossain, M.B.; Quinn, J.M.W.; Moni, M.A. Recognition of human emotions using EEG signals: A review. Comput. Biol. Med.; 2021; 136, 104696. [DOI: https://dx.doi.org/10.1016/j.compbiomed.2021.104696] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34388471]

22. Liao, P.; Zhou, X.; Chong, H.; Hu, Y.; Zhang, D. Exploring construction workers’ brain connectivity during hazard recognition: A cognitive psychology perspective. Int. J. Occup. Saf. Ergon.; 2023; 29, pp. 207-215. [DOI: https://dx.doi.org/10.1080/10803548.2022.2035966] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35098890]

23. So, W.; Wong, S.; Mak, J.N.; Chan, R. An evaluation of mental workload with frontal EEG. PLoS ONE; 2017; 12, e174949. [DOI: https://dx.doi.org/10.1371/journal.pone.0174949]

24. Yang, J.; Liang, N.; Pitts, B.J.; Prakah-Asante, K.; Curry, R.; Yu, D. An Eye-Fixation Related Electroencephalography Technique for Predicting Situation Awareness: Implications for Driver State Monitoring Systems. Hum. Factors; 2024; 66, pp. 2138-2153. [DOI: https://dx.doi.org/10.1177/00187208231204570] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37851849]

25. Ke, J.; Zhang, M.; Luo, X.; Chen, J. Monitoring distraction of construction workers caused by noise using a wearable Electroencephalography (EEG) device. Automat. Constr.; 2021; 125, 103598. [DOI: https://dx.doi.org/10.1016/j.autcon.2021.103598]

26. Kumar, S.P.; Selvaraj, J.; Krishnakumar, R.; Sahayadhas, A. Detecting Distraction in Drivers using Electroencephalogram (EEG) Signals. Proceedings of the 2020 Fourth International Conference on Computing Methodologies and Communication (ICCMC); Erode, India, 11–13 March 2020.

27. Jin, C.Y.; Borst, J.P.; van Vugt, M.K. Predicting task-general mind-wandering with EEG. Cogn. Affect. Behav. Neurosci.; 2019; 19, pp. 1059-1073. [DOI: https://dx.doi.org/10.3758/s13415-019-00707-1] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30850931]

28. Gray, H.M.; Ambady, N.; Lowenthal, W.T.; Deldin, P. P300 as an index of attention to self-relevant stimuli. J. Exp. Soc. Psychol.; 2004; 40, pp. 216-224. [DOI: https://dx.doi.org/10.1016/S0022-1031(03)00092-1]

29. Hao, H.; Conway, A.R.A. The impact of auditory distraction on reading comprehension: An individual differences investigation. Mem. Cogn.; 2022; 50, pp. 852-863. [DOI: https://dx.doi.org/10.3758/s13421-021-01242-6]

30. Chan, R.C.K. A further study on the sustained attention response to task (SART): The effect of age, gender and education. Brain Inj.; 2001; 15, pp. 819-829. [DOI: https://dx.doi.org/10.1080/02699050110034325]

31. Mognon, A.; Jovicich, J.; Bruzzone, L.; Buiatti, M. ADJUST: An automatic EEG artifact detector based on the joint use of spatial and temporal features. Psychophysiology; 2011; 48, pp. 229-240. [DOI: https://dx.doi.org/10.1111/j.1469-8986.2010.01061.x]

32. Barry, R.J.; Clarke, A.R.; Johnstone, S.J.; Mccarthy, R.; Selikowitz, M. Electroencephalogram θ/β Ratio and Arousal in Attention-Deficit/Hyperactivity Disorder: Evidence of Independent Processes. Biol. Psychiat.; 2009; 66, pp. 398-401. [DOI: https://dx.doi.org/10.1016/j.biopsych.2009.04.027] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/19500774]

33. Jap, B.T.; Lal, S.; Fischer, P.; Bekiaris, E. Using EEG spectral components to assess algorithms for detecting fatigue. Expert Syst. Appl.; 2009; 36, pp. 2352-2359. [DOI: https://dx.doi.org/10.1016/j.eswa.2007.12.043]

34. Zhang, Y.; Zhang, M.; Fang, Q. Scoping Review of EEG Studies in Construction Safety. Int. J. Environ. Res. Public Health; 2019; 16, 4146. [DOI: https://dx.doi.org/10.3390/ijerph16214146]

35. Chawla, N.V.; Bowyer, K.W.; Hall, L.O.; Kegelmeyer, W.P. SMOTE: Synthetic minority over-sampling technique. J. Artif. Intell. Res.; 2002; 16, pp. 321-357. [DOI: https://dx.doi.org/10.1613/jair.953]

36. Golmohammadi, R.; Darvishi, E.; Faradmal, J.; Poorolajal, J.; Aliabadi, M. Attention and short-term memory during occupational noise exposure considering task difficulty. Appl. Acoust.; 2020; 158, 107065. [DOI: https://dx.doi.org/10.1016/j.apacoust.2019.107065]

37. Lavie, N. Distracted and confused?: Selective attention under load. Trends Cogn. Sci.; 2005; 9, pp. 75-82. [DOI: https://dx.doi.org/10.1016/j.tics.2004.12.004] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/15668100]

38. Lavie, N. Attention, Distraction, and Cognitive Control Under Load. Curr. Dir. Psychol. Sci.; 2010; 19, pp. 143-148. [DOI: https://dx.doi.org/10.1177/0963721410370295]

39. Awais, M.; Badruddin, N.; Drieberg, M. A Hybrid Approach to Detect Driver Drowsiness Utilizing Physiological Signals to Improve System Performance and Wearability. Sensors; 2017; 17, 1991. [DOI: https://dx.doi.org/10.3390/s17091991] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28858220]

40. Cusinato, R.; Alnes, S.L.; van Maren, E.; Boccalaro, I.; Ledergerber, D.; Adamantidis, A.; Imbach, L.L.; Schindler, K.; Baud, M.O.; Tzovara, A. Intrinsic neural timescales in the temporal lobe support an auditory processing hierarchy. J. Neurosci.; 2023; 43, pp. 3696-3707. [DOI: https://dx.doi.org/10.1523/JNEUROSCI.1941-22.2023] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37045604]

41. Ke, J.; Du, J.; Luo, X. The effect of noise content and level on cognitive performance measured by electroencephalography (EEG). Automat. Constr.; 2021; 130, 103836. [DOI: https://dx.doi.org/10.1016/j.autcon.2021.103836]

42. Klimesch, W. Alpha-band oscillations, attention, and controlled access to stored information. Trends Cogn. Sci.; 2012; 16, pp. 606-617. [DOI: https://dx.doi.org/10.1016/j.tics.2012.10.007]

43. Squire, R.F.; Noudoost, B.; Schafer, R.J.; Moore, T. Prefrontal Contributions to Visual Selective Attention. Annu. Rev. Neurosci.; 2013; 36, pp. 451-466. [DOI: https://dx.doi.org/10.1146/annurev-neuro-062111-150439] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23841841]

44. Wen, W.; Guo, S.; Huang, H.; Li, S. Causal investigation of mid-frontal theta activity in memory guided visual search. Cogn. Neurosci.; 2023; 14, pp. 115-120. [DOI: https://dx.doi.org/10.1080/17588928.2023.2227787]

45. Craig, A.; Tran, Y.; Wijesuriya, N.; Nguyen, H. Regional brain wave activity changes associated with fatigue. Psychophysiology; 2012; 49, pp. 574-582. [DOI: https://dx.doi.org/10.1111/j.1469-8986.2011.01329.x] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/22324302]

46. Wang, Y.; Huang, Y.; Gu, B.; Cao, S.; Fang, D. Identifying mental fatigue of construction workers using EEG and deep learning. Automat. Constr.; 2023; 151, 104887. [DOI: https://dx.doi.org/10.1016/j.autcon.2023.104887]

47. Putman, P.; Verkuil, B.; Arias-Garcia, E.; Pantazi, I.; van Schie, C. EEG theta/beta ratio as a potential biomarker for attentional control and resilience against deleterious effects of stress on attention. Cogn. Affect. Behav. Neurosci.; 2014; 14, pp. 782-791. [DOI: https://dx.doi.org/10.3758/s13415-013-0238-7] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/24379166]

48. Dias Da Silva, M.R.; Gonçalves, Ó.F.; Branco, D.; Postma, M. Revisiting consciousness: Distinguishing between states of conscious focused attention and mind wandering with EEG. Conscious. Cogn.; 2022; 101, 103332. [DOI: https://dx.doi.org/10.1016/j.concog.2022.103332] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35453102]

49. Liu, N.; Chiang, C.; Chu, H. Recognizing the Degree of Human Attention Using EEG Signals from Mobile Sensors. Sensors; 2013; 13, pp. 10273-10286. [DOI: https://dx.doi.org/10.3390/s130810273] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23939584]

50. Wang, Y.; Jung, Y.; Lin, C. EEG-Based Attention Tracking During Distracted Driving. IEEE Trans. Neural Syst. Rehabil. Eng.; 2015; 23, pp. 1085-1094. [DOI: https://dx.doi.org/10.1109/TNSRE.2015.2415520] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25850090]