Skip to main content
PMC home page
Cognitive Neurodynamics logoLink to Cognitive Neurodynamics
. 2018 Apr 16;12(4):431–440. doi: 10.1007/s11571-018-9485-1

Automated detection of driver fatigue based on EEG signals using gradient boosting decision tree model

Jianfeng Hu 1,✉, Jianliang Min 1
PMCID: PMC6048010  PMID: 30137879

Abstract

Driver fatigue is increasingly a contributing factor for traffic accidents, so an effective method to automatically detect driver fatigue is urgently needed. In this study, in order to catch the main characteristics of the EEG signals, four types of entropies (based on the EEG signal of a single channel) were calculated as the feature sets, including sample entropy, fuzzy entropy, approximate entropy and spectral entropy. All feature sets were used as the input of a gradient boosting decision tree (GBDT), a fast and highly accurate boosting ensemble method. The output of GBDT determined whether a driver was in a fatigue state or not based on their EEG signals. Three state-of-the-art classifiers, k-nearest neighbor, support vector machine and neural network were also employed. To assess our method, several experiments including parameter setting and classification performance comparison were performed on 22 subjects. The results indicated that it is possible to use only one EEG channel to detect a driver fatigue state. The average highest recognition rate in this work was up to 94.0%, which could meet the needs of daily applications. Our GBDT-based method may assist in the detection of driver fatigue.

Keywords: Driver fatigue, Electroencephalogram (EEG), Gradient boosted decision tree (GBDT), Entropy

Introduction

A worldwide increase in traffic accidents have resulted in very large casualties with driver fatigue named as one of the most important causes. Driver fatigue is the transitory period between wakefulness and sleep, which can lead to sleep if not interrupted. Driver fatigue has been defined as a state that reduces psychological alertness, which can affect performance in cognitive and psychomotor tasks such as driving. The automatic detection of driver fatigue is a meaningful research field in the driving safety assistance system (Lal and Craig 2001; Saini and Saini 2014). Many methods have been proposed in the past few years, including various vehicle sensor parameters (Sahayadhas et al. 2012), eye state (Lin et al. 2015), video imaging techniques (Jo et al. 2014), and physiological signals. Furthermore, some methods have been specially developed to detect driver fatigue via using electromyograms (EMG) (Wang 2015; Fu and Wang 2014), electrocardiograms (ECG) (Fu and Wang 2014), electrooculograms (EOG) (Ma et al. 2014), and electroencephalograms (EEG) (Correa et al. 2014; Mu et al. 2017; Yin et al. 2017; Wang et al. 2018), where EEG is considered the most common and effective way to identify driver fatigue as the direct reaction of the brain states (Kar et al. 2010). Furthermore, EEG has been widely used in various fields nowadays, especially computer signal recognition (Dietrich and Kanso 2010; Donoghue et al. 2015; Shalbaf et al. 2015; Swapna et al. 2013; Mumtaz et al. 2017).

Various approaches based on EEG signals have been developed to detect driver fatigue. Chai et al. (2016) presented an autoregressive (AR) model for feature extraction and a Bayesian neural network for the classification algorithm, where they achieved a highest accuracy of 88.2%. Nguyen et al. (2017) proposed a new approach for combining EEG, EOG, ECG and NIRS signals to detect driver drowsiness based on fisher’s linear discriminant analysis method and achieved the mean result with 79.2 (± 9.4) of 9 subjects. Li et al. (2012) collected 16 channels of EEG data and computed 12 types of energy parameters, and obtained the highest accuracy of 91.5% from two channels (FP1 and O1). Chai et al. (2017) presented an improvement of classification performance for EEG-based driver fatigue classification from 43 participants by using AR feature extractor and sparse-deep belief networks, obtained an improved classification performance with a sensitivity of 93.9%, a specificity of 92.3%, an accuracy of 93.1%. Xiong et al. (2016) combined features of AE and SE with an SVM classifier and achieved the highest accuracy of 91.3% at channel P3. Fu et al. (2016) proposed a fatigue detection model based on the Hidden Markov Model (HMM), and achieved the highest accuracy of 92.5% based on EEG signals of two channels (O1 and O2) and other physiological signals. Silveira et al. (2016) proposed a method for assessing alertness level based on a single EEG channel (Pz–Oz) by using the normalized Haar discrete wavelet packet transform, found that the second index (γ + β)/(δ + α) achieved the best results. Shen et al. (2007) used random forest (RF) with a heuristic initial feature ranking scheme based on EEG signals from 12 subjects, which yielded the highest accuracy of 87.7%. With respect to driver fatigue detection based on EEG signals, the performance of many linear and non-linear single classifiers has already been assessed; however, it may be difficult to build a better single classifier as EEG signals are unstable and the training set is usually comparatively small. Consequently, such classifiers may have a poor performance or be unstable. Recent studies have shown that ensemble classifiers perform better than single classifiers (Hassan and Subasi 2016; Yang et al. 2016); however, few studies have been conducted on using ensemble classifiers based on EEG signals to study driver fatigue detection. Nevertheless, multichannel EEG acquisition systems such as the 32-channel EEG system used in this study, can only be used in the laboratory. Therefore, an EEG system with less channels, or even one channel, would be more convenient, and more comfortable.

In this study, a complete study of an EEG-based driver fatigue detection system was provided by using GBDT including data acquisition, data processing, feature extraction and classification. The study focused on applying a machine learning method to detect driver fatigue. The rest of this article was organized as follows. In “Methods” section, the gradient boosting decision tree algorithm was described in detail, and feature extraction was presented, including SE, FE, AE and PE. “Results” section showed that the experiments conducted on 22 subjects based on the GBDT method, and was followed by our conclusions in “Discussion” section.

Methods

Subjects

Twenty-two students (8 female, 17–24 years) participated in this experiment. All subjects had a driver’s license and were asked to abstain from alcohol, medications, or tea before and during the experiment. Prior to the experiment, they practiced the driving task for several minutes to become acquainted with the experimental procedures and purposes. This work was approved by the Academic Ethics Committee of Jiangxi University of Technology. The subjects provided their written informed consent as per human research protocol in this study.

Experimental paradigm

A sustained-attention driving task was performed by each subject on a static driving simulator (The ZY-31D car driving simulator, produced by Peking ZhongYu Co., Ltd). This equipment is an analog form of a real driving car, and contains all the driving capabilities of a vehicle. Using computer software technology, different driving environments and conditions can be constructed, such as sunny, foggy, or snowy weather and mountainous, highways, or the countryside. The driving environment selected for this experiment was a highway with low traffic density to more easily induce monotonous driving. Some researches had pointed out that our brains (under this type of driving environment) were more easily turned into a fatigue state and the EEG signals were more stable, which was good for recording data. All subjects in this experiment had approximate real driving experience.

Data recording and preprocessing

In summary, the total duration of the experiment was 40–120 min. The first step was to become familiar with the simulation software, then there was continuous monotonous driving until driver fatigue was determined, and the experiment was terminated.

When the driving lasted 10 min, the last 5 min of EEG signals were recorded as a normal state. When the continuous driving lasted between 30 and 120 min until self-reported fatigue, the questionnaire results showed that the subject was in driving fatigue (obeying the Chalder fatigue scale and Lee’s subjective fatigue scale) (Lee et al. 1991; Borg 1990) so the last 5 min of the EEG signals were labeled as the fatigue state. EOG was also used to analyze eye blink patterns as an objective validation of the fatigue state. It should be noted that the validation of the fatigue condition has also been conducted with a self-reported fatigue questionnaire as per the Chalder fatigue scale and Lee’s subjective fatigue scale in our recent researches (Mu et al. 2017; Yin et al. 2017), and the method of using a questionnaire for identifying fatigue condition has been used in many other studies (Li et al. 2012; Craig et al. 2012; Liu et al. 2010). The drivers were required to complete all tasks and ensure safe driving. Prior to the experiment, the drivers familiarized themselves with the operation of the driving simulator and the completion of the driving tasks.

All channel data were referenced to two electrically linked mastoids at A1 and A2, digitized at 1000 Hz from a 32-channel electrode cap (including 30 effective channels and two reference channels) based on the international 10–20 system and stored in a computer for offline analysis. Eye movements and blinks were monitored by recording the horizontal and vertical EOG.

After collecting EEG signals, the main steps of data preprocessing were carried out by the Scan 4.3 software of Neuroscan (Neuroscan, Compumedics, Australia). First, the raw signals were filtered by a 0.15–45 Hz band-pass filter to remove the noise. Next, 5 min of EEG data from thirty electrodes were sectioned into 1-s epoch, resulting in approximately 300 epochs. With 22 subjects, a total of 6600 units of datasets were formed for the normal state and another 6600 units for the fatigue state. A 10-fold cross validation approach for measuring classification results was used.

Feature extraction

As nonlinear parameters can quantify the complexity of a time series, it can be used to evaluate the non-linear, unstable EEG signals (Azarnoosh et al. 2011; Mateos et al. 2017). Spectral entropy (PE) was calculated by applying the Shannon function to the normalized power spectrum, which was described in detail in the literature (Kannathal et al. 2005). Approximate entropy (AE) was calculated as proposed by Pincus (1991). Like AE, sample entropy (SE) was calculated as proposed by Richman and Moorman (Richman and Moorman 2000). The calculation algorithm of AE and SE were defined clearly in the literature (Song et al. 2012), as was fuzzy entropy (FE) (Xiang et al. 2015).

In the four above-mentioned types of entropies, AE, SE and FE have parameters setting, m and r, which are the dimensions of phase space and similarity tolerance, respectively. Generally, too large an r will lead to a loss of useful information; however, if r is under estimated, the sensitivity to noise will increase significantly. In the present study, m = 2 while r = 0.2 * SD, where SD denotes the standard deviation of the time series, as refer the literature (Yentes et al. 2013).

For optimizing the detection quality, the feature sets were normalized for each subject and each channel by scaling between 0 and 1.

Classification

Since there is no uniform classification method suitable for all subjects and all applications, it is useful to test ensemble classification methods.

To assess the GBDT method for driver fatigue detection, three widely used classifiers (KNN, SVM and NN) were employed as a comparison. KNN is a supervised learning technique where a new instance is classified based on the closest training samples present in the feature space. KNN implements learning based on the k-nearest neighbors of each query point, where k is 10 in this study, if not otherwise specified. In the case of nonlinear classification, kernels (such as radial basis functions (RBF)) are used to map the data into a higher dimensional feature space where a linear separating hyper-plane could be found. When training an SVM classifier with the RBF kernel (Mu et al. 2016), two parameters must be considered: c and γ. A lower c makes the decision surface smooth, while a higher c aims at classifying all training examples correctly. γ defines how much influence a single training example has. In this study, γ = 2, and c = 1. NN is trained via using gradient descent and the gradients calculated by using Backpropagation (BP).

In this work, a recently developed ensemble method named GBDT, originally derived by Friedman (Friedman 2001; Hastie et al. 2009), and was introduced into driver fatigue detection. The GBDT-based method is based on a greedy strategy (called gradient boosting), which is different from random forest which is based on bootstrap aggregating (Bauer and Kohavi 1999). A parameter regularization process can prevent such over-fitting and improve prediction accuracy by optimizing four parameters: number of boosting stages (N), learning rate (L), maximum depth (M), and fraction of samples (F).

Performance metrics

To estimate the potential application performance of a detector, it is very important to properly examine the detection quality. The total average accuracy based on a feature set and some classifiers was the average of the accuracy of all single channel based on the same feature and the same classifier. The classification capabilities of different classifiers were comprehensively investigated with several indexes including Accuracy, Precision, Recall, F1-score, Matthews Correlation Coefficient (MCC), and the Brier score (Sokolova et al. 2006). These indexes are given as follows: Accuracy is the percentage of normal predictions corresponding to all samples; Precision is the percentage of normal predictions corresponding to the normal samples; and Recall is the percentage of fatigue predictions corresponding to the fatigue samples. Furthermore, F1-score was used to appraise both Precision and Recall. The MCC was used as a measure of the quality of binary classifications as it considers true and false positives and negatives, and is generally regarded as a balanced measure which can be used even if the classes are of very different sizes. The Brier score measures the mean squared difference between the predicted probability assigned to the possible outputs and the actual output. Therefore, a lower Brier score value and higher Precision, Recall, F1-score, and MCC values relates to higher performance.

Receiver Operating Characteristic (ROC) analysis is a kind of reliability estimation, and ROC curves regard each detection result as the possible critical value of diagnosis. The area under the ROC curve (AUC) is accepted as a fair indicator of measuring classifier performance since it is invariant to the operating conditions such as different misclassification costs and skewed class distribution (Bradley 1997).

Results

To validate the model performance of different combinations of regularization parameters, a series of GBDT models were built with various boosting stages (N = 50–2000), learning rate (L = 1.0–0.001), maximum depth (M = 1–7), and fraction of samples (F = 0.5–1.0). For the four feature sets, the performances of GBDT models based on different combination of regularization parameters are described in the following figures.

To examine the effectiveness of the GBDT model used for driver fatigue detection, a comparison was conducted with other commonly used classifiers including KNN, SVM, and NN under the same conditions (using the same data and feature sets). Figure 1 shows a comparison between the results for the four feature sets, respectively. In this study, Accuracy, Precision, Recall, F1-score, MCC, and the Brier score were used as the model performance indicators. A comparison of the results of different prediction methods and feature sets indicated that the GBDT model was statistically different to any of other techniques, and received the best model performance out of all four feature sets. This finding further confirmed the advantages of the GBDT model in modeling complex relationships between EEG signals and the fatigue state.

Fig. 1.

Fig. 1

Open in a new tab

Performance of different classifiers and the GBDT method with different number of boosting stages. The left vertical coordinate is for average accuracy, while the horizontal coordinate is for classifier: KNN, SVM, NN, GBDT (with N = 50), GBDT (with N = 100), GBDT (with N = 200), GBDT (with N = 500), GBDT (with N = 1000) and GBDT (with N = 2000), respectively. Where L = 0.1, M = 3, F = 1. GB50, GB100, GB200, GB500, GB1000 and GB2000 represent the GBDT ensemble classifier with 50, 100, 200, 500, 1000 and 2000 boosting stages, respectively

As shown in Fig. 1, the feature set FE outperformed the other three feature sets in all views of the six indexes. The average accuracy using feature set FE achieved 0.919 (when N = 500), which was more than seven percentage points off set PE (accuracy = 0.848).

The influence of the number of boosting stages (N) on model performance can also be seen in Fig. 1. Given other parameters, as N increased, the model obtained a better performance; however, when the value of the shrinkage parameter reached a certain level, the model performance reached equilibrium. Taking feature set FE as an example, the model performance became better (accuracy increasing from 0.850–0.919) as the learning rate parameter value increased from N = 50 to N = 500. However, increasing the value of the learning rate parameter from N = 500 to N = 2000 led to few changes (paired t test, p > 0.05).

The highest average accuracy of a single channel reached 0.920, which was very competitive compared to research described in Introduction. Even for the weak feature set (such as feature set PE), GBDT could significantly improve the performance from 0.541 (using classifier NN) to 0.851 (GBDT1000). Thus, it appears that GBDT could reduce the difference between the different feature sets, for example, the difference in average accuracy between feature set FE and feature set PE based on classifier SVM was 0.247 while the difference in average accuracy between feature set FE and feature set PE based on GBDT was 0.069. This may result from the GBDT model being able to effectively enhance the effect of the weaker feature sets.

The parameter learning rate strongly interacted with the number of boosting stages. Hastie recommended that the learning rate was set to a small constant (e.g., learning rate < = 0.1) (Hastie et al. 2009). The influence of the learning rate parameter on model performance can be seen in Fig. 2. Given the boosting stage (N = 500), an increased value of the learning rate parameter required fewer trees and less computational time to achieve its minimum error. Taking feature set FE as an example, the model performance became better (accuracy increased from 0.816 to 0.908) as the learning rate increased from L = 0.001 to L = 0.1. However, increasing the value of the learning rate parameter from L = 0.1 to L = 1 led to an insignificant result (paired t test, p > 0.05).

Fig. 2.

Fig. 2

Open in a new tab

Performance of GBDT method with different learning rates. The left vertical coordinate is for Accuracy, while the horizontal coordinate is for classifier: GBDT (with L = 0.001), GBDT (with L = 0.005), GBDT (with L = 0.01), GBDT (with L = 0.05), GBDT (with L = 0.1), GBDT (with L = 0.5) and GBDT (with L = 1), respectively. LR0001, LR0005, LR001, LR005, LR01, LR05 and LR1 represent the GBDT ensemble classifier with learning rates of 0.001, 0.005, 0.01, 0.05, 0.1, 0.5 and 1.0, respectively. Where N = 500, M = 3, F = 1

Even for the weak feature set, such as feature set PE, GBDT could significantly improve the performance from 0.582 (LR0001) to 0.851 (LR05). The difference in average accuracy between feature set FE and feature set PE based on LR0001 was 0.234, while the difference in average accuracy between feature set FE and feature set PE based on LR05 was 0.069.

The influence of the maximum depth on model performance can be seen in Fig. 3. For a given boosting stage (N = 500) and learning rate (L = 0.1), as maximum depth changed, the model obtained different performances. Taking feature set FE as an example, the model performance became better (accuracy increased from 0.814 to 0.920) as the learning rate parameter value increased from M = 1 to M = 3. However, an increase in the value of maximum depth from M = 3 to M = 7 led to no significant changes (paired t test, p > 0.05). Even for the weak feature set, such as feature set PE, GBDT could significantly improve performance from 0.826 (md1) to 0.919 (md3). The difference in average accuracy between feature set FE and feature set PE based on Md1 was 0.16, while the difference in average accuracy between feature set FE and feature set PE based on md3 was 0.068.

Fig. 3.

Fig. 3

Open in a new tab

Performance of the GBDT method with different maximum depths. The left vertical coordinate is for average accuracy, while the horizontal coordinate is the classifier: GBDT (with M = 1), GBDT (with M = 2), GBDT (with M = 3), GBDT (with M = 4), GBDT (with M = 5), GBDT (with M = 6) and GBDT (with M = 7), respectively. Md1, Md2, Md3, Md4, Md5, Md6 and Md7 represent the GBDT ensemble classifier with maximum depths of 1, 2, 3, 4, 5, 6 and 7, respectively. Where L = 0.1, N = 500, F = 1

The influence of fraction of samples on model performance can be seen in Fig. 4. For a given boosting stage (N = 500) and learning rate (L = 0.1), as fraction of samples increased, the model obtained a better performance. Taking feature set FE as an example, the model performance became better (accuracy increased from 0.588 to 0.920) as the fraction of samples increased from F = 0.5 to F = 1. Even for the weak feature set, such as feature set PE, GBDT could significantly improve performance from 0.549 (Sub05) to 0.851 (Sub1). Unlike the preceding example, the difference in average accuracy between feature set FE and feature set PE based on Sub05 was 0.038, while the difference in average accuracy between feature set FE and feature set PE based on Sub1 was 0.069.

Fig. 4.

Fig. 4

Open in a new tab

Performance of the GBDT method with a different fraction of samples. The left vertical coordinate is for average accuracy, while the horizontal coordinate is the classifier: GBDT (with F = 0.5), GBDT (with F = 0.6), GBDT (with F = 0.7), GBDT (with F = 0.8), GBDT (with F = 0.85), GBDT (with F = 0.9) and GBDT (with F = 1), respectively. Sub05, Sub06, Sub07, Sub08, Sub085, Sub09 and Sub1 represent the GBDT ensemble classifier with the fraction of samples being 0.5, 0.6, 0.7, 0.8, 0.85, 0.9 and 1.0, respectively, where L = 0.1, M = 3, N = 500

The highest accuracy of a single channel reached 0.940 based on a combination of channel TP7 and feature set FE, or a combination of channel C4 and feature set AE. The highest average accuracy of a single channel reached 0.919 based on feature set FE.

The ROC curve is a plot of true positive rate on the Y axis and false positive rate on the X-axis by varying different threshold ratios. A random performance of a classifier would have a straight line connecting (0, 0) to (1, 1). A ROC curve of the classifier appearing in the upper left triangle suggest that it has a superior performance classification. The results of ten independently rounds were used to draw mean ROC curves. The performance was analyzed by ROC curves and areas under ROC curves (AUC). Figure 5 shows the ROC curve for all subjects based on optimal GBDT classifiers, and the corresponding AUC was 0.946.

Fig. 5.

Fig. 5

Open in a new tab

ROC Curve

Discussion

In this paper, a GBDT-based approach was proposed to detect driver fatigue in an EEG-based system. Results showed that it is a promising system to detect driver fatigue, and achieved a higher success rate with only one channel. With the purpose of providing a more efficient ensemble method for detecting driver fatigue, it was found that: (1) it was possible to use only one electrode for driver fatigue detection, where the highest recognition rate of one electrode could be up to 0.940, which was able to meet the needs of daily applications; and (2) the GBDT method could obviously improve the performance of the detector, especially for the weaker feature sets. This is different from the traditional computational intelligence algorithms (e.g., KNN, SVM, and NN).

But are there significant differences among them? A paired-samples t test was conducted to compare the performance using different classifiers and results show in Table 1. By using the paired-samples t test (with double tail), we can statistically conclude whether the factor has significantly improved the average accuracy or not. It can be seen that there is a significant difference between the majorities of classifiers (p < 0.05). Performance of classifier GB200 and GB500 are significantly better than that of other classifiers, while performance of classifier GB2000 significantly worse than that of other classifiers, as also shown in Fig. 1.

Table 1.

The paired-samples t test analysis between different classifiers

KNN SVM NN GB50 GB200 GB500 GB1000
SVM *
NN * **
GB100 ** * **
GB200 ** ** ** **
GB500 ☆ * ** ** *
GB1000 * ☆ ☆ ☆ * ☆
GB2000 ☆ ☆ * ** ☆ ☆ ☆
Open in a new tab

*p < 0.05; **p < 0.01; ☆p > 0.05

Furthermore, as seen in Table 2, it was found that the classification performance of the proposed method was better than that in previous research using different classification methods based on fewer channels EEG signals. Although the present study is based on the existing EEG data, the high performance of detecting driving fatigue by using a GBDT-based classification indicated that it can be applied well in the real-time detection of driving fatigue.

Table 2.

Studies regarding driver fatigue detection using different methods

Research group Feature method Classifiers EEG channels Average accuracy (%)
Xiong et al. (2016) AE and SE SVM P3 90.0
Correa et al. (2014) Multimodal Analysis NN C3-O1, C4-A1 and O2-A1 83.6
Nguyen et al. (2017) Statistical tests Fisher’s linear discriminant analysis All channels 88.6
Ko et al. (2015) Fast Fourier Transformation The linear regression A single frontal channel 90.0
Chai (Chai et al. 2017) AR Sparse-DBN All channels 93.1
This paper FE GBDT TP7 94.0
Open in a new tab

As different brain characteristics probably exists between different subjects, the EEG features may be different for different subjects; therefore, different subjects using the same feature extraction method or the same classifier may have different performances. It is possible to choose a combination of subject-specific features, which are different from the subjects using a different combination, thus improving the recognition rate of each subject. These subject-specific EEG features can be distinguished from different subjects for identification or authentication of an individual, that is, the EEG password or biometrics (Hu et al. 2015).

Acknowledgements

This work was supported by Supported by National Natural Science Foundation of China (61762045), Natural Science Foundation of Jiangxi Province, China (Nos. 20151BBE50079, 20171BAB202031), Educational Commission of Jiangxi Province, China (Nos. GJJ151146, GJJ161143) and Postdoctoral Assistance Project of Jiangxi Province, China (2017KY33). Thanks ZD Mu and P Wang for collecting and preprocessing EEG data.

Compliance with ethical standards

Conflict of interest

The author declares no conflicts of interest.

Ethical approval

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

References

  1. Azarnoosh M, Nasrabadi AM, Mohammadi MR, Firoozabadi M. Investigation of mental fatigue through EEG signal processing based on nonlinear analysis: symbolic dynamics. Chaos Solitons Fractals. 2011;44(12):1054–1062. doi: 10.1016/j.chaos.2011.08.012. [DOI] [Google Scholar]
  2. Bauer E, Kohavi R. An empirical comparison of voting classification algorithms: Bagging, boosting, and variants. Mach Learn. 1999;36(1–2):105–139. doi: 10.1023/A:1007515423169. [DOI] [Google Scholar]
  3. Borg G. Psychophysical scaling with applications in physical work and the perception of exertion. Scand J Work Environ Health. 1990;16:55–58. doi: 10.5271/sjweh.1815. [DOI] [PubMed] [Google Scholar]
  4. Bradley A. The use of the area under the ROC curve in the evaluation of machine learning algorithms. Pattern Recogn. 1997;30(7):1145–1159. doi: 10.1016/S0031-3203(96)00142-2. [DOI] [Google Scholar]
  5. Chai R, Naik G, Nguyen TN, Ling S, Tran Y, Craig A, Nguyen H. Driver fatigue classification with independent component by entropy rate bound minimization analysis in an EEG-based system. IEEE J of Biomed Health Inf. 2016 doi: 10.1109/JBHI.2016.2532354. [DOI] [PubMed] [Google Scholar]
  6. Chai R, Ling SH, San PP, et al. Improving EEG-based driver fatigue classification using sparse-deep belief networks. Front Neurosci. 2017;11:103–117. doi: 10.3389/fnins.2017.00103. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Correa AG, Orosco L, Laciar E. Automatic detection of drowsiness in EEG records based on multimodal analysis. Med Eng Phys. 2014;36(2):244–249. doi: 10.1016/j.medengphy.2013.07.011. [DOI] [PubMed] [Google Scholar]
  8. Craig A, Tran Y, Wijesuriya N, Nguyen H. Regional brain wave activity changes associated with fatigue. Psychophysiology. 2012;49:574–582. doi: 10.1111/j.1469-8986.2011.01329.x. [DOI] [PubMed] [Google Scholar]
  9. Dietrich A, Kanso R. A review of EEG, ERP, and neuroimaging studies of creativity and insight. Psychol Bull. 2010;136(5):822–848. doi: 10.1037/a0019749. [DOI] [PubMed] [Google Scholar]
  10. Donoghue S, Garcia M, Jordan D, et al. Transition dynamics of EEG-based network microstates during mental arithmetic and resting wakefulness reflects task-related modulations and developmental changes. Cogn Neurodyn. 2015;9(4):371–387. doi: 10.1007/s11571-015-9330-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Friedman JH. Greedy function approximation: a gradient boosting machine. Ann Stat. 2001;29(5):1189–1232. doi: 10.1214/aos/1013203451. [DOI] [Google Scholar]
  12. Fu RR, Wang H. Detection of driver fatigue by using noncontact EMG and ECG signals measurement system. Int J Neural Syst. 2014;24(24):478–491. doi: 10.1142/S0129065714500063. [DOI] [PubMed] [Google Scholar]
  13. Fu RR, Wang H, Zhao WB. Dynamic driver fatigue detection using hidden Markov model in real driving condition. Expert Syst Appl. 2016;63:397–411. doi: 10.1016/j.eswa.2016.06.042. [DOI] [Google Scholar]
  14. Hassan AR, Subasi A. Automatic identification of epileptic seizures from EEG signals using linear programming boosting. Comput Methods Prog Biomed. 2016;136:65–77. doi: 10.1016/j.cmpb.2016.08.013. [DOI] [PubMed] [Google Scholar]
  15. Hastie T, Tibshirani R, Friedman JH. Elements of Statistical Learning. Berlin: Springer; 2009. [Google Scholar]
  16. Hu JF, Mu ZD, Wang P. Multi-feature authentication system based on event evoked electroencephalogram. J Med Imaging Health Inf. 2015;5(4):862–870. doi: 10.1166/jmihi.2015.1471. [DOI] [Google Scholar]
  17. Jo J, Lee SJ, Park KR, Kim IJ, Kim J. Detecting driver drowsiness using feature-level fusion and user-specific classification. Expert Syst Appl. 2014;41(4):1139–1152. doi: 10.1016/j.eswa.2013.07.108. [DOI] [Google Scholar]
  18. Kannathal N, Choo ML, Acharya UR, Sadasivan P. Entropies for detection of epilepsy in EEG. Comput Methods Prog Biomed. 2005;80(3):187–194. doi: 10.1016/j.cmpb.2005.06.012. [DOI] [PubMed] [Google Scholar]
  19. Kar S, Bhagat M, Routray A. EEG signal analysis for the assessment and quantification of driver’s fatigue. Transp Res F. 2010;13(5):297–306. doi: 10.1016/j.trf.2010.06.006. [DOI] [Google Scholar]
  20. Ko LW, Lai WK, Liang WG, Chuang CH, Lu SW, Lu YC, et al. Single channel wireless EEG device for real-time fatigue level detection. Int Jt Conf Neural Netw. 2015 [Google Scholar]
  21. Lal SKL, Craig A. A critical review of the psychophysiology of driver fatigue. Biol Psychol. 2001;55:173–194. doi: 10.1016/S0301-0511(00)00085-5. [DOI] [PubMed] [Google Scholar]
  22. Lee KA, Hicks G, Nino-Murcia G. Validity and reliability of a scale to assess fatigue. Psychiatry Res. 1991;36(3):291–298. doi: 10.1016/0165-1781(91)90027-M. [DOI] [PubMed] [Google Scholar]
  23. Li W, He QC, Fan XM, Fei ZM. Evaluation of driver fatigue on two channels of EEG data. Neurosci Lett. 2012;506(2):235–239. doi: 10.1016/j.neulet.2011.11.014. [DOI] [PubMed] [Google Scholar]
  24. Lin LZ, Huang C, Ni XP, Wang JW, Zhang H, Li X, Qian ZQ. Driver fatigue detection based on eye state. Technol Health Care. 2015;23:S453–S463. doi: 10.3233/THC-150982. [DOI] [PubMed] [Google Scholar]
  25. Liu J, Zhang C, Zheng C. EEG-based estimation of mental fatigue by using KPCA–HMM and complexity parameters. Biomed Signal Process Control. 2010;5:124–130. doi: 10.1016/j.bspc.2010.01.001. [DOI] [Google Scholar]
  26. Ma JX, Shi LC, Lu BL. An EOG-based vigilance estimation method applied for driver fatigue detection. Neurosci Biomed Eng. 2014;2(1):41–51. doi: 10.2174/2213385202666141218104855. [DOI] [Google Scholar]
  27. Mateos DM, Erra RG, Wennberg R, et al. Measures of entropy and complexity in altered states of consciousness. Cogn Neurodyn. 2017;1:1–12. doi: 10.1007/s11571-017-9459-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Mu ZD, Hu JF, Min JL. EEG-based person authentication using a fuzzy entropy-related approach with two electrodes. Entropy. 2016;18:432. doi: 10.3390/e18120432. [DOI] [Google Scholar]
  29. Mu ZD, Hu JF, Yin JH. Driving fatigue detecting based on EEG signals of forehead area. Int J Pattern Recognit Artif Intell. 2017 [Google Scholar]
  30. Mumtaz W, Vuong PL, Malik AS, et al. A review on EEG-based methods for screening and diagnosing alcohol use disorder. Cogn Neurodyn. 2017;3:1–16. doi: 10.1007/s11571-017-9465-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Nguyen T, Ahn S, Jang H, et al. Utilization of a combined EEG/NIRS system to predict driver drowsiness. Sci Rep. 2017;7:43933. doi: 10.1038/srep43933. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Pincus SM. Approximate entropy as a measure of system complexity. Proc Natl Acad Sci USA. 1991;88(6):2297–2301. doi: 10.1073/pnas.88.6.2297. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Richman JS, Moorman JR. Physiological time-series analysis using approximate entropy and sample entropy. Am J Physiol Heart Circ Physiol. 2000;278(6):H2039–H2049. doi: 10.1152/ajpheart.2000.278.6.H2039. [DOI] [PubMed] [Google Scholar]
  34. Sahayadhas A, Sundaraj K, Murugappan M. Detecting driver drowsiness based on sensors: a review. Sensors. 2012;12(12):16937–16953. doi: 10.3390/s121216937. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Saini V, Saini R. Driver drowsiness detection system and techniques: a review. Comput Sci Inf Technol. 2014;3:4245–4249. [Google Scholar]
  36. Shalbaf R, Behnam H, Jelveh MH. Monitoring depth of anesthesia using combination of EEG measure and hemodynamic variables. Cogn Neurodyn. 2015;9(1):41–51. doi: 10.1007/s11571-014-9295-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Shen KQ, Ong CJ, Li XP, Hui Z. A feature selection method for multilevel mental fatigue EEG classification. IEEE Trans Biomed Eng. 2007;54(7):1231–1237. doi: 10.1109/TBME.2007.890733. [DOI] [PubMed] [Google Scholar]
  38. Silveira TLTD, Kozakevicius AJ, Rodrigues CR. Automated drowsiness detection through wavelet packet analysis of a single EEG channel. Expert Syst Appl. 2016;55:559–565. doi: 10.1016/j.eswa.2016.02.041. [DOI] [Google Scholar]
  39. Sokolova M, Japkowicz N, Szpakowicz S (2006). Beyond accuracy, F-score and ROC: a family of discriminant measures for performance evaluation. In: AI2006: advances in artificial intelligence. Springer, pp 1015–1021
  40. Song Y, Crowcroft J, Zhang J. Automatic epileptic seizure detection in EEGs based on optimized sample entropy and extreme learning machine. J Neurosci Methods. 2012;210(2):132–146. doi: 10.1016/j.jneumeth.2012.07.003. [DOI] [PubMed] [Google Scholar]
  41. Swapna G, Swapna G, Swapna G, et al. Automated EEG analysis of epilepsy: a review. Knowl Based Syst. 2013;45(3):147–165. [Google Scholar]
  42. Wang H (2015). Detection and alleviation of driving fatigue based on EMG and EMS/EEG using wearable sensor. In: Proceedings of the 5th EAI international conference on wireless mobile communication and healthcare, pp 155–157
  43. Wang H, Dragomir A, Abbasi NI, et al. A novel real-time driving fatigue detection system based on wireless dry EEG. Cogn Neurodyn. 2018;8:1–12. doi: 10.1007/s11571-018-9481-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Xiang J, Li CG, Li HF, Cao R, Han XH, Chen JJ. The detection of epileptic seizure signals based on fuzzy entropy. J Neurosci Methods. 2015;243:18–25. doi: 10.1016/j.jneumeth.2015.01.015. [DOI] [PubMed] [Google Scholar]
  45. Xiong Y, Gao J, Yang Y, Yu X, Huang W. Classifying driving fatigue based on combined entropy measure using EEG signals. Int J Control Autom. 2016;9(3):329–338. doi: 10.14257/ijca.2016.9.3.30. [DOI] [Google Scholar]
  46. Yang T, Chen WT, Cao GT. Automated classification of neonatal amplitude-integrated EEG based on gradient boosting method. Biomed Signal Process Control. 2016;28:50–57. doi: 10.1016/j.bspc.2016.04.004. [DOI] [Google Scholar]
  47. Yentes JM, Hunt N, Schmid KK, Kaipust JP, Mcgrath D, Stergiou N. The appropriate use of approximate entropy and sample entropy with short data sets. Ann Biomed Eng. 2013;41(2):349–365. doi: 10.1007/s10439-012-0668-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. Yin JH, Hu JF, Mu ZD. Developing and evaluating a mobile driver fatigue detection network based on electroencephalograph signals. Healthc Technol Lett. 2017 doi: 10.1049/htl.2016.0053. [DOI] [PMC free article] [PubMed] [Google Scholar]

Articles from Cognitive Neurodynamics are provided here courtesy of Springer Science+Business Media B.V.

RESOURCES