Bayesian classifier with multivariate distribution based on D-vine copula model for awake/drowsiness interpretation during power nap
Introduction
Power nap is a short sleep process where the person wakes up before entering the deep sleep [1]. Nap is a supplement to the regular sleep at night. Taking a nap is helpful to repair the damage by sleep deprivation [2]. According to clinical studies, subjects who have a nap have lower coronary mortality than those without napping [3]. Nap is also beneficial for human's long term memory [4], as well as the consolidation of relevant memories [5] etc. In addition, nap is recommended as a good fatigue countermeasure. It makes people quickly recover from the tired state to the normal mental state to maintain a high vigilance level and good working performance [6].
The process of nap may involve the states of being awake and drowsiness. Although the nap is benefit to the mental and physical health, the proper duration is better to be regulated to prevent people from entering the deep sleep [7]. Because waking up from a deep sleep would cause sleep inertia. Sleep inertia is associated with grogginess and disorientation, and people feel more tired than before sleeping [8]. There are several investigation and survey researches on the best sleep timing and length to minimize the sleep inertia effects. However, few have focused on the interpretation of the states from awake to drowsy which are of significance to regulate the nap duration, avoid the fatigue of sleep inertia and maintain the efficient working condition of the body.
The neurophysiological signal contains the characteristic information, which is closely related to the change of the human body state. Physiological states can be diagnosed from neurophysiological signals, e.g. the automatic detection of epilepsy based on EEG (electroencephalography) has the applicant significance for clinical diagnosis [9]. The features extracted from time domain, frequency domain, time-frequency domain of the neurophysiological signals, and the nonlinear dynamic features are feasible for the analysis of physiological states [10]. During the nap, the level of alertness shows a gradual decline process from awake state to the drowsy or sleepy state. The change of the characteristics corresponding to certain frequency activities of EEG and EOG (electrooculography) are typically utilized for analysis [11].
Usually, the extracted features are sent to the classifiers for different interpretation tasks. However, there are complex correlations among the features especially the data from real cases. It would be difficult to obtain desirable performance by sending the features directly to the classifiers. The neurophysiological signals are obtained by placing the electrodes on the scalp. The electrical activities of neurons in the corresponding functional areas are recorded. It is known that the structure made up of millions of neuronal cells is complex. Even for separated recorded positions, there are correlations among the signals at different locations. Therefore, the complex characteristics of neurophysiological signals is worth to be considered to improve the performance of the signal processing or pattern recognition methods for real applications. The methods of feature fusion or feature selection methods can be one of the solution to generate more effective features for classification [12]. On the other hand, with the help of parametric modeling methods, the internal structure of features can be interpreted to improve the classification performance [13]. Comparing the two kinds of solutions, the parametric modeling method is more feasible to make a well relationship between the feature extraction and classification procedures for the neurophysiological signals.
Bayesian classifier is a useful method for various application. The multivariate distribution of parameter is the important prior knowledge for classification. In order to apply the Bayesian classifier for awake/drowsy state interpretation, a well fit of the multivariate distribution is necessary to be developed to deal with the complex characteristics of the neurophysiological signals. The common methods of Bayesian classifier include Gaussian discriminant analysis, naive Bayesian classifier, etc. However, the relationships of multivariable are almost not considered or simplified when implementing those methods. Therefore, how to make accurate estimation of multivariate probability density function for the Bayesian classifier is important to improve its classification performance.
Copula theory is a statistical tool to describe the correlation among random variables. The copula-based model begins with the application in the field of economics [14]. It is also utilized for time series analysis in several fields. Jovanovic et al. proved that copula can capture and visualize the spatial and temporal fluctuations in dependency structures of cardiovascular signals [15]. Ozdemir et al. [16] presented a data-driven classifier fusion method based on the copula statistical theory. Cui et al. [17] introduced a wind power ramp forecast model based on copula theory. Hazarika et al. [18] developed an uncertainty modeling strategy based on multivariate copula. The vine copula is developed based on the theory of copula. Kraus et al. [19] introduced a semi-parametric quantile regression method based on vine copula model. Vernieuwe et al. [20] explored the use of vine copulas in the rainfall field among all storm-related variables. Stübinger et al. [21] developed a statistical arbitrage strategy by using vine copula to interpret the relationships among the random variables. The above presented works showed the application significance and effectiveness of copula and vine copula especially for the data analysis and modeling. Unlike those studies, we focused on the development of classifier which can be adaptive to the real neurophysiological signals for drowsiness detection purpose. The application of parametric model by vine copula for Bayesian classifier is explored and evaluated.
In copula theory, the joint distribution function of multiple random variables is decomposed into the products of several one-dimensional distribution functions and a copula function. The copula function connects several one-dimensional distribution functions to describe the correlation among the random variables. The copula-based Bayesian classification has been realized by some of the authors for drowsiness detection [22]. The constructed multivariate copula model enables the developed Bayesian classifier to obtain better classification performance than the traditional Bayesian classifiers. Due to the structure, the copula function still has limited ability to have a well fit dependency. In terms of optimizing the correlation structure and computational complexity of multi-dimensional variables, the vine copula is better than the traditional multivariate copula. The vine copula builds multi-dimensional dependence structures on the basis of multiple binary copula functions. The realization is more flexible in constructing the dependencies for multivariate copula functions. According to different decomposition strategies, vine copula is mainly divided into C-vine copula and D-vine copula [23]. The D-vine copula is linear structure while the C-vine copula is star structure. The D-vine copula is less complex and more convenient to be optimized than the C-vine copula [24].
In this study, a Bayesian classifier with multivariate distribution is developed based on the D-vine copula model. The multivariate distribution of features is decomposed into independency and dependency products. The marginal probability density of features refers to the independency product that is estimated by the kernel function method and k-nearest-neighbor density estimation method. The pair-copula function is used to construct the correlation among the features, where the parameters are estimated using the maximum likelihood estimation method. The proposed classifier was validated with the actual neurophysiological signals recorded during one's nap after lunch. The obtained results were compared with the Gaussian discriminate analysis, naive Bayesian classifier and supported vector machine model to evaluate the classification performance.
Section snippets
Data acquisition
In total, 8 healthy subjects were participated in this study. All subjects are male with an average age of 22 years old (3 subjects is 22 years old and 5 subjects is 23 years old). Their neurophysiological signals were recorded in the Department of Advanced Control System, Saga University, Japan. Detailed explanation was done for all subjects before recordings, and informed consent was obtained. The ethic statement was approved by Saga University, Saga, Japan. The subject was leaned in a
Results
The recorded sleep data were divided into consecutive 5-s segments for awake/drowsiness interpretation. In the process of daytime short nap for about 20–30 min, the amounts of samples of awake and drowsy state are not balanced. Typically, the awake state is more than the drowsy state. Additionally, the change between awake and drowsiness does not have regular pattern among the subjects. In order to ensure the amount of the balanced training samples, 70% of the segments was randomly selected as
Bayesian classifier based on D-vine copula
Bayesian decision theory has been widely used, such as Gaussian discriminant analysis and naive Bayesian classifier. The basic idea of Bayesian classifier can be considered as a process of inferring posterior information from prior information. In Gaussian discriminant analysis, it is assumed that the features follow multivariate Gaussian distribution. This hypothesis is commonly used since it can approximately simulate the distribution of many kinds of data in practical application and
Conclusion
A Bayesian classifier with a multivariate distribution based on the D-vine copula model was developed. The multivariate distribution is decomposed into the products of the marginal probability density function and pair-copula functions. In the classification process based on the neurophysiological signal, the proposed classifier in this paper performs better than the Gaussian discriminative analysis and naive Bayesian classifier. The advantage of the developed D-vine copula Bayesian classifier
Acknowledgments
This study is supported by the National Natural Science Foundation of China under Grants 61773164, the National Key Research and Development Program of China2017YFB13003002, and the Natural Science Foundation of Shanghai under Grant 16ZR1407500.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (35)
Sleep deprivation due to shift work
Handb. Clin. Neurol.
(2015)- et al.
Sleep inertia
Sleep Med. Rev.
(2000) - et al.
Measuring neurophysiological signals in aircraft pilots and car drivers for the assessment of mental workload, fatigue and drowsiness
Neurosci. Biobehav. Rev.
(2014) - et al.
A new maximum relevance-minimum multicollinearity (MRmMC) method for feature selection and ranking
Pattern Recognit.
(2017) A review of copula models for economic time series
J. Multivar. Anal.
(2012)- et al.
Copula as a dynamic measure of cardiovascular signal interactions
Biomed. Signal Process. Control
(2018) - et al.
D-vine copula based quantile regression
Comput. Stat. Data Anal.
(2017) - et al.
Mixture of D-vine copulas for modeling dependence
Comput. Stat. Data Anal.
(2013) - et al.
Evading the curse of dimensionality in nonparametric density estimation with simplified vine copulas
J. Multivar. Anal.
(2016) - et al.
Hard to wake up? The cerebral correlates of sleep inertia assessed using combined behavioral, EEG and fMRI measures
Neuroimage
(2019)
Miracle Sleep Cure: The Key to a Long Life of Peak Performance
Siesta in healthy adults and coronary mortality in the general population
Arch. Intern. Med.
The long-term memory benefits of a daytime nap compared with cramming
Sleep
The benefit of directed forgetting persists after a daytime nap: the role of spindles and rapid eye movement sleep in the consolidation of relevant memories
Sleep
Effects of two 15-min naps on the subjective sleepiness, fatigue and heart rate variability of night shift nurses
Industry Health
A 30-minute, but not a 10-minute nighttime nap is associated with sleep inertia
Sleep
Classification of EEG signals for detection of epileptic seizures based on wavelets and statistical pattern recognition
Biomed. Eng. Appl. Basis Commun.
Cited by (11)
Clustered Bayesian classification for within-class separation
2022, Expert Systems with ApplicationsCitation Excerpt :Górecki et al. (Górecki et al., 2016) proposed a hierarchical Archimedean copula-based Bayesian classifier and gave a comparison with other classifiers. Qian et al. (Qian et al., 2017) and Wang et al. (Wang et al., 2020) used Copula-based Bayesian classifier for drowsiness detection. Ozdemir et al. (Ozdemir et al., 2018) proposed a classifier which is a fusion of multiple Copula-based Bayesian classifiers.
Study of functional connectivity of central motor system in Parkinson's disease using copula theory
2021, Biomedical Signal Processing and ControlCitation Excerpt :Copula measures arbitrary margin distributions using a rich set of dependency structures and control scale parameters and obtains the dependency power. Copula can accommodate higher-order interaction [18–21]. The copula method has already been used extensively on EEG data [22–24].
Automatic classification methods for detecting drowsiness using wavelet packet transform extracted time-domain features from single-channel EEG signal
2021, Journal of Neuroscience MethodsCitation Excerpt :To achieve this, there are several drowsy detection (DD) methods are available, such as subjective-based (Miley et al., 2016), vehicle-based (Vinckenbosch et al., 2020), behavior-based (Ramzan et al., 2019), and physiological-based (Hu and Lodewijks, 2020). Due to the ease of extracting information from the human body during drowsiness, the physiological parameters based method has proven to be more efficient than other methods for detecting drowsiness (Wang et al., 2020; Halim and Rehan, 2020). The signals produced in the brain are more correlated to the physiological signs of human (Wei et al., 2020; Anuragi and Sisodia, 2020), and are reliable to determine the onset of drowsiness.
State of Health of Liquid Electrolyte Batteries and Its Relationship with the Change in Transmittance at a Frequency of 254 nm
2023, Journal of Sustainable Development of Energy, Water and Environment SystemsA hybrid machine learning framework for e-commerce fraud detection
2022, Model Assisted Statistics and Applications