Effective connectivity of mental fatigue: Dynamic causal modeling of EEG data

Abstract

BACKGROUND:

Recognition of sources in the brain and their interaction with mental fatigue states are interesting subjects for researchers.

OBJECTIVE:

The aim of this study was to investigate the mental fatigue effects on brain areas by dynamic causal modeling (DCM) parameters that are extracted from event-related potential (ERP) signals which were then estimated based on mental fatigue data with visual stimulation.

METHODS:

ERP were recorded based on a Continuous Performance Task in four consecutive trials. Active regions and brain sources were extracted by a Multiple Sparse Priors algorithm.

RESULTS:

Four models are proposed for DCM. The parameters and the structure of the best model were obtained by SPM software for ERP in each of the four trials.

CONCLUSION:

The results illustrate that an increase of mental fatigue through trials leads to increased likelihood of choosing forward models.

Keywords

Dynamic causal modeling effective connectivity active cortical regions mental fatigue

1. Introduction

One of the most noticeable mental activities is mental fatigue, which has recently been extensively studied. When an individual has to pay attention to something carefully for a certain period of time, they are faced with a situation termed mental fatigue [1]. Such fatigue leads to decreasing performance in human beings, lack of managing plans, lack of control and supervision of work; this lack of concentration is evident in all of these situations and is considerable [2]. In the case of mental fatigue, the regional cortex of the brain is inactive and as a result the connections between different parts of the brain are lost. Recognizing these inactive regions leads to detecting mental fatigue [3, 4]. Recognizing the model of mental fatigue leads to improving existing treatment methods due to its significance in our lives. One of the most important models, used as a tool for evaluating attention in patients suffering from nerve damage, was developed by Solhberg and Mater [5]. So far, many studies have been done with the aim of proposing a model which can describe the behavior and activity of the brain during mental fatigue situations (e.g. the connectivity between mental fatigue and the activity of the brain when an error exists, as well as brain anterior cingulate cortex (ACC) activity [6], using coherence calculation of EEG channels and non-linear analysis of EEG signals based on time and frequency features [1]). Moreover, different situations for creating mental fatigue have been investigated, such as during driving [7, 6] and with respect to visual attention [8] and visional fatigue [9]. In the 20 ${}^{\text{th}}$ century, numerous studies were undertaken to provide a better understanding of the impact the brain has on communication between cortical regions and how these regions interact during certain cognitive-sensory-motor activities [10]. Brain connectivity is usually considered to be of three different types: structural, functional, and effective. Structural connectivity is related to anatomical connectivity and also involves the pattern of the links in the brain. Effective and functional connectivity refers to the crossing of signals inside brain regions [11]. Effective connectivity measures the causal connectivity inside the brain and also refers to those parameters of the model that describe the observed dependencies. This type of connectivity is based upon the model that describes the effects. In other words, effective connectivity concerns the effects of the nervous system on other systems, whether at synapse or neural population levels [12]. So far, many computational methods based on some EEG data available for the purpose of learning the structures of effective cortical connectivity have been utilized, which can be categorized into two groups. The first group considers the known structure and model for coupling (e.g. dynamic causal modeling and structural equations modeling). The second one comprises methods that investigate the brain regions without any initial models (e.g. directed transfer function and Granger causality [10]). The main idea behind dynamic causal modeling (DCM) is that brain processing can be considered an algebraic dynamic non-linear system, which is based upon some inputs that create some outputs; it can also be considered a Bayesian overall framework with the purpose of concluding the hidden nervous states by measuring cerebral activity. The main purpose of DCM is the description and estimation of interpretive quantities from a neurobiological point of view (e.g. the power of synaptic coupling among the population of neurons and modeling of concept-dependent populations). Furthermore, it can be considered a reverse solution for the problem of source localization because the process of reversing DCM needs both the recreation of sources and its dynamics [13]. The reverse model comprises two sub models, in a manner that the first sub model is the direct sub model and shows the interaction between the brain regions. The second one simulates the activity of a neuron, which is a dynamic subpopulation and is referred to as the neural mass model [14, 15]. Considering the coupling interactions model and combining this model with the neural mass model leads to finding accurate information relevant to involved sources at the cortical level [15, 2] performed their records in two cases of electrocorticography (ECOG) and EEG with the aim of receiving coupling connections architecture between two sources: frontal and parietal [16] also investigated the improvement in performance of dynamic causal modeling for event-related potential by using a unit processor of a graphics processing unit (GPU). The main idea of DCM for ERP is matching data, as a time series to a position-time model. In the time-position DCM model, the potential of an event is assumed to be a result of connection changes or coupling power in each cortex region in the DCM position time model [9].

Up to now, most research in the field of dynamic causal modeling and effective connectivity in the brain has been based on fMRI. Analyzing DCM based on EEG and ERP is more complicated than cerebral imaging methods such as fMRI [15]. EEG provides a much more realistic and accurate account of the actual interactions between neurons than fMRI. The time resolution of EEG is vastly superior to that of fMRI. Neurons convey information to other neurons by firing in patterns that are meaningful only when viewed with fine time resolution. The EEG signal contains vastly more useful data than fMRI signals, which is why the EEG-based analysis is more complicated. Using EEG allows for realistic modeling of how the neurons interact and convey information to one another. By contrast, fMRI and other methods with excellent spatial resolution but poor time resolution can only provide generalized information about what areas of the brain are interacting with which other areas. Quantifying effective connectivity and the models based on EEG gives a distinct advantage over fMRI and similar methods with respect to the goal of recognizing and investigating causality in neural processes as well as the causes of neurobiological and psychological diseases. In this paper, some EEG and ERP signals are extracted, which are as follows: recognizing active cortical regions during mental fatigue with visional stimulation, recognizing the kind of effective connectivity among them, and finally, presenting a dynamic causal modeling during mental fatigue for these signals, which, to the best of the authors’ knowledge, have not been investigated in previous studies. Although it must be noted that in this study active cortical regions during mental fatigue with visual stimulation and their interaction are specified, the aim of this work was how quantitative values of involved sources could be calculated. The rest of this paper is subdivided as follows: Section 2 describes our database, Section 3 describes the proposed method and its details, and in Section 4 the results of the modeling are outlined, followed by a conclusion of the paper in Section 5.

2. Database

The data employed in the present investigation was hitherto published along with a standard EEG analysis [1]. The participants included twenty male students ranging from 18 to 22 years old. Informed consent was obtained from all individual participants included in the study. The task utilized in the current study was grounded in one type of continuous performance task (CPT) algorithms called “Sustained Attention Dots” from Amsterdam Neuropsychological Tasks (ANT). To experience a long-term attentive task and guarantee that subjects would reach mental fatigue, this algorithm was repeated in four trials. In each of the four trials, subjects were presented with 600 patterns. Each pattern consisted of 3, 4, or 5 white dots on a black background computer screen. Equal number of patterns from each type of pattern (3, 4, or 5 dots) was randomly distributed during each trial. To prevent fixed pattern application and dependence upon memory, the patterns were presented in random configurations. Subjects were required to reply differently when a 4-dot pattern, rather than a 3-dot or 5-dot pattern, was displayed. Each pattern was maximally displayed for 8 seconds or until the subject pressed a keyboard button. The interval between each response and the next pattern stimulus was set to 250 ms. During the recognition memory task, the cranial voltages were collected based upon 10–20 and 24 channel-standard high-input impedance amplifier. Amplified analog voltages (0.1–100 Hz band-pass) were digitized at 200 Hz. Recording started 100 ms before stimulus onset and continued 1,000 ms after that. The EEG was digitally low-pass filtered at 40 Hz. Normalization was performed based on the z-score method. EMG and EOG artifacts were removed through an independent component analysis (ICA) from the original signal [17]. ERP epochs were extracted from the continuous data stream offline, and the EEG signals were referenced to linked ears.

3. Materials and methods

The DCM method for EEG/MEG/LFP is more complicated than DCM for fMRI. This complication arises from time-dependent information that exists in the dynamics of electrophysiological signals which should only be described by models that are able to depict electrophysiological mechanisms in more detail [15]. Using the fundamentals of DCM, EEG signals are fitted by models. To choose the best model, a Bayesian method is applied. DCM mathematical theory and its application in EEG signal analyses are described in the next section.

3.1 Dynamic causal modeling

DCM uses a neural mass model to describe dynamic activities, which are excitatory and also inhibitory among neural subpopulations. In this model, the activity of a source is emulated using three neural subpopulations, each of which is one of the cortical layers. An excitatory subpopulation is present in the granular layer, an inhibitory subpopulation in the supra-granular layer and a population of deep pyramidal cells in the infra-granular layer [18]. A forward model uses extrinsic connections among different sources, matching with the connectivity rules described in [19]. These rules allow the construction of a network of coupled sources linked by extrinsic connections. In this model, bottom-up or forward connections begin in the infra-granular layers and end in the granular layer; top-down or backward connections link in granular layers and lateral connections originated in the infra-granular layers are distributed among all the layers. The DCM is specified in terms of some state equations that summarize the average synaptic dynamics in terms of spike-rate dependent current and voltage changes, for each subpopulation, which is as follows:

$\displaystyle\dot{x}=f(x,u,\theta)$ (1)

$F$ is some nonlinear function describing the neurophysiological impacts on $x$ , which is the evolution of the neuronal state; $\theta$ is the model parameter, and $u$ is the sensory input to the granular layer [14].

3.1.1 Multiple sparse priors method

Crude data on EEGs can be considered a set of signals and/or topography of the skull. In order to retain skull topography, usually specified brain sources are considered. Each of those consists of a group of neurons with membrane potentials fluctuating in a specified period of time. The signal created by the neural group is shown through a bipolar current that increases the difference in electrical potential on the skull; it can also create a measurable magnetic electrical field outside the brain. Active sources are collected through a toolbox called statistical parametric mapping (SPM) [20].

In this method, there is a linear mapping between the instant distributed bipolar inside the brain and the set of signals recorded through electrodes on the head [21], which is shown according to the following equation:

$\displaystyle y=L(\theta)x_{0}+\varepsilon$ (2)

Equation (2) shows the neural state model in data channels, in such a way that $y$ is a set of EEG data and a matrix with dimension of $N_{c}\times N_{n}$ , where $N_{c}$ is number of sensors and $N_{n}$ is sample time. $x_{0}$ , is the activity of sources with dimension of $N_{d}\times N_{n}$ , where $N_{d}$ depicts the bipolar domain. The lead field $L(\theta)$ varies based on the position and direction of sources, whileis an error that may occur during the recording process. Moreover, in Eq. (2) if $N_{c}\leqslant\leqslant N_{n}$ , this equation will be non-invertible and it is recommended to use a Bayesian solution to solve it [22].

3.1.2 Bayesian model selection

Using Bayesian theory and experimental knowledge leads to finding initialized values of probability and as a result the posterior probability on the parameters is derived as follows [15]:

$\displaystyle p(\theta|y)=\frac{p(y|\theta)p(\theta)}{p(y)},p(\theta|y)\alpha p% (\theta|y)p(\theta)$ (3)

3.1.3 Expectation-Maximization

The approximation mentioned in the previous section, Eq. (3), uses varied Bayes that is formally identical to Expectation-Maximization (EM), as described in [23]. The EM can be formulated in analogy to statistical mechanics as a gradient descent on the free energy $F$ , of a system. The aim is to minimize the free energy with respect to a varied density $q(\theta)$ . Minimizing free energy leads to the following equations:

$\displaystyle q(\theta)=p(\theta|y)$

(4) $\displaystyle F-Ln(p(y))$

The free energy $F$ is the negative marginal log-likelihood or negative log-evidence. After convergence and minimization of the free energy, the variable density is used as an approximation for the desired conditional density and the log-evidence is used for model comparison [24, 25].
4. Results

4.1 Active cortical regions

EEG recorded data is created from the separated cortical sources. Each of these sources consists of a group of neurons, the membrane potential of which simultaneously fluctuates during a certain period of time. A signal that is created by this neural group can be shown through a current dipole that increases the difference in electrical potential on the scalp. These regions are extracted based on the MSP method in three steps with three, four, and five dots. The obtained results are shown in Fig. 1 and Table 1.

Figure 1.

Locations of sources during fatigue situation with 3-, 4-, and 5-dot patterns. Sources of activity, modelled as dipoles (estimated posterior moments and locations), are superposed in an MRI of a standard brain in MNI space. The second subject is in three dimensional planes (i.e. sagittal, frontal, and horizontal with MNI coordinates).

Table 1

Prior coordinates per the locations of the equivalent current dipoles in MNI space (mm)

Left occipital (A1)	$-$ 28, $-$ 94, 2
Right occipital (A2)	22, $-$ 92, 4
Left prefrontal (A3)	$-$ 10, 64, 10
Right prefrontal (A4)	14, 62, 10
Left central (A5)	$-$ 50, $-$ 8, 10
Right central (A6)	44, $-$ 4, 12
Motor cortex (A7)	0, 12, 12
Left frontal (A8)	$-$ 26, 44, 10
Right frontal (A9)	38, 46, 10

4.2 Dynamic causal modeling for event-related responses

After the necessary processes and extraction of active regions, the means of connectivity and the level of effectiveness among them were investigated. Four dynamic causal models, differentiated through how they connect various regions, were defined for each step of the attention test for each subject and all three presentations of the pattern, and the models’ parameters and probabilities were approximated. The maximum connectivity model shown in Fig. 2d was not selected as the best model in most cases (95% of subjects), so it was neglected in the comparison between the models. In this step, all models for all steps and participants that were under consideration in 3 kinds of 3-, 4-, and 5-dot states are defined and, moreover, parameters based on the Expectation-Maximization algorithm are estimated and results for all participants in three states are obtained; finally, all of the models for all participants were set based on corresponding data. Ultimately, through choosing a Bayesian model, the most probable model was selected.

Figure 2.

Presented dynamical model. The sources including the network are connected with forward (black), backward (red), or lateral (broken lines) connections as shown. A1, 2: Occipital; A3, 4: Prefrontal; A5, 6: Central; A7: Motor Cortex; A8, 9: Frontal. Four different models were tested within the same architecture (a and d), allowing for learning-related changes in forward F, forward and backward FB, forward, backward and lateral FBL and full forward, backward and lateral FBL connections respectively.

Figure 3.

Bayesian model selection among DCMs for three models, F, FB, FBL. The graphs reveal the free energy approximation to the probability. Probability for models F, FB, and FBL for each subject. The diamond attributed to each subject identifies the best model on the basis of the subject’s highest probability. Labels of each column demonstrate the best model. Simple repetitive tasks in four consecutive trials were tested (a and d) in turn among 19 subjects.

The results of the Bayesian model selection (BMS) are shown in Fig. 3. At the time of describing the 3-dot pattern forward, backward and lateral and the model with forward and backward connectivity in the first step of the test were chosen as providing the best dynamic causal modeling.

From the results of BMS, according to Fig. 3, at the time of describing the 4-dot pattern of the model with forward and backward connectivity in the first step of the test, all three connectivity models in the second step of the test, the forward connectivity model in the third step of the test, and also similar to the third part, the forward connectivity model in the fourth step of the test was chosen as offering the best dynamic causal modeling for most of the participants.

From the results of BMS, according to Fig. 3, at the time of describing the 5-dot pattern of the model with forward, backward and lateral connectivity in the first step of the test, all three connectivity models in the second step of the test, the forward connectivity model and also the model with both forward and backward connectivity in the third step of the test, the model with forward, backward and lateral connectivity in the fourth step of the test was chosen as giving the best dynamic causal modeling in most of the participants. The results show that displaying all three patterns of presentation to participants led to activation of nine regions of the brain. Investigating the obtained results of the winning model at the time of displaying patterns to the participants leads to a decrement in coupling from steps 3 to 4 and sometimes from steps 2 to 3 and 4 in 70% of participants.

In a limited number of participants, the expectations of connectivity from steps 3 to 4 were not obtained in all three models of presentation pattern, which, according to the results of [1], these people at the time of testing understood that the test is finished and mental fatigue happened to them until the third step of the test; moreover, the approach of connectivity up to that point until then was according to expectation. From steps 3 to 4, due to being aware of the final step, the approach of connectivity was inversed and increasing.

5. Conclusion

To the best of the authors’ knowledge, the exact location of active cortical sources during mental fatigue with visional stimulation has not been investigated yet. Moreover, the cortical connections and the effect of sources on each other during this situation have not been investigated either [8]. In this paper, after some necessary preprocessing on mental fatigue data, the active locations of the brain during mental fatigue in all three presentation patterns (i.e. 3, 4 and 5 dot) were obtained for subjects tested using the MSP method in MNI coordinates. The results show that the prefrontal part is related to short- and long-term memory [26] and also that the executive circuits exist in the frontal part [27]. The sensory part of the motor is the place where orders and cortical commands are formed [27]. Previous research provides evidence of the validation of obtained regions in this paper. Furthermore, it can be seen that the sources were different in the three presentation patterns; however, there were some similar sources too, which demonstrates that the regions and the means of connectivity are dependent on the state.

Three dynamic causal models with different connectivity among regions in all steps of the test in three presentation patterns are presented, and the model parameters are calculated through an EM algorithm. Then, the best model for each participant in all four steps and all three presentation patterns was selected with the BMS method. The results show that there is a decrement in connectivity among sources in presentation patterns (3-, 4-, 5-dot) for most participants (70%). In the first step of the test forward and backward connectivity model and also the model with forward, backward and lateral connectivity was chosen as providing the best dynamic causal modeling. In the second step of the test, all three connectivity models and in steps 3 and 4, forward connectivity models were chosen as offering the best dynamic causal modeling. In a limited number of participants, the expectations of connectivity from steps 3 to 4 were not obtained in all three models of presentation pattern, which, according to the results of [1] these people at the time of testing understood that the test is finished and mental fatigue happened to them until the third step of the test. Moreover, the approach of connectivity until then was according to expectation. From steps 3 to 4, due to being aware of the final step, the approach of connectivity was inverse and increasing.

Footnotes

Acknowledgments

Special thanks goes to the Science and Research Branch, Islamic Azad University, Tehran, Iran.

Conflict of interest

The authors have no conflict of interest to report.

References

Azarnoosh

Nasrabadi

Mohammadi

Firoozabadi

. Investigation of mental fatigue through EEG signal processing based on nonlinear analysis: Symbolic dynamics. Chaos, Solitons & Fractals. 2011; 44(12): 1054-62.

Papadopoulou

Friston

Marinazzo

. Estimating directed connectivity from cortical recordings and reconstructed sources. Brain Topography. 2015: 1-12.

Lorist

Boksem

Ridderinkhof

. Impaired cognitive control and reduced cingulate activity during mental fatigue. Cognitive Brain Research. 2005; 24(2): 199-205.

Muzur

Pace-Schott

Hobson

. The prefrontal cortex in sleep. Trends in Cognitive Sciences. 2002; 6(11): 475-81.

Sohlberg

Mateer

. Introduction to cognitive rehabilitation: Theory and practice: Guilford Press; 1989.

Zhao

Liu

Zheng

. Electroencephalogram and electrocardiograph assessment of mental fatigue in a driving simulator. Accident Analysis & Prevention. 2012; 45: 83-90.

Borghini

Vecchiato

Toppi

Astolfi

Maglione

Isabella

, et al., eds. Assessment of mental fatigue during car driving by using high resolution EEG activity and neurophysiologic indices. Engineering in Medicine and Biology Society (EMBC), 2012 Annual International Conference of the IEEE; 2012: IEEE.

Boksem

Meijman

Lorist

. Effects of mental fatigue on attention: an ERP study. Cognitive Brain Research. 2005; 25(1): 107-16.

H-CO

Seo

Kham

Lee

, eds. Measurement of 3D visual fatigue using event-related potential (ERP): 3D oddball paradigm. 3DTV Conference: The True Vision-Capture, Transmission and Display of 3D Video, 2008; 2008: IEEE.

10.

Sameshima

Baccala

. Methods in brain connectivity inference through multivariate time series analysis: CRC press; 2014.

11.

Hosseini

Wang

Bell

Brinton

Simpson

, eds. Is the Software Package Embedding Dynamic Causal Modeling Robust? XIII Mediterranean Conference on Medical and Biological Engineering and Computing 2013; 2014: Springer.

12.

Stone

Baron-Cohen

Knight

. Frontal lobe contributions to theory of mind. Journal of Cognitive Neuroscience. 1998; 10(5): 640-56.

13.

Friston

. Eigenimages and multivariate analyses. Human Brain Function. 1997.

14.

Friston

Harrison

Penny

. Dynamic causal modelling. Neuroimage. 2003; 19(4): 1273-302.

15.

Himpe

. Implementation and Analysis of Dynamic Causal Modeling for EEG MEG Data 2014.

16.

Wang

W-J

Hsieh

I-F

Chen

C-C

. Accelerating computation of DCM for ERP in MATLAB by external function calls to the GPU. PloS One. 2013; 8(6): e66599.

17.

Delorme

Makeig

. EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. Journal of Neuroscience Methods. 2004; 134(1): 9-21.

18.

David

Harrison

Friston

. Modelling event-related responses in the brain. NeuroImage. 2005; 25(3): 756-70.

19.

Felleman

Van Essen

. Distributed hierarchical processing in the primate cerebral cortex. Cerebral Cortex (New York, NY: 1991). 1991; 1(1): 1-47.

20.

Ashburner

Barnes

Chen

Daunizeau

Flandin

Friston

, et al. SPM12 Manual. Wellcome Trust Centre for Neuroimaging, London (UK). 2014.

21.

Dale

Sereno

. Improved localizadon of cortical activity by combining EEG and MEG with MRI cortical surface reconstruction: a linear approach. Journal of Cognitive Neuroscience. 1993; 5(2): 162-76.

22.

Kiebel

David

Friston

. Dynamic causal modelling of evoked responses in EEG/MEG with lead field parameterization. NeuroImage. 2006; 30(4): 1273-84.

23.

Friston

. Bayesian estimation of dynamical systems: an application to fMRI. NeuroImage. 2002; 16(2): 513-30.

24.

Penny

Stephan

Mechelli

Friston

. Comparing dynamic causal models. Neuroimage. 2004; 22(3): 1157-72.

25.

Penny

Friston

Ashburner

Kiebel

Nichols

. Statistical parametric mapping: the analysis of functional brain images: Academic press; 2011.

26.

Kovacs

Fausett

Page

Kim

T-W

Moir

Merriam

, et al. Alzheimer-associated presenilins 1 and 2: neuronal expression in brain and localization to intracellular membranes in mammalian cells. Nature Medicine. 1996; 2(2): 224-9.

27.

Slotnick

Thompson

Kosslyn

. Visual memory and visual mental imagery recruit common control and sensory regions of the brain. Cognitive Neuroscience. 2012; 3(1): 14-20.