Self-Regulation Phenomenon Emerged During Prolonged Fatigue Driving: An EEG Connectivity Study

Driving fatigue is a common experience for most drivers and can reduce human cognition and judgment abilities. Previous studies have exhibited a phenomenon of the non-monotonically varying indicators (both behavioral and neurophysiological) for driving fatigue evaluation but paid little attention to this phenomenon. Herein, we propose a hypothesis that the non-monotonically varying phenomenon is caused by the self-regulation of brain activity, which is defined as the fatigue self-regulation (FSR) phenomenon. In this study, a 90-min simulated driving task was performed on 26 healthy university students. EEG data and reaction time (RT) were synchronously recorded during the whole task. To identify the FSR phenomenon, a data-driven criterion was proposed based on clustering analysis of individual behavioral data and the FSR group was determined as having non-monotonic increase trend of RT and the drops of RT during prolonged driving were more than two levels among the total five levels. The subjects were then divided into two groups: the FSR group and the non-FSR group. Quantitative comparative analysis showed significant differences in behavioral performance, functional connectivity, network characteristics, and classification performance between the FSR and non-FSR groups. Specifically, the behavioral performance exhibited apparent non-monotonic development trend: increasing-decreasing-increasing. Moreover, network characteristics presented similar self-regulated development trends. Our study provides a new insight for revealing the complex neural mechanisms of driving fatigue, which may promote the development of practical techniques for automatic detection method and mitigation strategy.

of RT and the drops of RT during prolonged driving were more than two levels among the total five levels.The subjects were then divided into two groups: the FSR group and the non-FSR group.Quantitative comparative analysis showed significant differences in behavioral performance, functional connectivity, network characteristics, and classification performance between the FSR and non-FSR groups.Specifically, the behavioral performance exhibited apparent non-monotonic development trend: increasingdecreasing-increasing.Moreover, network characteristics presented similar self-regulated development trends.Our study provides a new insight for revealing the complex neural mechanisms of driving fatigue, which may promote the development of practical techniques for automatic detection method and mitigation strategy.Index Terms-EEG, fatigue self-regulation (FSR), brain network, machine learning.

I. INTRODUCTION
D RIVING fatigue can be induced by long-time driving tasks, which can lead to a functional decline of the driver's decision-making and control ability, thereby increasing the risk of traffic accidents.Investigation indicates that 21% of global annual traffic accidents are caused by driving fatigue [1].In fact, driving fatigue has become one of the main causes of major traffic accidents.According to a report from Canada, more than 50% of drivers, e.g., nearly 20 million Canadians, were driving under fatigue, and 20% were driving in semisleep [2].It has been reported that 16.5% of fatal traffic accidents have an obvious causal relationship with driving fatigue in the United States [3].On a specific highway in China, this ratio reaches a staggering 40%, according to the China Expressway Network.Consequently, researchers have been continuously exploring detection methods and neural mechanisms of driving fatigue to mitigate its detrimental effects.
Due to its unique advantages of high temporal resolution and portability [4], [5], EEG has been widely adopted in recent researches for driving fatigue detection.Of note, two broad research directions of practicability and accuracy were repeatedly-emphasized in these studies.For instance, instead of using EEG data from whole brain, Wei et al., proposed a non-hair-bearing (NHB, containing Fp1, Fp2, F7, and F8 EEG channels) strategy for fatigue detection [6] and achieved an accuracy of 80%.More recently, we provided further evidence to support the effectiveness of the NHB strategy [7] and optimized the classification method to obtain an accuracy of 90.12% by fusing multiple features (including power spectrum density, functional connectivity (FC), and entropy) from only three EEG channels (i.e., AFp3h, AFpz, and AFp4h).On the other hand, researches improve the accuracy of driving fatigue classification/detection method by developing new feature extraction algorithms and classification models have been emerged [8], [9].For instance, Gao and colleagues constructed a multiplex recurrence network (RN) to fuse the original EEG signals and used a convolutional neural network (CNN) to extract the features of multi-channel RN [10].The proposed RN-CNN method achieved an accuracy of 92.95%.The same team later introduced an EEG-based spatio-temporal CNN to automatically learn effective EEG features and further improved the classification accuracy to 97.37% [11].Most recently, we have proposed an CNN-Attention framework that could concurrently perform dual-task of driving fatigue detection and authentication, which may have potential in autonomous driving [12].In sum, significant progress has been made in practical and accurate driving fatigue detection, the understanding of its underlying neural mechanisms however remains relatively limited.
In recent years, research on the neural mechanism of driving fatigue mainly focused on revealing fatigue-related EEG biomarkers through a two-way (vigilant vs. fatigued) classification approach [13].Only until recently, attempts have been made to show the complex characteristics of EEG measures with the increase of time-on-task [14].It is noteworthy mentioning that recent studies showed non-monotonic development trends in various fatigue evaluation indexes (i.e., quantitative subject assessment [15], reaction time [13], [16], EEG-based energy index [17], the weight of FC [18], and brain functional network characteristics [4], [19]), reiterating the complex mechanisms underlying driving fatigue.For instance, a non-linear fashion of behavioral metrics (i.e., accuracy and reaction time (RT)) has been repeatedly reported by different research groups using divergent experimental protocols which is accompanied by substantial inter-individual differences [18], [20].Moreover, in our recent study of net-work reorganization during mental fatigue, we reported nonmonotonic development of network metrics in both cognitive fatigue and driving fatigue conditions [4].According to [19], such non-monotonic phenomenon was interpreted as the adaptive regulation of physiological rhythm during mental fatigue.Specifically, this non-monotonic variation contains additional information suggesting that the fatigue evolution process is complex and not solely a fatigue accumulation process.
In this study, the phenomenon of non-monotonic development trend is defined as the fatigue self-regulation (FSR).To our knowledge, no research to date has investigated the neural mechanism during FSR in driving fatigue.The reasons may be as follows: first, instead of analyzing the data from the entire fatigue process, most studies employed a two-class classification approach (i.e., pre-vs.post-task [4], or vigilant vs. fatigued [21], [22]).Also the fatigue evolution process exhibits significant individual differences [23] (including variations in the timing and duration of the FSR among individuals), the widely-adopted group-level statistical/classification analysis may overlook the patterns associated with the FSR.Second, in the studies where cognitive tasks were employed for fatigue inducing (i.e., sustained attention task [24]), a typical duration would be around 30 minutes which may not be long enough to trigger the FSR.
Of note, the FSR phenomenon and fatigue recovery may represent different neural mechanisms.Heuristically, fatigue recovery is the recovery of mental and/or physical resources induced by natural rest or specific activities during the break [25].For instance, Blasche et al., have reported that natural rest can reduce fatigue and increase vitality, while adding specific activities during rest (i.e., participating in sports activities or relaxation exercise) can improve performances in tasks with high attention requirements [26].Other approaches, including listening to relaxing music during rest [27], [28] and use of low-frequency electrical pulses to stimulate acupoints [29], [30], also showed the effectiveness in relieving fatigue.Unlike the abovementioned fatigue recovery process where a typical block effect was administrated via various approaches, the FSR phenomenon shows the self-regulation during the primary task performing (i.e., long-term continuous driving in this work without break).In fact, insights of FSR can be gained from fatigue studies of individual differences.In our previous studies, substantial inter-individual differences in behavioral impairment were revealed [7].More importantly, we showed cognitive resilient participants whose performance declines less rapidly as a function of time-on-task [31], [32], indicating a trait-like predisposition to mental fatigue vulnerability [7], [33].Taken into the above into consideration, we posited that the FSR phenomenon may contribute to such fatigue resilience.
In the current work, we set out to quantitatively investigate the complex neural mechanisms of driving fatigue in terms of adaptive fatigue self-regulation.Specifically, using a long-term simulated driving paradigm, we proposed a data-driven quantitative approach to determine the FSR through clustering analysis of the individual behavioral data.Then, subjects within this study were categorized into two groups: the FSR group and the non-FSR group.Moreover, in the FSR group we further defined four states (namely, vigilant state (VS, corresponding to the vigilant state), fatigue state (FS, corresponding to the fatigue state prior to the FSR), FSR state (FSRS, corresponding to the FSR state showing apparent self-regulated behavioral performance improvement), and fatigue deepening state (FDS, corresponding to the most fatigued state)) based upon the behavioral measures pattern.Of note, in the non-FSR group, only two states (i.e., VS and FDS) would be appear due to the monotonic variation of the behavioral measures pattern.We conducted a comprehensive analysis of the differences in behavioral and FC between these two groups throughout the fatigue evolution process.

II. METHODS AND MATERIALS A. Subjects
In this study, 26 healthy university students (Male/Female = 16/10) from the National University of Singapore were enrolled to participate in a driving simulation experiment.They were 23.2 ± 2.8 years old.None of the subjects had previously participated in a similar test.All subjects were right-handed with normal or corrected-to-normal vision, and had never participated in a simulated driving experiments.Participants diagnosed with psychiatric disorder or sleep disorders (such as epilepsy, schizophrenia or chronic insomnia, etc) were excluded.Participants were prohibited from consuming alcohol, caffeine, and strong tea for 2 hours before the experiment.The scalp and earlobes were washed 2 hours before the experiment, and no gel, wax, or conditioner was used to avoid distortion of the EEG waveform due to excessive scalp resistance.This experiment was conducted in the Cognitive Engineering Laboratory at the Singapore Institute of Neurotechnology (SINAPSE) from 1:00 pm to 5:00 pm.This study was approved by the Institutional Review Board of the National University of Singapore, and all subjects signed an informed consent form before the study.

B. Experimental Design
The experiment setup was composed of simulated driving equipment and wireless dry electrode EEG acquisition equipment.Subjects were instructed to perform a 90-minute simulated driving task at a pre-set constant speed of 80 km/h using simulation software and a racing wheel (Model: Logitech G27).The driving road environment was displayed by three 65-inch LCD screens and obey to the left driving rules according to Singapore standards.During the driving task, a guide car was positioned in front of the participant's car.The guide car would illuminate the brake light at a certain random time.The subjects were asked to step on the brake as soon as possible to maintain a safe distance (around 15 m) without crash.The time interval, from the time the brake light of the pilot car came on until the participant stepped on the brake, was recorded as the reaction time (RT).Here, the RT values of less than 0.2 seconds or greater than 2 seconds were considered as outliers and would be removed for the following analyses.Each subject was required to drive as stably as possible.The traffic conditions consisted of a pristine rural straight two-lane road with minimal vehicle density, intentionally designed to induce a sense of monotony and reduced challenge among the participants.The test was carried out in the afternoon between 1 and 5 pm to reduce the potential interference from time-of-day variations, and set in a quiet environment that was set up to simulate real road conditions as much as possible.For more experimental design details, please refer to our previous work [34], [35].

C. Data Recording and Preprocessing
During the driving simulation experiment, a 24-channel (including AFF3, AFF4, FFC5, FFC3h, FFCZ, FFC4h, FFC6h, CCP5h, CCP1, CCPZ, CCP2, CCP6h, PO7, PO3, POZ, PO4, PO8, O1h, OZ, and O2h) wireless dry electrode cap (model: HD-72, Cognionics Inc., USA) was used to collect EEG data.All electrodes were placed in accordance with the 10-20 international standard system.The reference electrode was left and right mastoid.The electrode impedance was controlled below 20 k and the sampling frequency was set as 250 Hz.The horizontal and vertical electrooculograms (EOG) were recorded for artifact removal.A previously-validated standard EEG preprocessing process was adopted for raw EEG signals.Specifically, a second-order Butterworth band-pass filter (1 -40 Hz) was used to remove the low-frequency baseline drift and 50 Hz power frequency interference.Then, the method of average reference electrode was reset for the filtered EEG data.The independent component analysis (ICA) was applied to remove the high correlation artifacts with EOG.Subsequently, a band-pass filter was used to extract four standard bands: θ (4 -8 Hz), α (8 -13 Hz), β (13 -30 Hz), and γ (30 -40 Hz).All data preprocessing was performed using customized codes in Matlab 2019b (The MathWorks Inc., U.S.) and EEGLAB [36].

D. Determination of the FSR Process
The evolution process of driving fatigue exhibited notable variations among individuals.In Fig. 1, we showed two typical driving fatigue evolution processes of the individual behavioral data.In order to determine the FSR process in a data-driven fashion, we provided a quantitative framework based upon the clustering-analysis of individual behavioral performance.Specifically, a 5-min sliding window with 50% overlap was administered to the 90-min behavioral data, the average RT within each 5-min window was then estimated, leading to 35 RTs for each subject.A normalization approach was then employed to normalize the RTs between 0 and 1 (corresponding to the maximum RT) for each participant.Then a K-means clustering model with Euclidean distance was adopted on the RT vector of each subject to reach an optimal clustering of K.If the K levels of RT exhibited a non-monotonic increase trend and the drops of RT were more than two levels (from level 5 to level 2 in Fig. 1(a)), the subject was determined as having the FSR process and would be divided into FSR group.Otherwise, if the K levels of RT exhibited a monotonic increase trend (Fig. 1(b)), then the subject was determined as having no The experimental setup and protocol where a 90-min prolonged driving task was administered.Behavioral measures (i.e., reaction time, RT) were estimated as mean RT within a 5-min window.Through sliding the window with 50% overlap over the entire course of driving, we obtained the development trend of RT for each subject.Of note, a normalization approach was adopted to normalize the RTs between 0 and 1 (corresponding to the maximum RT) for each participant.(b) The development trends of the normalized RT for 2 representative subjects with the FSR phenomenon (left panel) and without the FSR phenomenon (right panel).The RT vector was then subjected to a K-means clustering analysis where the RTs were clustered into 5 levels.Subject with the FSR phenomenon was determined as having RT drops ≥ 2 levels.With the FSR group, four states (VS, FS, FSRS, and FDS) that corresponding to different fatigue processes were further introduced.In the non-FSR group, a monotonic development trend of RT was revealed and two states (VS and FDS) that corresponding to the most vigilant and fatigued state were also illustrated.
FSR process and would be divided into non-FSR group.In order to assess the generalizability of the FSR determination process, we have performed clustering analyses with different number of clusters (K = 4, 5, 6) and found the FSR process intact (supplementary materials).Considering the inputs of 35 RTs for the clustering analysis, K = 5 was employed in the current work to avoid too narrow dynamic range of RTs at K = 4 or small sample size per clusters at K = 6.In additional to the FSR process determination, we further defined four states based upon the RT clustering pattern, i.e., vigilant state (VS, corresponding to the most vigilant state), fatigue state (FS, corresponding to the fatigued state prior to the FSR process), FSR state (FSRS, corresponding to the regulation state), and fatigue deepening state (FDS, corresponding to the most fatigued state).Of note, since the non-FSR group did not have the FSR process, subjects within the non-FSR group only have two states (i.e., VS and FDS).

E. FC Computation
In order to reveal the complex neural mechanisms underlying the FSR process, EEG data of the abovementioned four states were extracted from participants in the FSR group and EEG data of the VS and FDS was also extracted from participants in the non-FSR group.Then each 5-min EEG data was divided into 150 epochs of 4-sec time window with 50% overlap.Within each window, the phase lag index (PLI) [37] was estimated to assess the FC of all pairs of EEG channels for its superiority in minimising the influence of common source and volume conduction.
For a given pair of two EEG time series s k (t) and s l (t), the instantaneous phase is calculated using the Hilbert transform: where z k and z l are instantaneous amplitude, φ k (t) and φ l (t) are the instantaneous phases at t moment, sk (t) is the Hilbert transform of each time series.The phase difference between two EEG time series is expressed as: then, PLI can be defined as: where | • | means the absolute value, and ⟨•⟩ denotes the mean value and sign stands for the signum function: The value of PLI is in the range of [0, 1].A large value indicates a strong degree of phase synchronization between the two channels of EEG signals, e.g., PLI = 1 means that these two EEG signals are in complete phase synchronization.After the functional connections of all pairs of EEG channels were estimated for an epoch using PLI, a 24 × 24 weighted adjacency matrix (of note, 24 is corresponding to the number of EEG channels) could be obtained for each frequency band.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
Of note, we obtained four matrices that were corresponding to the VS, FS, FSRS and FDS for each subject within the FSR group whereas for the subject within the non-FSR group, we obtained two matrices corresponding to the VS and FDS at each frequency band for the following analyses.

F. Brain Network Analysis
Once we have obtained the brain network, graph theoretical analysis was conducted to quantitatively assess the driving fatigue and FSR-related brain network reorganization.Prior to the graph theoretical analysis, a previously-validated sparsity approach (e.g., the ratio of the current number of connections to the total number of all possible connections) was employed.The sparsity approach normalized all resultant networks to have the same number of nodes and edges, which exclude the suspicious FCs and minimize the effect of possible discrepancies in the overall FC weight between groups/states, therefore enabling us to explore the intrinsic differences in the topological structure of the brain networks.Here, instead of arbitrarily selecting a single threshold, we thresholded each brain network over a wide range of sparsity level (i.e., 20% ∼ 40% with a step of 1%) with the following two criteria: 1) ensure the network reachability while including the strong FCs and 2) allow for prominent small-world properties.The weighted network measures to assess global integration (e.g., characteristic path length, L w , and global efficiency, E w global ) and local segregation (e.g., clustering coefficient, C w , and local efficiency, E w local ) were estimated for each step of sparsity using the brain connectivity toolbox [38].To avoid multiple comparisons at individual sparsity threshold and to reduce the dependency of any significant difference sin network topology on the arbitrary choice of a single threshold, an integrated network metric was estimated for all network metrics over the predefined sparsity range.The following statistical comparisons were conducted on the integrated network metrics.
For a weighted network G w with N nodes, the network measures are defined as follows: 1) Characteristic Path Length (L w ): L w means the shortest path length between different nodes: where l w i j represents the weighted path length between node i and j.Path length of an edge conceptualized to weight networks is defined as the reciprocal of the edge weight (l w i j = 1/w i j ), and min{l w i j } is the shortest path length between node i and j.L w is a key parameter to depict the speed of information transmission in a brain network, e.g., a shorter L w indicates increased communication between nodes.
2) Clustering Coefficient (C w ): C w reflects the possibility that all neighbor nodes of a node in the network are neighbor to each other and it could be estimated as: where k i is the degree of node i in the binary network (only count the FC linking the node i without considering the edge weight), deg i is the weighted degree of node i in the weighted network (the sum of edge weights of FCs linking the node i), w i j is the edge weight of FC linking node i and j.The a denotes the connection between two nodes (if the connection is existed, a = 1; otherwise, a = 0).C w measures the extent of a local density or cliquishness of a network.
3) Global Efficiency (E w global ): E w global is a comprehensive index to measure the efficiency of information transmission on a network: E w local measures the information exchange at the clustering level.The concurrent higher E w global and E w local (i) indicate a system with high balance between local fault tolerance and wide-scope interactions.

G. Classification Models
In additional to the quantitative graph theoretical analysis, a data-driven classification analysis framework was also performed.Specifically, two classification schemes were performed, including 1) state classification between two common states (VS and FDS) in each group to assess the fatigue effect; and 2) subject classification between both groups (non-FSR and FSR) in the two common states to assess the intrinsic individual differences.Here, three widely-used conventional classification algorithms (including k-nearest neighbor (KNN), support vector machine (SVM), and BP neural network based on Adaboost algorithm (BP_Adaboost)) were employed for the classification to reduce the possible bias caused by the performance of a single model.The inputs for the classification models were group/state labels and significantly-altered PLI features determined via statistical comparisons.Detailed information about the shape of PLI input features for the classification models were listed in supplementary materials.A 5-fold cross-validation strategy was implemented to make the classification results more reliable.The settings of the three classifiers were briefly introduced as follows: 1) KNN: KNN classifier is an algorithm for classifying samples in feature space based upon distance measurement [39].In this study, K was set as 5, and Euclidean distance was used.2) SVM: SVM is a classifier that constructs the optimal segmentation hyperplane on the feature space [40].The commonly-used radial basis function (RBF) kernel was used in this work.
3) BP_Adaboost: BP_Adaboost is an integrated classification algorithm [41] and 20 BP neural networks were used as the base classifier.It is noteworthy mentioning that the classification analyses adopted here was to provide a comprehensive Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.appreciation of the neural mechanisms of the FSR process instead of boosting the classification accuracy.

H. Statistical Analysis
In order to assess the driving fatigue effect, separate paired t-test was initially employed to compare the behavioral metrics (i.e., RT), FCs, and network metrics between the VS and FDS state in both groups.Then separate one-way analysis of variance (ANOVA) was implemented for between-group comparison at two common states.Moreover, a repeated-measure ANOVA (with the factor state) was further administered on the behavioral and network metrics to reveal the complex neural mechanisms of FSR phenomenon.The performance of classification was assessed via a previously-validated permutation test through assessing whether the classification accuracy was better than the chance level.As such, the classification procedure was repeated with randomized labels and the p-value of classification accuracy was determined as the quotient between the number of iterations that showed better results than the actual labels and the total number of iterations (i.e., 1000 in this work).The statistical differences were considered significant at the p = 0.05 threshold.All statistical analyses were conducted using in-house codes implemented in MATLAB 2019b software (Mathworks Inc., New York).

A. Behavioral Performance
According to the proposed FSR determination criteria, we found that 18 out of 26 subjects exhibited a salient FSR process.Hence, the subjects were divided into the FSR group and the non-FSR group.In Fig. 2, we showed the comparisons of behavioral metrics between both groups (Fig. 2(a)) and across four states in the FSR group (Fig. 2(b)).It can be

TABLE I CLASSIFICATION BETWEEN THE VS AND FDS IN BOTH GROUPS
seen that significantly lower (F = 4.517, p < 0.05) RT in the FDS state was revealed in the FSR group in comparison with the non-FSR group.Further cross-state comparison within the FSR group showed an "N" shape trend, that is significantly increased RT between the VS and FS state, following by a significantly decreased RT in the FSRS, then a significant increase of RT in the FDS.We further showed the distribution of time window that was categorized into four states in the FSR group (Fig. 2(c)).A significant state effect (F = 46.824,p < 0.001) was revealed in the RT of FSR group.Inspection of the distribution, we found a clear evolution pattern with VS and FDS presented towards the beginning and the end of the task while the FS and FSR states were distributed in the middle of the task.Of note, the time window distribution of each state further illustrate the apparent inter-individual differences in the evolution of driving fatigue and support the hypothesis of individual variability of the FSR phenomenon.

B. Driving Fatigue Effect in Both Groups
In order to assess the driving fatigue effect in both groups, we then look into the reorganization of brain network between two common states in both groups.In Fig. 3, we showed the functional connections with significant fatigue effect (p < 0.05) between the VS and FDS states in both groups.As expected, the number of significantly altered FCs was less in the FSR group in comparison with that in the non-FSR group (i.e., 23 (θ :α:β:γ = 4:3:2:14) vs. 45 (θ:α:β:γ = 8:7:14:16)).Moreover, We found most of the altered FCs were long-range Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.connections that exhibited an increased trend (FDS > VE, i.e., 43 out of 45 in the non-FSR group, and 21 out of 23 in the FSR group).These FCs with significant fatigue effect were set as inputs for the state classification analysis and obtained satisfactory classification accuracy (Table I, p < 0.001 for all models, 1000 permutations).Of note, three classification models exhibited relatively comparable performance, suggesting the intrinsic nature of the fatigue-related significantly altered FCs.The fatigue-related topological reorganization of brain network was quantitatively assessed through graph theoretical analysis.In Fig. 4, we showed the network metrics of four frequency bands between the VS and FSD state in each group.We found statistical differences in L w and E w global of γ band in the non-FSR group and significantly increased C w of β and γ band in the FSR group.

C. Divergent Patterns of VS and FDS in Both Groups
Following the investigation of fatigue effect within each group, we set analysis to look into the complex neural mechanisms of mental fatigue between both groups.Firstly, we performed direct statistical comparisons of FCs within the VS and FDS state between both groups (Fig. 5).Apparent differences were found between both groups in the two states (i.e., VS state, θ :α:β:γ = 50:85:48:47; FDS state, θ :α:β:γ = 31:61:39:23), where most of these FCs exhibited a unified pattern, that is, non-FSR > FSR.The identified FCs were set as input for the subject classification.Interestingly and as expected, subjects could be detected with high accuracy (Table II, p < 0.001 for all models, 1000 permutations).In Fig. 6, we showed the network measures of four frequency bands between both groups in the VS and FDS state.In the VS state, subjects in the FSR group exhibited a less integrated brain network architecture (as shown in increased L w and reduced C w , E w global , and E w local mainly in θ , α, and β frequency bands) in comparison with those in the non-FSR group.While in the FDS state, a less integrated network topology was revealed in α and β bands in the FSR group.Together, these findings showed divergent brain network patterns between both groups, therefore highlighting the individual-dependent complex neural mechanisms of mental fatigue.

D. Evolution Process in the FSR Group
We then investigated the evolution process of brain network reorganization in the FSR group.Significant state effect (p < 0.05) was found in all network metrics at β and γ bands and E w global at α band (Table III).In Fig. 7, we showed the network measures across the four states.Further inspection of the development trend of the network metrics, we revealed Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.and reduced L w ).Unlike the behavioral measures, we did not observe a clear "N" shape trend in all network metrics.
IV. DISCUSSION Considering the phenomenon of non-monotonically changes in driving fatigue evaluation indexes, this study provided a comprehensive investigation of the FSR phenomenon.To verify the FSR phenomenon, we proposed a data-driven approach for the quantitative determination of the FSR process and attempted to explore the neural mechanism of the FSR phenomenon from multiple perspectives including behavioral performance, FC, network characteristics, and classification performance.The main findings are as follows.Firstly, the RT value of FSR group in FDS was significantly lower than that of non-FSR group, which indicated that the FSR group had lower fatigue degree than the non-FSR group.Secondly, FSR group and non-FSR group had significant differences in brain functional network structures.Thirdly, the network characteristics of FSR group showed complex development trend among the four states.These findings will be discussed in details below.
Subjects were required to maintain sustained attention during the long-term driving task.Heuristically, continuous allocation and occupation of attention would lead to depletion of the finite cognitive resources [42] and lack of supplements would inevitably cause mental fatigue [20].It has long been recognized that reaction time represents a robust indicator of real-world lapses related to mental fatigue [43].In line with previous findings [5], we also showed significant performance decline as seen in slowed RT between the VS and FDS state at group level.However, we found a non-monotonic development trend in most of the participants through inspecting the behavioral impairments at individual level.Interestingly, such the non-monotonic development trend had repeatedly reported in fatigue studies [16], [20], [44].For instance, Lim and colleagues investigated the fatigue-related effect on behavioral performance and brain activities during performing sustained attention tasks (i.e., 20-min psychomotor vigilance test (PVT) [20], and 65-min auditory oddball task [44]).A non-linear fashion of behavioral metrics (including accuracy and RT) was reported.Similar findings of non-monotonic development trend were also revealed in our previous fatigue studies using similar experimental paradigm [16].According to [19], the non-monotonic development trend was interpreted as adaptive regulation.In the current work, we defined it as the FSR of driving fatigue that is regulated autonomously and internally.That is, during the FSR, subject would recover the behavioral performance by spontaneously, passively, and without active human involvement to adjust itself to fatigue, which is unlike the self-regulation process in a recent work by Love and colleagues [45], where the self-regulation was caused by shift of attentional control through extra activities (including advance road familiarization, talking to oneself, listening to music or media, talking to others, opening windows, readjusting sitting position, etc).Heuristically, these activities were considered as intentionally attention-switch, which may lead to fatigue recovery in comparison with the passively FSR.
Of note, we moved a step forward to provide a data-driven approach based upon clustering analysis of the behavioral metrics and set a qualitative criteria for the FSR determination (i.e., non-monotonically drop of RT is more than 2 out of  5 clustering levels).Most of the subjects (i.e., 18 out of 26) exhibited apparent FSR phenomenon.Statistical results of significantly reduced RT in the FSR group during the FDS state and the clear "N"-shape development trend (that is increase−decrease−increase) suggested that the FSR can adaptively adjust the brain state and restore behavioral performance.Possible explanation for the FSR phenomenon could stem from self-regulation and the adaption to the simulated driving task [46], [47], which may lead to the adjusting state of the subject to restore behavioral performance.
Functional connections have recently been considered as sensitive biomarkers for driving fatigue detection [5], [35], coinciding with the recent advent of human connectome studies [48].In order to comprehensively reveal the complex neural mechanisms underlying driving fatigue, we set out two comparisons: 1) driving fatigue effect in both group and 2) divergent patterns of both groups in two common states.In the study of driving fatigue effect, we found that the FSR group enhanced fewer FCs than the non-FSR group.Previous study has demonstrated that brain requires multiple regions to coordinate in information processing [49].Thus, in comparison with the FSR group, participants in the non-FSR group needs to enhance more FCs to improve the the efficiency of the cerebral cortex in processing and transmitting information at fatigue state when performing the same driving tasks [50], [51].Using the FCs with significant state effect as input for the classification model, we achieved satisfactory performance in both groups (94.25% and 84.41% for the non-FSR group and FSR group respectively).Of note, the classification accuracy of the state classification in the FSR group was much lower than that of the non-FSR group despite the different number of features (no. of FCs in FSR vs. non-FSR = 23 vs. 45).We have performed additional classification analysis to only use the top 23 FCs with high between-state statistics for state classification in the non-FSR group and found the superior classification performance was maintained (data not shown).Moreover, we found that significantly altered FCs exhibited a long-range frontal-related pattern, leading to increased global and local properties of brain network [38].These results are consistent with previous findings of other researchers in the study of fatigue-related alterations in brain network [52], [53].
Meanwhile, we observed that the non-FSR group and FSR group showed apparent differences in FCs and network characteristics in the VS and FDS states.Specifically, most of the FCs of the non-FSR group were found to be stronger than those in the FSR group, and these differential FCs had high classification performance.According to the neural efficiency hypothesis [54], this difference can be explained by the fact that the non-FSR group needs to activate more and stronger FCs to complete the task when performing the same intensity driving fatigue task [55], which means Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
that the FSR group can execute the same task more easily.In addition, we also compared the network characteristics between both groups.The network metrics (including C w , E w global , and E w local ) of the FSR group were lower than those of the non-FSR group, and L w was higher than non-FSR group in both states.It is well known that the brain needs more attention resources to maintain brain structure more efficiently [5].This indicates that the brain of non-FSR group requires to generate a more efficient structure and consume more resources to respond quickly to internal and external environmental changes during driving tasks [56], as proved by the results of more and stronger FCs activated to complete driving tasks.Furthermore, prolonged high-intensity resource consumption can lead to resource depletion [57], leading to worse behavioral performance of non-FSR group during FDS.Of note, the between-group differences in FCs and network characteristics showed good consistency in both states.In summary, the above results further demonstrate the feasibility and theoretical nature of our proposed FSR phenomenon and the divergent patterns between the FSR group and non-FSR group.
In the FSR group, we found FSR-related evolution process of the network reorganization.Further inspection showed the reorganizations were mainly resided between frontal and parietal regions in β and γ bands (supplementary materials).Prefrontal and parietal activity has been demonstrated to be enhanced during the performance of attention-related tasks, showing a positive correlation with fatigue levels [58].The network of frontal-parietal area is responsible for maintaining sustained attention [7], supporting cognitive function, and executive control.This pattern of attention can be divided into two processes [59], [60], [61].One is the goal-directed and controlled attention process, which is more mediated by the intra-parietal cortex and the superior frontal cortex, exerting a top-down influence on the perceptual and sensorimotor regions.The other is the stimulus-driven attention process (more automated), more mediated by the inferior frontal cortex and temporal cortex and more subcortical structures, mainly lateral to the right hemisphere, exerting top-down effects on perceptual and sensorimotor regions.Goal-directed attentional patterns consume attentional resources leading to ego depletion, without showing in automatic or stimulus-driven processes [62], [63].According to attention recovery theory [64], to supplement the consumed resources, capturing attention in the environment in a bottom-up manner can promote the recovery of attention resources [65], [66].Therefore, different attention modes have different information processing capabilities, which can regulate the degree of fatigue by enhancing and inhibiting the control mode of brain regions.It has also been reported that the endogenous enhancement of activation in the frontal-parietal region can modulate task performance [67], and the activation of the right dorsolateral prefrontal cortex can relieve fatigue symptoms and improve tolerance to tasks [68].The above findings indicate that FSR significantly improves behavioral performance by activating brain FC in the frontal-parietal region in the high frequency range.More optimized frontal-parietal structure after reorganization seems to enhance the information processing ability and resist the decline of the brain's overall function, so that behavioral performance can be restored.Some considerations should be taken into account when interpreting our findings.First, 26 healthy young subjects have been enrolled in this study.Further grouping into the FSR and non-FSR group may inevitably lead to small samples in each group.Also the unbalanced subject number within each group may influence the classification performance.We have performed the classification analysis through employing three previously-validated classification methods and the permutation analysis was adopted to further assess the classification performance.The statistically significant results lead us to believe the analysis framework in this work may reveal the intrinsic complex neural mechanisms of the driving fatigue.Nevertheless, previous studies have reported age-related differences in mental work capacity [69] and mental fatigue [70].Hence, the observations of the FSR phenomenon here may not be generalizable to people with other age range.Future work with larger independent datasets with broader age range is needed to validate our findings.Second, a datadriven analysis approach was introduced in this work for the determination of the FSR phenomenon.That is, the individual behavioral performance in terms of window-average RT was set as input for a clustering analysis.Here, the duration of the window length might influence the clustering results.We have performed additional analysis with different window length (i.e., 4-min, 5-min, and 6-min).We found the main findings of clustering intact (supplementary materials).Therefore, to ensure the number of events within the window for a stable average RT estimation as well as maintain the temporal dynamics of the behavioral metrics, the 5-min window was adopted here for the presentation of main findings.Future work that incorporating other behavioral metrics (i.e., lane deviation, speed variation) could achieve optimal clustering when extending our work to different experimental protocols that mimic long-term driving in real-world environments.

V. CONCLUSION
In this study, we introduced a data-driven approach for the FSR determination during a prolonged simulated driving task and participants were further grouped into the FSR and non-FSR group prior to the following comparisons.To the best of our knowledge, this is the first study to comprehensively investigate the neural mechanisms underlying the driving fatigue, providing evidence to support the notion that participants exhibited substantial individual differences in behavioral performance and brain network reorganization during driving fatigue.In this context, the findings of the current work shed some of the first quantitative insights on the complex neural mechanisms underlying driving fatigue and provides an important aspect that need to be considered when developing techniques for the automatic detection of driving fatigue.

Fig. 1 .
Fig.1.Data-driven determination of the FSR process.(a) The experimental setup and protocol where a 90-min prolonged driving task was administered.Behavioral measures (i.e., reaction time, RT) were estimated as mean RT within a 5-min window.Through sliding the window with 50% overlap over the entire course of driving, we obtained the development trend of RT for each subject.Of note, a normalization approach was adopted to normalize the RTs between 0 and 1 (corresponding to the maximum RT) for each participant.(b) The development trends of the normalized RT for 2 representative subjects with the FSR phenomenon (left panel) and without the FSR phenomenon (right panel).The RT vector was then subjected to a K-means clustering analysis where the RTs were clustered into 5 levels.Subject with the FSR phenomenon was determined as having RT drops ≥ 2 levels.With the FSR group, four states (VS, FS, FSRS, and FDS) that corresponding to different fatigue processes were further introduced.In the non-FSR group, a monotonic development trend of RT was revealed and two states (VS and FDS) that corresponding to the most vigilant and fatigued state were also illustrated.

Fig. 2 .
Fig. 2. Behavioral (a) RTs of the VS and FDS between both groups.(b) RTs of the four states (VS, FS, FSRS, and FDS) in the FRS group.Each bar represents the mean ± standard deviation.*, p < 0.05, **, p < 0.01.(c) Time window distribution of the four states in the FSR group.Each bar represents the stacked histogram of the number of the state appearing in each time window over the subjects in the FSR group.

Fig. 3 .
Fig. 3.The functional connections of (a) the non-FSR group and (b) the FSR group between the VS and FDS state.Connections are determined with statistical difference (p < 0.05, multiple-comparison corrected).Red edges indicate significantly-increased edge weight (VS < FDS) whereas blue edges indicate significantly-decreased edge weight (VS > FDS).Most of the altered FC exhibited a significantly increased pattern, suggesting a compensation effect to attenuate the effects of driving fatigue.

Fig. 4 .
Fig. 4. Brain network measures between the VS and FDS states in (a) the non-FSR group and (b) the FSR group.Each bar represents mean ± standard deviation.*, p < 0.05.

Fig. 5 .
Fig. 5. Statistically different FCs between the non-FSR and FSR group in (a) the VS and (b) the FDS state.The red edge indicates that the weight of FC in the non-FSR group is higher than that in the FSR group, while the blue edge indicates the opposite.The proportion of red and blue edges within each frequency band was given in (c) for the VS state and (d) for the FDS state.

TABLE II CLASSIFICATION
BETWEEN THE NON-FSR AND FSR GROUP IN BOTH GROUPS IN THE TWO COMMON STATESthat the architecture of the brain network tend to become more efficient at both global and local levels towards the end of task (as seen in significantly increased C w , E w global , and E w local

TABLE III STATISTICAL
COMPARISONS OF BRAIN NETWORK METRICS IN THE FSR GROUP