Sex Differences in Time-Domain and Frequency-Domain Heart Rate Variability Measures of Fatigued Drivers

The effects of fatigue on a driver’s autonomic nervous system (ANS) were investigated through heart rate variability (HRV) measures considering the difference of sex. Electrocardiogram (ECG) data from 18 drivers were recorded during a simulator-based driving experiment. Thirteen short-term HRV measures were extracted through time-domain and frequency-domain methods. First, differences in HRV measures related to mental state (alert or fatigued) were analyzed in all subjects. Then, sex-specific changes between alert and fatigued states were investigated. Finally, sex differences between alert and fatigued states were compared. For all subjects, ten measures showed significant differences (Mann-Whitney U test, p < 0.01) between different mental states. In male and female drivers, eight and four measures, respectively, showed significant differences between different mental states. Six measures showed significant differences between males and females in an alert state, while ten measures showed significant sex differences in a fatigued state. In conclusion, fatigue impacts drivers’ ANS activity, and this impact differs by sex; more differences exist between male and female drivers’ ANS activity in a fatigued state than in an alert state.


Introduction
It has been reported that nearly 1.3 million people are killed and 50 million people are injured in road traffic collisions each year [1], and driving fatigue is estimated to be responsible for 20-30% of all road fatalities [2][3][4]. Some studies on driving impairment have considered sleepiness and fatigue as similar mental conditions [5], and the term "fatigue" is often used as an overarching term, which includes sleepiness and mental fatigue [6]. There are many psychophysiological symptoms such as tiredness, lack of energy, difficulties to concentrate, loss of interest, and so on, caused by fatigue that dangerously affect driving [7]. Driving requires mental and physical attention and alertness to be performed effectively [8], and fatigue may affect a driver's attention and vigilance when controlling a vehicle and may result in a disastrous consequence [9].
Predicting a driver's mental state by measuring his or her fatigue before or during driving is a method that may be used to reduce traffic accidents [9]. Substantial efforts have been devoted to this topic; employed methods can be divided into three categories [10,11]: vehicle-based methods, behavior-based methods and physiological-based methods. Vehicle-based methods detect the level of a driver's fatigue usually by steering wheel movement or standard deviation in the lane position.
Behavior-based methods determine a driver's mental state by his or her facial movements through a camera. These two categories of methods become apparent only after the driver starts to sleep with limited time left to prevent an accident. Physiological-based methods, in which physiological signals, such as electrocardiogram (ECG) [8][9][10], electroencephalograph (EEG) [8][9][10] and functional near-infrared spectroscopy (fNIRS) [9], are employed, are regarded as reliable and accurate methods for fatigue detection in the early stages. However, these methods are less acceptable for receiving physiological signals because they involve placing electrodes on the body. With the development of noncontact ECG technology [10], the ECG-based method is expected to detect fatigue in a nonintrusive manner while maintaining the advantages of reliability, accuracy and timely detection. The autonomic nervous system (ANS) controls heart rate by balancing the sympathetic and parasympathetic nervous systems [12]. Several physiological and pathological changes could exert an impact on the ANS. Heart rate variability (HRV) has emerged as the most valuable noninvasive test to assess ANS function [13], and RR interval time series derived from ECGs are the source information for HRV analyses [14]. For the quantitative evaluation of ANS function, time and frequency domain methods have been widely employed in HRV analyses. Several factors, such as smoking [15], stress [16], and anxiety [17], could alter HRV measures. ANS function is also influenced by driving fatigue and could be expressed by HRV measures. The HRV measures absolute power of the VLF band (P VLF(abs) ) and power of the HF band in normalized units (P HF(nu) ) were reported to be significantly increased and power of the LF band in normalized units (P LF(nu) ) and ratio of the LF band power to the HF band power (r HF/LF ) to be significantly decreased in drivers in a fatigued state compared with those in an alert state by Awais et al. [8,10]. Tran et al. [18] reported that a driver's absolute power of the LF band (P LF(abs) ), P LF(nu) and r LF/HF increased significantly and that the mean RR interval, M RR , and P HF(nu) decreased significantly. Abtahi et al. [19] reported that the M RR , standard deviation of RR interval (D RR ), number of successive RR interval pairs that differ by more than 50 ms (N NN50 ), P LF(abs) , absolute power of the HF band (P HF(abs) ) and absolute power of all three bands (P tot(abs) ) increased significantly. These studies contribute to our understanding of the relationship between driving fatigue and ANS function and clarify the development of driving fatigue detection algorithms. A summary of the experimental designs, HRV measures and HRV measure extraction methods of six prior studies related to driver fatigue is provided in Table 1. Sex differences exist in HRV measures [20][21][22], which should be considered when employing these measures.
For the studies related to driving fatigue and HRV measures, two limitations exist. One limitation is that sex differences have been taken into consideration by few (if any) studies, and the other limitation is that the change tendency of some time and/or frequency HRV measures from the driver's alert state to fatigued state had contradictory results based on the available reports. For studies on the sex differences in HRV measures, to the best of the authors' knowledge, all of them were researched with subjects in an alert state, leaving the sex differences in HRV time and frequency measures in a fatigued state unknown. Our experiments were performed in a high-fidelity driving simulator, and fatigue was induced by prolonged monotonous, simulated driving. With being aware of the limitations in this field, the aim of the study was three-fold: First, as many HRV measures as possible were extracted using time domain and frequency domain methods to provide a comprehensive evaluation of the impact of fatigue on the driver's ANS and to provide references for the development of driving fatigue detection methods; second, the HRV measures of drivers of different sexes in different mental states were compared to provide a basis for developing driver-specific fatigue detection algorithms for specific sexes; and finally, the similarities and differences of HRV measures of males and females in both an alert and a fatigued state were compared to provide a new perspective for understanding the sex differences in the ANS. The subjects were asked to provide their Karolinska Sleepiness Scale (KSS) score every 5 min during driving.
HRV analysis was performed following reference [23], and an AR model was used for the PSD estimation of the HRV time series. [19] 22 ( A driver in the 5 min prior to or after driving was scored as being in an alert or fatigued state.
Peak-to-peak intervals were determined through a blood pressure waveform produced by the radial artery. The HRV analysis followed reference [23] and used an FFT for PSD estimation. [24] 10 (unknown) Unknown 41 ± 9 Driving on highways for a mean duration of 223 min Observers classified the state of the driver each minute as either alert or drowsy through video recordings. Ten minutes before and after a drowsy minute was defined as a drowsiness period; the other periods were defined as alert periods.
HRV measures were extracted from a 300-beat window with a step of one beat. The Hodrick-Prescott filter was used for detrending, and a periodogram was estimated using a Hanning window. [25] 12 (12) 9/3 24-30 Two simulated driving sessions lasting 15-20 min: one in the morning and one in the afternoon after lunch Two observers scored the drowsiness level according to recorded videos of the drivers' faces.
HRV measures were extracted from one-minute segments of the ECG.

Subject
Eleven male and nine female college students (age: 25.95 ± 2.67 years) with legal driver's licenses were recruited to participate in the experiment. A sum of RMB 150 was provided to the subject as compensation for their participation in the experiment. Two of the male subjects were excluded from the ECG-related research. One was excluded due to poor connectivity of the ECG electrodes during the driving period, and the other was excluded due to fatigue throughout the experiment, mainly caused by extensive studying the night before. Subjects were asked to refrain from consuming caffeine, alcohol or tea and from smoking on the testing day. All subjects provided informed consent for inclusion before participating in the study. The study was conducted in accordance with the Declaration of Helsinki, and the protocol was approved by the Ethics Committee of Tsinghua University (20160006). Details of the methods have been described in our previous studies [26].

Driving Simulator
A driving simulator with six degrees of freedom (see Figure 1) was used in the experiment. A passenger car was mounted on the motion base, and five large projections surrounded the car. The driving simulator provided a realistic driving experience to the drivers. To accelerate the development of the driver's fatigue symptoms, a simulated, straight highway-driving scenario was selected, and there were no other cars in the simulation other than the virtual experimental car during the driving period.
. J. Environ. Res. Public Health 2020, 17, x 5 of used by extensive studying the night before. Subjects were asked to refrain from consumin ffeine, alcohol or tea and from smoking on the testing day. All subjects provided informed consen r inclusion before participating in the study. The study was conducted in accordance with th claration of Helsinki, and the protocol was approved by the Ethics Committee of Tsinghu iversity (20160006). Details of the methods have been described in our previous studies [26].

. Driving Simulator
A driving simulator with six degrees of freedom (see Figure 1) was used in the experiment. ssenger car was mounted on the motion base, and five large projections surrounded the car. Th iving simulator provided a realistic driving experience to the drivers. To accelerate th velopment of the driver's fatigue symptoms, a simulated, straight highway-driving scenario wa lected, and there were no other cars in the simulation other than the virtual experimental car durin e driving period.

. Experimental Protocol
After the experimental details were introduced, single-channel EEG and ECG electrodes wer aced on the subject. The Biopac MP 150 system (Biopac Inc., Goleta, CA, USA) was used for sign

Experimental Protocol
After the experimental details were introduced, single-channel EEG and ECG electrodes were placed on the subject. The Biopac MP 150 system (Biopac Inc., Goleta, CA, USA) was used for signal acquisition. The EEG and ECG signals were acquired at a sampling frequency of 1 kHz. For the ECG, a modified lead II configuration was adopted. The left leg (LL)-equivalent electrode was placed in the upper-left quadrant, and the right arm (RA)-equivalent electrode was placed in the right infraclavicular fossa. Then, a laboratory assistant guided the subject into the car. A pretest of approximately 10 min was performed to familiarize the subject with the experimental environment, followed by a 60 min period of uninterrupted driving. In addition to the EEG and ECG signals, the driver's facial expressions were recorded by a camera mounted on the center console during the test. The subjects were required to maintain a speed of approximately 75 km/h in the simulated highway scenario. The tests were performed between 9:30 and 17:30, as these are conventional daily working hours. The ECG signals and videos were analyzed offline, and the EEG signals were left for further research.

Fatigue Assessment
The 60-min videos of the drivers' facial expressions were each divided into 60 one-minute segments. Then, five well-trained experts scored each one-minute segment a score of 1, 2 or 3 according to the criteria in Table 2. Table 2. Subjective scores for driver's mental state assessment.

Alert 1
The eyes open normally and blink quickly, the eyes are active, the attention appears focused, and attention to the outside world is maintained. The head is upright, and the facial expression is changing frequently.

Fatigued 2
The eyes appear to be partially closed, eyes appear to be partially closed, blinking duration is extended, blinking speed is decreased, eye activity is decreased, or eyes become sluggish; the subject yawns, takes a deep breath, sighs, swallows, rubs the eyelids using their hands, shakes his or her head, scratches his or her face or performs any other action that suggests fatigue or reduced concern with the environment.

Very fatigued 3
The eyes appear to be half or fully closed, the eyelids are so heavy that they are unable to open, the eyes are closed for a long period of time, there is head nodding and head tilting, and the ability to continue driving is lost.
Since the number of segments with a score of 3 was low, the scores of 3 were rescored as 2. In accordance with the length of the ECG signal recommended for short-term HRV analyses [23], the 60-min videos and the ECG signals were divided into 19 five-minute segments, with 40% overlap at each end with the start/end of the adjoining five-minute segments. In each five-minute segment, the scores of every one-minute segment assessed by each expert were summed over. The minimum and maximum of the summed score in a five-minute segment were 25 and 50, respectively. If the summed score is greater than or equal to 38, according to the simple majority principle, the driver can be considered to be in a fatigued state during the five-minute segment, and if the summed score is less than or equal to 37, the driver can be considered to be in an alert state during the five-minute segment.
To reduce misjudgments, we adopted the principle of the effective majority, which stipulated that if the summed score was greater than or equal to 40, then the five-minute segment was defined as a fatigued segment; if the score was less than or equal to 35, then the five-minute segment was defined as an alert segment. Otherwise, the segment was defined as a discordant segment. In this manner, we obtained 88 alert segments and 223 fatigue segments for the following analyses. We also obtained 31 discordant segments, which were excluded from further analysis. Figure 2 compares the summed scores of all subjects in 19 five-minute segments. The figure shows that as time elapsed, all subjects essentially changed from an alert state to a fatigued state. obtained 88 alert segments and 223 fatigue segments for the following analyses. We also obtained 31 discordant segments, which were excluded from further analysis. Figure 2 compares the summed scores of all subjects in 19 five-minute segments. The figure shows that as time elapsed, all subjects essentially changed from an alert state to a fatigued state.
. Figure 2. Boxplot of the summed scores of all subjects in 19 five-minute segments.

Extraction of HRV Time and Frequency Measures
We detected R peaks in every segment of the ECG, and abnormal beats were detected and corrected with normal R beats. Then, the occurrence times of the R peaks were denoted by a time series (t0, t1…, ti,…), and the RR interval series was obtained. The RR interval series was unevenly

Extraction of HRV Time and Frequency Measures
We detected R peaks in every segment of the ECG, and abnormal beats were detected and corrected with normal R beats. Then, the occurrence times of the R peaks were denoted by a time series (t 0 , t 1 . . . , t i , . . . ), and the RR interval series was obtained. The RR interval series was unevenly sampled, and the HRV signal was assumed to be a continuous signal derived from the RR interval series via cubic spline interpolation. The HRV time series was obtained by sampling the HRV signal evenly at a given sampling frequency of 4 Hz.
The HRV time domain measures were calculated from the RR interval series. To calculate the HRV frequency domain measures, the smoothness priors approach (SPA) with λ = 500, which corresponds to a high-pass filter with a cutoff frequency of 0.035 Hz, was applied to the RR interval series, and the power spectral density (PSD) was obtained by Welch's periodogram method [27]. Then, the PSD was divided into three bands: very low frequency (VLF, 0-0.04 Hz), low frequency (LF, 0.04-0.15 Hz) and high frequency (HF, 0.15-0.4 Hz).
The abovementioned processes were completed by employing Kubios 3.0.2 (Kubios Oy, Kuopio, Finland) [27]. All the measures were calculated in the manner recommended by the Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology [23]. The abbreviations of the measures were redefined (see Table 3) in accordance with the requirements of the Chinese national standards. 2.6. Statistical Analysis SPSS 19 (SPSS Inc., Chicago, IL, USA) was used for the statistical analyses. We calculated the means, standard deviations, medians, 1st quartile (Q1) and 3rd quartile (Q3) to describe the distributions of HRV measures for the alert group and fatigued group and for the male-alert (MA) group, male-fatigued (MF) group, female-alert (FA) group and female-fatigued (FF) group to facilitate comparisons with other studies in this field. The Mann-Whitney U test was used to assess whether the time domain and frequency domain measures significantly differed between (1) the alert group and fatigued group, (2) the MA and MF groups and the FA and FF groups, and (3) the MA and FA groups and the MF and FF groups. According to Fritz et al. [29], a common effect size statistic for the Mann-Whitney U test is r using the following formula: where N is the total number of observations, and when runs the test SPSS report the appropriate z value. The difference between the means and the 95% confidence interval of the mean difference are also presented for a difference between means convey effect size information [30]. Table 4 shows the drivers' HRV measures during the alert and fatigued states. Being fatigued showed significantly (p < 0.01) increases in M RR , D RR , D RMS , N NN50 , p NN50 , P VLF(abs) , P LF(abs) , P HF(abs) and P tot(abs) and decreases in M HR compared with being alert. Tables 5 and 6 presents the HRV measures in the male-alert (MA), male-fatigued (MF), female-alert (FA) and female-fatigued (FF) groups. The MF group showed significantly (p < 0.01) increases in M RR , D RR , D RMS , p NN50 , P VLF(abs) , P LF(abs) and P tot(abs) and decreases in M HR compared with the MA group. The FF group showed significantly (p < 0.01) increases in D RR , P VLF(abs) , P LF(abs) and P tot(abs) compared with the FA group.

Results
The HRV measures were also compared between the two sexes. The FA group showed significantly (p < 0.01) lower values of P LF(nu) , P HF(nu) and r LF/HF compared with the MA group. The FF group showed significantly (p < 0.01) higher values of M HR and P HF(nu) , and lower values of M RR , D RR , D RMS , P VLF(abs) , P LF(abs) , P LF(nu) , r LF/HF and P tot(abs) compared with the MF group. The mean difference is the fatigue group mean minus the alert group mean.

Discussion
Using a driving simulator, this study investigated ANS activity in alert and fatigued driving states through HRV time and frequency domain methods. We focused especially on the sex difference in both mental states. For the comparison with existing studies, assessment of the results with and without considering sex differences led to the following points.

HRV Time Domain and Frequency Domain Measures of Drivers without Considering the Difference of Sex
Our results comparing the HRV time and frequency domain measures of drivers in different mental states are shown in Table 4. Tables 7 and 8 show comparisons of the results of seven studies (six previous studies and this study) related to HRV time and frequency domain measures of drivers in different mental states.
As shown in Table 4, all six HRV time domain measures were significantly different between the alert and fatigued states. As shown in Table 7, M HR and N NN50 are consistent across all the included studies with respect to the change direction from the alert state to the fatigued state.
Although Tran et al. [18] obtained the opposite result, three studies demonstrated that M RR is significantly higher in the fatigued state than in the alert state, which is also supported by a review from Ismail et al. [31]. Thus, the influence of fatigue in terms of M RR can be determined. D RR is likely to increase significantly from the alert state to the fatigued state based on four of the six studies, and p NN50 is highly likely to increase significantly because it is related to M RR and N NN50 , which definitely increase significantly, although the results of this study and that by Tran et al. [18] do not fully agree. The change direction of D RMS is still unclear from Table 7.
As shown in Table 8, an increased change direction in P VLF(abs) and P tot(abs) is consistent across all the included studies, and regarding the powers of the VLF, LF and HF bands, the absolute values are better than the normalized values for discriminating between mental states. D RR reflects the ebb and flow of all the factors that contribute to HRV [28] and correlates with P tot(abs) , with r > 0.9 [32]. Fatigue can be considered a factor that contributes to HRV since significant increases in D RR and P tot(abs) were confirmed in drivers' fatigued state compared with their alert state. Given that p NN50 and P VLF(abs) are significantly increased from the alert state to fatigued state, more parasympathetic and sympathetic activities can be inferred to be present in the fatigued state as p NN50 is often interpreted as a proxy for cardiac parasympathetic activity and P VLF(abs) may mirror sympathetic activity [33].

The HRV Time and Frequency Domain Measures of Drivers Considering the Difference of Sex
To evaluate the sex difference, the differences in the impact on the change in mental states in terms of the HRV measures were compared between males and females. Table 6 shows that four HRV measures exhibited significant differences between the alert state and fatigued state in both males and females. For males, another four HRV measures were significantly different between the two mental states, while for females, no other HRV measures were significantly different between the two mental states.
This situation may imply that ANS activities are more sensitive to the change in mental state for males than for females, as HRV is regarded as the most valuable non-invasive test to assess ANS function [13]. The HRV time and frequency domain measures were adopted by Huang et al. [34] and Patel et al. [35] for the establishment of driving fatigue detectors. It could be reasonable to infer that a better result would be obtained for such detectors if the sex factor was considered.
The sex-related influence on the HRV measures was observed by Ryan et al. [20] early in the 1990s. Recently, a comprehensive study on both age and sex differences in HRV measures was reported by Voss et al. [22]. In their study, an HRV analysis was performed on a 5-min ECG recording obtained in the supine position. According to their study, in the 25-to 34-year-old age group, there were four HRV time and frequency domain measures (M RR , P LF(abs) , P LF(nu) and r LF/HF) that were significantly higher in males than in females, one HRV measure (P HF(nu) ) that was significantly lower in males than in females, and five HRV measures (D RR , D RMS , N NN50 , P HF(abs) and P tot(abs) ) that showed no significant difference between the two sexes. Our study confirmed these results, except for those for M RR and P LF(abs) , which showed a higher median value in males than in females without statistical significance. The difference in the two studies may be that subjects in the study by Voss et al. [22] were in a supine position without a workload, but they were in a sitting position with a driving workload in our study. It should be noted that all studies, including those by Ryan et al. [20] and Voss et al. [22], to our knowledge, have researched the sex difference of HRV measures of subjects in only the alert state.
Investigating the sex-related influence on HRV measures in a fatigued state could enable further understanding of sex differences. However, little is known in this field. In our study, a comparison of the sex-related influence on the HRV measures between an alert state and a fatigued state is shown in Table 6. The sex differences are larger in the fatigued state than in the alert state: seven of the thirteen HRV time and frequency domain measures that show no significant difference between males and females in the alert state become significantly different in the fatigued state. The relationship between stress and HRV measures was reported by Tharion et al. [36], and stress is associated with fatigue [7,32]. Thus, sex differences of HRV measures in subjects in the stress state are highly expected.
The mechanism of the sex differences in the HRV measures has not been clearly disclosed, and our contribution is a new question proposed to the field of neurophysiology: What is the mechanism underlying the sex differences in healthy subjects' ANS activity in the fatigued state but not in the alert state? From the application point of view, since the HRV time and frequency domain measures in the state of mental fatigue could better characterize the physiological characteristics of sex, whether these measures can better characterize more physiological and pathological characteristics in the state of mental fatigue is a topic worthy of investigation.
In our opinion, an ideal study related to HRV measures and driving fatigue should (1) involve real vehicles with real-world driving scenarios to ensure a real driving experience, (2) assess fatigue in an offline method to allow careful inspection, and (3) extract HRV measures by established software to increase the reproducibility of the experiment by other investigators. None of the abovementioned studies fulfilled all these requirements, thus leading to some disagreements regarding the results to an extent. One limitation of this study is that the experiments were performed in driving simulators rather than in real vehicles, and some differences may exist between our results and those from real driving conditions. Another limitation is that the number of participants is somewhat scarce and presents similar sociodemographic characteristics (age, vocation, etc.), which can limit the generalization of results. Further research using a large sample size with different sociodemographic characteristics will help to further elucidate the findings of this study. Besides, Our experiment was performed between 9:30 and 17:30; however, since the time of day may affect the perception of fatigue, further experiments performed at the same time of day may yield a more elaborate result regarding the relationship between HRV measures and driving fatigue.

Conclusions
This study investigated the effects of both mental states and sex factors on drivers' ANSs by extracting thirteen time and frequency domain HRV measures from ECG signals. The main conclusions are as follows: (1) For drivers of both sexes, there are ten HRV time domain and/or frequency domain measures that are significantly different between the alert state and the fatigued state, implying that the HRV time domain and/or frequency domain measures have the ability to characterize drivers' mental states; (2) For male drivers, nine HRV time domain and/or frequency domain measures are significantly different between the different mental states, and for females, four measures are significantly different between the different mental states. This finding indicates that the HRV time and frequency domain measures have sex differences in characterizing the mental states of drivers; and (3) There are three HRV time domain and/or frequency domain measures in the alert state that show significant differences between male and female drivers, while ten measures in the fatigued state show significant differences between the two sexes. It can be seen that there are more sex differences in autonomic activities in a fatigued state than in an alert state.