Predicting Perceived Realism in Virtual Reality Driving Simulations Using Participants’ Personality Traits, Heart Rate Changes, and Risk Preference

Virtual reality (VR) has recently been adopted for driving simulations to enhance their realism and thus improve the validity of the simulation results. However, given that perceived realism is a subjective factor that varies by individual, understanding and predicting perceived realism in VR driving simulations are prerequisites for enhancing their validity. Studies on VR have investigated how individual factors such as psychophysiological metrics are associated with perceived realism. However, how these psychophysiological metrics are associated with perceived realism in VR driving simulations has not yet been investigated. To address this problem, this study investigated the relationship between perceived realism and psychophysiological metrics, including individual characteristics (sex, age), personality traits (psychopathy, Machiavellianism, sensation seeking, impulsivity), heart rate changes during the event, and risky decision-making during the event, across three driving simulations. The results indicated that psychopathy, Machiavellianism, heart rate changes during the event, and risky decision-making during the event were significantly correlated with the perceived realism of VR driving simulations. In addition, we tested three types of machine learning models to find the appropriate ones for predicting perceived realism, showing that the tree-based algorithm had the highest prediction accuracy.


I. INTRODUCTION
Improving the realism of driving simulations is crucial for ensuring the validity of their outcomes.Consequently, various methods have been employed to enhance perceived realism in driving simulations such as using monitors or projectors in conjunction with physical mockups of vehicle cockpits [1].This combination helps shape the simulation environment to closely replicate real-world driving The associate editor coordinating the review of this manuscript and approving it for publication was Xiaogang Jin .conditions.However, despite conventional simulations being costly and space-consuming [2], their validity remains questionable because participants may be conscious of their surrounding environment [3], which might reduce perceived realism.
One possible solution to solve this problem is to adopt virtual reality (VR) driving simulations [2].Although previous studies have not found that VR driving simulations have better validity than conventional ones [4], [5], recent advancements in VR technology, particularly concerning resolution and tracking capabilities, suggest that VR can create driving simulations of comparable quality to conventional simulations while offering additional advantages owing to its enhanced immersion [6].The first advantage of using VR driving simulations over conventional driving simulations is that they can isolate participants in the real world, which can improve immersion [7].Second, VR driving simulations can be run with less space and at a lower cost than conventional driving simulations [5], which improves the accessibility and reproducibility of the simulation results.Third, several sensory tracking systems such as eye-tracking [8], heart rate tracking [9], and facial tracking [10] have recently been merged with VR, making it easier to assess various sensory signals for individual psychophysiological metrics.
Because VR can induce realistic behavior, several studies have investigated its perceived realism using psychophysiological metrics to identify potential relationships with presence (defined as the sense of being inside a virtual environment [11], which is closely associated with perceived realism [12]) using personality traits [13], facial muscle activity with eye-tracking [14], and neural activity [15], [16].Conversely, studies on perceived realism in driving simulations have primarily concentrated on the influence of external factors rather than on the psychophysiological metrics of drivers.In subsections A and B, we review previous studies that have investigated the factors influencing perceived realism in driving simulations and the individual factors used in the present study that may be associated with perceived realism in VR, respectively.

A. RELATED WORK ON PERCEIVED REALISM IN VR DRIVING SIMULATIONS
Several previous studies have explored the factors that influence perceived realism in driving simulations to enhance perceived realism.Among those studies investigating factors influencing perceived realism in non-VR driving simulations, many have examined external factors.Nesti et al. [17] changed the tilt rate of a driving simulation used to implement sustained acceleration and found that when the tilt limit of the driving simulation increased, perceived realism also rose.Bertollini et al. [18] examined yaw rate turning maneuvers and compared three scenarios: without motion cueing, with motion cueing but without yaw rate turning maneuvers, and with both motion cueing and yaw rate turning maneuvers.They found that perceived realism was higher in the motion cueing condition than that without motion cueing.However, they observed no significant effect of yaw rate turning maneuvers.
Rock et al. [19] investigated perceived realism with and without surrounding traffic.They found that perceived realism significantly increased when surrounding traffic existed and that realistic virtual traffic agent behavior also increased perceived realism.Himmels et al. [20] compared the perceived realism of lower-and higher-fidelity simulators and found that the latter, which had a full-vehicle mockup with a 360-degree field of view, achieved higher perceived realism than the former, which had a full-vehicle mockup with a 180-degree field of view.By contrast, Shi et al. [21] manipulated a driving scenario that changed the details of buildings and the surrounding environment to investigate the differences in perceived realism and found no significant effect of these details of VR scenarios.Finally, recent research investigating the influences of external factors on perceived realism in VR driving simulations includes the study by Colley et al. [22], who found that swivel seat rotation increased perceived realism.Similarly, Yu et al. [23] found that vibrotactile feedback from a headband enhanced perceived realism.In general, several driving simulation studies have investigated the influence of external factors on perceived realism to enhance realism.However, to the best of our knowledge, none have investigated the association between perceived realism in VR driving simulations and the internal factors of drivers, including their individual characteristics, personality traits, and psychological metrics.Since perceived realism in VR is characterized as one's own assessment of one's realism level [24], it is also important to understand the influence of individuals' internal factors to enhance perceived realism in driving simulations.To address this gap, our study aimed to investigate the relationship between perceived realism and various psychological metrics.In the following section, we review the psychophysiological metrics that may be associated with perceived realism in VR driving simulations taken from other VR studies and conventional driving simulations as well as introduce the extracted factors used in the present study.

B. RELATED WORK ON THE FACTORS INFLUENCING PERCEIVED REALISM
First, several prior studies have shown that user experiences in VR are influenced by personal characteristics, although there have been cases showing contrasting results.For example, one previous study showed that older adults had higher presence in virtual environments than younger adults [25], whereas another study showed that individuals aged 35-45 years reported lower immersion in VR than those aged 10-20 years [26].Similarly, one study showed that male participants had a higher sense of presence in VR [27], whereas another study showed that when vibratory stimuli were included, only female participants had an increased presence [28].In a similar vein, VR driving simulation studies have shown that sex and age influence user experiences; for instance, female participants showed increased simulation sickness in driving simulations than male participants [29] and older adults showed increased simulator sickness than younger adults [30], which was a factor negatively associated with presence in VR.Owing to these previous findings, we include sex and age as potential factors associated with perceived realism.
Second, various previous studies have shown that individuals' personality traits are another factor that may affect perceived realism in VR.For instance, personality traits such as  anxiety [13], introversion, [31] agreeableness, and conscientiousness [32] have been shown to be associated with a sense of presence in VR.However, in this study, we do not adopt the personality trait questionnaires typically used in other VR studies since emergencies and accidents, which are commonly simulated in VR driving scenarios [33], [34], [35] and influence perceived realism, are typically ignored by these previous studies.Therefore, we use personality traits that are associated with driving behavior in an emergency and may be associated with perceived realism, including psychopathy, Machiavellianism, sensation seeking, and impulsivity.
First, psychopathy, which is a personality trait indicating a lack of empathy and impulsivity with antisocial behavior [36], has been shown to be associated with risky [37] and aggressive driving [38].Personality traits linked to psychopathy have also been found to predict the reported uncanny valleys of virtual characters [39] that degrade the realism of VR in driving simulations.As such, we include psychopathy as a factor that may be associated with perceived realism.
Second, Machiavellianism, which is defined as manipulating others for one's own interest [40], has been shown to be associated with driving behavior such as anger, aggression [41], and speeding [42].Moreover, together with psychopathy, Machiavellianism represents cold-heartedness [43], which, associated with cynical attitude [44], may influence perceived realism.Thus, we also include Machiavellianism as a factor that may be associated with perceived realism.
Third, sensation seeking and impulsivity are widely used in driving behavior studies and have been shown to be associated with risky driving [45] and accident prevalence [46], [47].Additionally, impulsivity is positively correlated with a sense of presence in VR [48] and high sensation seekers report relatively low spatial presence in robotics studies [49].We thus add sensation seeking and impulsivity as factors that may also be associated with perceived realism in VR.
Third, since presence has been shown to be associated with emotional arousal [50], several previous studies have explored the association between physiological signals and perceived realism.These studies have found that a sense of presence is positively correlated with physiological signals [51] and heart rate [52].Additionally, when comparing heart rate, skin conductance, and skin temperature in stressful virtual environments, changes in heart rate have been found to be an objective measure of presence [53].Since previous studies have shown associations between heart rate and perceived realism, we include heart rate as a factor that may be associated with perceived realism in VR.
Finally, decision-making during an emergency may be associated with perceived realism in VR driving simulations.A previous study suggested that the high perceived realism of driving simulations results in realistic driving behavior [54] and that decision-making in an emergency may change across the range of perceived differences in VR driving simulations.Therefore, this study includes the potential association between perceived realism and decision-making in an accident situation.

C. RESEARCH MOTIVATION AND HYPOTHESES
To enhance perceived realism and ensure the validity of driving simulation outcomes, it is important to know whether the above-mentioned psychophysiological metrics influence perceived realism as well as whether perceived realism can be predicted by such factors.To this end, VR driving simulations were run in three experiments and the potential relationship between the psychophysiological metrics and perceived realism was evaluated.Additionally, three types of machine learning models were applied to identify the best models to predict perceived realism.
The present study is novel owing to the following reasons.First, no previous studies have used psychophysiological metrics to predict perceived realism in VR driving simulations.Second, the manner in which perceived realism is associated with risky decision-making in VR driving simulations has not been investigated in prior research.Finally, no previous studies have compared different types of machine learning models to identify those suitable for predicting perceived realism in VR driving simulations using offline personal background data (individual characteristics and personality traits) and online data (psychophysiological data measured in real time).
Based on prior studies of the factors influencing perceived realism, the following four hypotheses were proposed: 1. Individual characteristics are correlated with perceived realism during VR driving simulations.2.An increased heart rate is positively correlated with perceived realism during VR driving simulations.3. Perceived realism influences risky decision-making during VR driving simulations (i.e., those who experience higher and lower realism behave differently).4. Machine learning models significantly predict perceived realism using psychophysiological metrics.

A. DATASET
This study's datasets were partially derived from previous VR decision-making studies that have investigated sacrificial decisions in an accident situation (Dataset 1) [34] and situational awareness and decision-making differences based on different sensory warnings in an accident situation (Dataset 2) [35].Additionally, the present study included an ongoing study that investigates the same sacrificial decision-making in Dataset 1 from a conditionally automated driving perspective (Dataset 3; see Table 1 for a detailed description of the datasets).

B. GAME DESIGN
Unity3D (Unity Technologies, San Francisco, USA) 5.4.1.f1(Dataset 1), 2017.2.0.f3 (Dataset 2), and 2018.2.14f1 (Dataset 3) were used to develop the vehicle driving simulation.All the driving simulation baselines were implemented using the same assets freely available at https://www.assetstore.unity3d.com/en/#!/content/10.The same head-mounted display (Oculus Rift CK1; Irvine, USA; resolution = 1080 × 1200 px for each eye at 90 frames per second) and wheel-pedal interface (Joystick, Power Racer 270 DX; Seoul, KR) were used to deliver the VR visualization.To simplify the process, only the accelerator and brake pedals were used when moving both forward and backward (e.g., brake pedals were used when going backward).The earphones of the VR headset played the sounds of the vehicle engine and navigation to better immerse users in the driving simulation.In the driving simulations, all the datasets had the same driving course before entering the session where the event occurred.The differences in the events and driving conditions are presented in Table 1.Based on these driving simulations, all the experiments in the dataset consisted of three training sessions to train participants to adjust to the driving conditions before the solitary test session.To facilitate this, participants were briefed on the objectives of the training before the start of the session.The objectives encompassed familiarizing participants with VR driving simulations and acclimating them to the complexity of the course design.The course had multiple forks with randomly selected paths at each juncture, leading to a cliff.Entry onto this path resulted in participants' failure during the session.Before each fork, visual (Dataset 2), auditory (Dataset 3), or visual and auditory warnings (Datasets 1 and 2) were presented to participants to advise them on the direction of the cliff to prevent them from falling.The auditory warning signal was a human voice announcing the cliff's location and the visual signal was a typical cliff danger sign.Using the built-in ''Yuna'' voice of Mac OS X, the auditory announcement was delivered in Korean, for example: ''Cliff to the left.''Before starting the test session, participants were informed that their lap time for the next session (the test run) would be recorded and that they would need to try again if they were unsuccessful in reaching the goal.During the test session, unforeseen incidents occurred (pedestrians blocking the road or trees falling and blocking the road) on the side of the fork that did not lead to the cliff to examine how individuals make decisions in highrisk situations.Regarding Datasets 1 and 2, participants drove manually and could choose any direction during the event.Regarding Dataset 3, the event started with autonomous driving and participants received additional information that autonomous driving would choose the direction to fall down the cliff.In this situation, participants could take over and manually choose the direction of driving or leave the decision to the autonomous vehicle and enter the direction that leads to the cliff.The experiment was automatically terminated before the vehicle either collided with obstacles or plummeted from the cliff (Figure 1).
Participants were instructed to remove their headsets and stop the test if they experienced dizziness or motion sickness at any point during the experiment.Building on previous studies that have examined the association between drivers' personality traits and decision-making [34], [35] as well as the influences of sensory warnings [35], this study focused on the perceived realism of VR driving simulations as well as the impact of psychophysiological metrics.

C. PARTICIPANTS
A total of 354 participants were recruited for this study (Dataset 1: 66, Dataset 2: 180, Dataset 3: 108; 257 men, 97 women, mean age = 23.0;SD = 2.78).All participants were recruited from university community sites and none had problems watching 3D in VR or reported any psychological or neurological problems.The experiments on Datasets 1 and 2 were approved by the local ethics committee of Korea University (1040548-KU-IRB-16-127-A-1, KU-IRB-2018-0096-01) and that on Dataset 3 was approved by the local ethics committee of Kyung Hee University (KHSIRB-21-328 [RA]).

D. MEASURES OF PERCEIVED REALISM, PERSONALITY TRAITS, INDIVIDUAL CHARACTERISTICS, AND RISKY DECISION-MAKING
First, participants rated the perceived realism of the VR driving simulation on a five-point Likert scale from 1 ''Very unrealistic'' to 5 ''Very realistic.''Second, Levenson's selfreport psychopathy scale [55], Machiavellianism [56], sensation seeking [57], and impulsivity [58] were used to assess participants' personalities.To enhance the accuracy of the questionnaire, Korean versions of the psychopathy scale [59], interpersonal reactivity index [60], Machiavellianism [61], sensation seeking scales [62] that translated Zuckerman's sensation seeking scale V, and impulsivity scales that translated Barratt Impulsiveness Scale-11 [63] were used, which were validated for Korean participants.Third, information related to sex and age was collected to determine their potential relationships with perceived realism.Finally, based on our previous study on risk evaluation in the same event situation [35], we assigned participants who chose to fall down a cliff to a high-risk decision-making group.

E. MEASURE OF HEART RATE CHANGES
An MP 35 System (Biopac Inc., Goleta, CA) with BSL Pro 4.1 software was used to assess participants' heart rate changes.For this purpose, a modified lead II configuration was used, which involved placing three electrodes underneath both the clavicles and bottom-left rib.Heart rate changes (in beats per minute) were determined 3 s before and after the trees fell or the pedestrians appeared and within a 3 s window before and at the end of the event (either colliding or falling down the cliff) because autonomic arousal and heart rate changes are highly related (event onset).In addition, because the duration of the event changed depending on participants' driving speeds, unity-controlled parallel-port timestamps were used to precisely time the event.

F. STATISTICAL ANALYSES
First, the correlations among the personality traits, individual characteristics, and perceived realism were investigated to identify the factors significantly associated with the perceived realism of VR driving simulations.Second, to understand the detailed impact of participants' personality traits on perceived realism, we adopted three subscales of the psychopathy measure (callousness, egocentric, and antisocial) [64], three subscales of the impulsivity measure [58], and four subscales of the sensation seeking measure [65].Third, to enhance our understanding of individuals' differences in perceived realism, we divided participants into high (rated 4 or 5 on perceived realism), medium (rated 3), and low (rated 1 or 2) groups.

G. MACHINE LEARNING MODELS FOR PREDICTING PERCEIVED REALISM
To find appropriate models to predict perceived realism based on the psychophysiological metrics, we evaluated three types of machine learning models.The first type was general linear model-based regression, which includes multiple linear regression, Ridge regression, and Lasso regression.Second, we selected tree-based regression models, which may be better for categorizing discrete variables such as perceived realism instead of continuous variables.These include decision trees, random forest [66], and gradient boosting [67].Finally, since the relationship between perceived realism and the psychophysiological metrics could be neither linear nor categorical, we added non-linear machine learning algorithms, including support vector regression, K-nearest neighbor, and Gaussian process regression.
The analysis first included all the psychophysiological metrics as factors that may predict perceived realism.Next, features that did not show either significance or trends in the correlation analysis were removed and tested using an offline model (except heart rate changes and risky decision-making), an online-only model, and an offline with online model to investigate the improvement in prediction accuracy compared with the baseline model.
In all the analyses, the data were split into 80% training data and 20% test data, and five-fold cross-validation was conducted to calculate participants' predicted perceived realism.Using the actual and predicted perceived realism values, Pearson's correlation analysis was run to determine whether the predicted values were significantly correlated with the actual values.

H. OVERALL PROCEDURE
First, the completed personality trait questionnaires were independently gathered one to five days before the experiment.Before the actual experiment started, the researchers indicated that the goal of the study was to examine possible connections between driving behavior and heart rate changes.Participants were also supplied with written instructions on how to operate the vehicle (manually or with conditionally automated driving) and determine the direction of the cliff before reaching the crossroads.They were tasked with aiming to drive as quickly and safely as possible without failing.
After reading the written directions and confirming that they understood them, the experimenter fitted the electrocardiography apparatus and head-mounted display.Participants in Dataset 1 were randomly assigned to either the experimental or the control group; however, all participants in Datasets 2 and 3 were assigned to the same group.Three training sessions were then completed, with the total training time ranging from 5 to 10 min across the three experiments.
Following each training session, the experimenter examined participants' conditions, asked whether they had experienced cybersickness, and suggested a brief rest.After the last training session, the experimenter informed participants that the upcoming lap would be the test session and that their lap time would be recorded.During the test session, a sudden event occurred and the simulation ended after the event.After the experiment, the researcher re-examined participants' conditions and asked them to complete the questionnaire on the perceived realism of the VR driving simulation.

A. DESCRIPTIVE STATISTICS
First, one-way analysis of variance was applied to investigate whether the personality traits differed across the datasets.The results indicated that none of the personality trait scales differed significantly (Psychopathy: F(2,351) = 0.391, p = 0.677; Sensation seeking: F(2,351) = 0.391, p = 0.677; Machiavellianism: F(2,351) = 0.391, p = 0.677; Impulsivity: F(2,351) = 0.391, p = 0.677).Second, the perceived realism of the VR driving simulation was analyzed.Perceived realism was rated approximately 2.9 out of 5 on average (Table 2).Third, the heart rate change data showed that participants' heart rate did not significantly change during the event (t(353) = 0.28, p = 0.78).

TABLE 2.
Descriptive statistics for this study.Personality traits were rescaled to 1-100% for a better comparison.

B. CORRELATION ANALYSIS
Pearson's correlation analysis was used to investigate the relationship between perceived realism and the psychophysiological metrics.The results indicated that psychopathy (r p = 0.034), Machiavellianism (r = −0.18,p < 0.001), heart rate changes during the event (r = 0.15, p = 0.010), and risky decision-making during the event (r = 0.17, p = 0.002) were significantly correlated with perceived realism in VR driving simulations, while sensation seeking showed a trend toward significance (r = −0.10,p = 0.071) (Table 3).As the results showed that psychopathy was significantly correlated with perceived realism, the three subscales of psychopathy were investigated.The results showed that callousness (r = −0.15,p = 0.004) was significantly correlated with realism, whereas egocentrism (r = −0.06,p = 0.234) and being antisocial (r = −0.03,p = 0.567) were not (an additional correlation analysis is presented in Table 7).Overall, the results showed that callousness, risky decision-making during the event, and heart rate changes during the event were significantly correlated with perceived realism in VR driving simulations.

C. DIFFERENCES IN THE PSYCHOPHYSIOLOGICAL METRICS BY PERCEIVED REALISM GROUP
To understand the influence of the psychophysiological metrics on perceived realism more in depth, we divided participants into high, medium, and low perceived realism groups.First, we ran chi-square tests and found significant differences in risky decision-making during the event by group (χ2 = 10.964,p = 0.004) but no significant differences by sex (χ2 = 2.823, p = 0.244).Next, one-way analysis of variance was run to investigate whether there were differences in personality traits and heart rate changes during the event between the groups.The results showed that psychopathy (F(2,351) = 4.439, p = 0.012), Machiavellianism (F(2,351) = 9.631, p < 0.001), and heart rate changes during the event (F(2,351) = 5.465, p = 0.003) were significantly different between the groups (Table 4).Next, post-hoc paired t-tests were run on the significant factors found from the previous analysis.We found that the high perceived realism group had significantly lower psychopathy than the medium group (t(210) = 3.119, p = 0.002).Moreover, the high perceived realism group had significantly lower Machiavellianism than the medium (t(210) = 3.7, p < 0.001) and low (t(261) = 3.776, p < 0.001) groups.Additionally, we found that the high perceived realism group showed a significantly increased heart rate during the event than the low group (t(261) = 3.060, p = 0.002) and that the low group exhibited significantly less risky decision-making during the event than the medium (t(261) = 2.528, p = 0.012) and high (t(261) = 3.120, p = 0.002) groups (Table 4).Overall, the results indicated that the different perceived realism groups showed different individual characteristics, heart rate changes during the event, and risky decision-making during the event.

D. PREDICTING PERCEIVED REALISM FROM THE PSYCHOPHYSIOLOGICAL METRICS
Three types of regression-based machine learning models were next applied to determine which types significantly predicted perceived realism when including all the features and including only online and offline features.The results showed that when including all the features, only random forest significantly predicted perceived realism (r = 0.29, p < 0.01).When including only offline features, only random forest showed trends to predict perceived realism (r = 0.21, p = 0.05).Finally, when including only online features, linear regression (r = 0.21, p = 0.08), Ridge regression (r = 0.21, p = 0.08), Lasso regression (r = 0.23, p = 0.05), and the decision tree (r = 0.23, p = 0.05) showed trends to predict perceived realism.The results indicated that when including all the features or offline features, tree-based models predicted perceived realism better, whereas general linear model-based regression showed better trends to predict perceived realism when including only online features (Table 5).

TABLE 4.
Descriptive statistics of the high, medium, and low perceived realism groups.Personality traits were rescaled to 1-100% for a better comparison.

TABLE 5.
Regression analysis for predicting perceived realism from the psychophysiological metrics.

E. FEATURE REALISM
Since the previous analyses indicated that including either all online or all offline features did not significantly predict perceived realism, we next employed feature selection-based prediction to explore potential enhancements.First, we selected factors that were significantly correlated with perceived realism in the correlation analysis and applied the same machine learning model to investigate potential enhancements in prediction accuracy.Since both online features (heart rate changes during the event, risky decision-making during the event) showed significant correlations and were already analyzed in the online only-based prediction in the previous section, we investigated only the offline with online features and offline-only models.We found that linear regression (r = 0.22, p = 0.06), Ridge regression (r = 0.22, p = 0.06), and random forest (r = 0.23, p = 0.05) showed trends toward predicting perceived realism when including both offline and online features (psychopathy, Machiavellianism, heart rate changes during the event, risky decision-making during the event), whereas decision tree (r = 0.29, p = 0.02), random forest (r = 0.41, p < 0.01), and gradient boosting (r = 0.31, p < 0.01) significantly predicted perceived realism when including only significantly correlated offline features (Table 4).Figure 2 shows the visualization results.
Finally, since offline with online features showed worse performance than when including all the features in the previous analysis, we added the additional offline feature of sensation seeking that showed a trend toward being correlated with perceived realism.The results showed that when including offline with online features, decision tree (r = 0.39, p < 0.01) and random forest (r = 0.27, p = 0.01) significantly predicted perceived realism.Moreover, when including only offline features, decision tree (r = 0.37, p < 0.01), random forest (r = 0.35, p < 0.01), gradient boosting (r = 0.28, p = 0.02), and K-nearest neighbor (r = 0.30, p = 0.01) significantly 12144 VOLUME 12, 2024 Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
predicted perceived realism (psychopathy, Machiavellianism, sensation seeking; Table 4).Overall, the results indicated that the tree-based algorithm showed the best prediction of perceived realism when using only offline features, while including sensation seeking enhanced prediction accuracy if including online with offline features.

IV. DISCUSSION
This study used individual characteristics (sex, age), personality traits (psychopathy, Machiavellianism, sensation seeking, impulsivity), heart rate changes during the event, and risky decision-making during the event to predict perceived realism in VR driving simulations.Importantly, we found that Machiavellianism, psychopathy, heart rate changes during the event, and risky decision-making during the event were significantly correlated with perceived realism in VR driving simulations.Additionally, we found that it is possible to significantly predict perceived realism when using significant offline features with the tree-based algorithm.
First, the study hypothesized that individual characteristics were correlated with perceived realism in VR driving simulations.The correlation analysis indicated that psychopathy and Machiavellianism were significantly (and negatively) correlated with perceived realism, supporting the first hypothesis.Additionally, based on the subscale analysis, callousness, a subscale of psychopathy, was significantly correlated with perceived realism, which is one of the subfactors of Machiavellianism (see Table 7) [68].This result concurred with that of a previous study in which callousness was significantly correlated with impaired realism [69].Callousness is defined as the absence of empathy or guilt [70].Those who have higher callousness may show a cold-hearted attitude toward a shortcoming in the driving simulation environment and event situation, and therefore evaluate perceived realism in VR driving simulations lower than other participants.
The second hypothesis suggested that an increased heart rate is positively correlated with perceived realism in VR driving simulations.This study's results supported the second hypothesis, finding that those who perceived a more realistic simulation experienced higher emotional arousal during the event, in line with several previous studies [51], [71], [72].Interestingly, the average heart rate before the event was significantly correlated with perceived realism (see Table 7).This implies that heart rate changes and the overall heart rate of participants who experience higher perceived realism may increase.This is in line with the aforementioned result suggesting that callousness is an important factor influencing perceived realism, as callousness is negatively correlated with one's resting heart rate [73].
Regarding the third hypothesis, the results indicated that risky decision-making during the event was positively correlated with perceived realism and that the low perceived realism group showed significantly less risky decision-making than the medium and high groups, supporting the third hypothesis.This is in line with a previous study that used electric shocks to imitate the pain of enemy fire, which can reduce shooting decision errors [74].Potentially, in this study, those participants who perceived the VR driving simulation as more realistic intuitively reacted to avoid obstacles during the event and chose to take more risks to avoid accidents.
For the final hypothesis of this study, the results indicated that tree-based regression machine learning models significantly predicted perceived realism when including all the features or only offline features.One potential reason why the tree-based models outperformed the other models was the possible non-linear relationship between perceived realism and the psychophysiological metrics.Hence, the interaction of several features could contribute to the accurate prediction of perceived realism.For example, psychopathy and Machiavellianism showed the highest correlations of the four personality traits and using these two factors showed the highest prediction accuracy of all the potential combinations.This implied that tree-based models are suitable for modeling the interaction between features [75] and can predict perceived realism better.Tree-based models might also be better when a collection of predictor or classification variables are paired with a single outcome variable [76], which is relevant to the present work.However, when only online features were included, the linear regression models outperformed the other algorithms.Hence, if only online data are available, linear models could be better than tree-based models.

V. LIMITATIONS OF THE STUDY
This study had several limitations.First, the event occurred only once and the heart rate changes used in the present study were only observed at specific time points.However, the heart rate before the event showed significant correlations with perceived realism (see Table 7), which implies that without an event, the overall heart rate may differ based on perceived realism.If participants are repeatedly exposed to events, the predictability of the situation may increase.Therefore, emotional arousal during an event could significantly differ from similar situations in the real world.Future studies should test the abovementioned results based on the differences in heart rate changes between the low and high perceived realism groups or run multiple repetitions of event scenarios to investigate the potential influence of predictability.
Second, this study's sample was biased toward young men, which limited the generalizability of the findings.However, young male drivers have a higher probability of vehicle accidents in the real world [77], [78].Investigating the relationship between perceived realism and the psychophysiological metrics of young adults would help enhance the ecological validity of the simulation results and could be used to prevent future accidents.To extend the results of the present study to the general population, future studies should recruit samples of various ages and investigate the potential differences between younger and older adults.
Third, in the current study, average perceived realism was around a medium level, suggesting that realism should be enhanced to test the validity of the prediction model in realistic driving simulations.A potential reason for the moderate level of realism was that the present study did not incorporate actual movement or acceleration, nor did it simulate realistic environmental backgrounds such as traffic flow, which might have limited perceived realism.To enhance perceived realism and thus improve the validity of the results, future work should include realistic movements and environmental factors.
Finally, although the experimental conditions were similar across the three datasets, some differences remained, which may have influenced perceived realism in VR driving simulations.The present study aimed to identify a relationship between perceived realism and psychophysiological metrics rather than the influence of situational factors on perceived realism, which may not be critically influenced by the association between psychophysiological metrics and perceived realism.In future studies, it would be interesting to design an experiment to investigate how manipulated situational factors influence perceived realism in VR driving simulations.

VI. CONCLUSION
The present study used individual characteristics, personality traits, heart rate changes during the event, and risky decision-making during the event to predict perceived realism in VR driving simulations.Since making driving simulations realistic is important to narrow the gap between driving behavior in the real world and simulations, this study sheds light on the factors influencing perceived realism in VR driving simulations.First, this study revealed that Machiavellianism and psychopathy influence perceived realism, which suggests that when evaluating or investigating factors that may enhance perceived realism, researchers must balance the psychopathy and Machiavellianism of participants to objectively evaluate effectiveness.Second, this study showed that perceived realism influences risky decision-making during the event, which also suggests that when designing a simulation study to investigate drivers' decision-making in an accident situation, perceived realism must be controlled for to validate the results of driving simulations.Third, this study found that tree-based models were better for predicting perceived realism when including offline data or the individual characteristics of drivers, which can used as a reference for accurately modeling perceived realism in driving simulations.Additionally, when including only online features, linear regression models showed better performance, which suggested that when predicting perceived realism using only physiological data, linear regression models should be used instead of tree-based models.For example, if only using data on individuals' personality traits, VR simulation-based training can predict perceived realism in real time using tree-based models.However, if only online features such as heart rate changes are available, it would be better to use linear models to predict perceived realism.This can be applied to other VR elements such as VR exposure therapy, which can manipulate the intensity of exposure and provide personalized content from psychophysiological metrics that can be used to enhance realism and thus increase the effectiveness of therapy [79].

FIGURE 1 .
FIGURE 1. Screenshot of the event (A) pedestrians appear (B) trees fall.

FIGURE 2 .
FIGURE 2. Random forest-based regression analysis when including only significant offline features.

TABLE 1 .
Description of the datasets.

TABLE 3 .
Correlations between perceived realism and the psychophysiological metrics.

TABLE 6 .
Regression analysis for predicting perceived realism when excluding non-significant features.