Assessment of threat and negativity bias in virtual reality

Negativity bias, i.e., tendency to respond strongly to negative stimuli, can be captured via behavioural and psychophysiological responses to potential threat. A virtual environment (VE) was created at room-scale wherein participants traversed a grid of ice blocks placed 200 m above the ground. Threat was manipulated by increasing the probability of encountering ice blocks that disintegrated and led to a virtual fall. Participants interacted with the ice blocks via sensors placed on their feet. Thirty-four people were recruited for the study, who were divided into High (HN) and Low (LN) Neuroticism groups. Movement data were recorded alongside skin conductance level and facial electromyography from the corrugator supercilii and zygomaticus major. Risk-averse behaviours, such as standing on ‘safe’ blocks and testing blocks prior to movement, increased when threat was highest. HN individuals exhibited more risk-averse behaviour than the LN group, especially in the presence of high threat. In addition, activation of the corrugator muscle was higher for HN individuals in the period following a movement to an ice block. These findings are discussed with respect to the use of room-scale VE as a protocol for emotion induction and measuring trait differences in negativity bias within VR.


Scientific Reports
| (2020) 10:17338 | https://doi.org/10.1038/s41598-020-74421-1 www.nature.com/scientificreports/ this negativity bias across a range of experimental paradigms 41,42 . Heightened sensitivity to negative stimuli is a characteristic associated with increased trait neuroticism 43 and high neuroticism has also been associated with a 'harm avoidant' style of coping 44 , leading to exaggerated psychophysiological reactivity to increased threat 45 and negative images and films 46,47 . Hence, the accelerated gradient of negativity and its disproportionate impact on behaviour and psychophysiology may be augmented for those with high trait neuroticism who are predisposed towards negative affectivity.
The goal of the current study was to explore behavioural and psychophysiological responses to systematic increases of threat within a room-scale VE. Participants were required to negotiate a virtual variant on the visual cliff 48 where they must traverse a grid of large, translucent blocks of ice that floated in the air at a minimum height of 200 m. Participants negotiated three levels of the VE from low to high threat. Patterns of movement were recorded alongside ambulatory psychophysiology, e.g., fEMG, SCL. It was predicted that movement patterns would reflect greater caution and risk aversion as threat increased. It was also hypothesised that increased threat would provoke higher sympathetic activation in the SCL and greater negative valence (increased activation of corrugator and reduced activation of zygomaticus). It was predicted that participants with higher trait neuroticism would exhibit a greater tendency towards risk-averse behaviour and higher psychophysiological reactivity to threat (e.g., higher activation/negative valence).

Methods
Participants. Thirty-four participants (21 female) were recruited with a mean age of 23.72 years (SD = 3.15).
Participants were recruited from a university population of postgraduates and undergraduates. The only criterion for inclusion was that participants must be aged over 18 years and able to walk without assistance. Participants were excluded if they were currently taking any medication. Participants completed the OCEAN.20 personality inventory 49 and the Situational Vertigo Questionnaire (SVQ) questionnaire 50 . The OCEAN.20 inventory includes twenty items, five of which are each dedicated to each of the Big Five personality traits 51 , e.g., Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism. For the purpose of the current study, we only measured trait Neuroticism where participants provided responses on a 7-point Likert scale (1 = very strongly disagree; 0 = neutral; 7 = very strongly agree) to five statements, e.g., 'My feelings are easily hurt' 'I am often nervous and tense' . The reliability of this trait as measured on the OCEAN.20 and indexed by Cronbach's alpha was 0.81. The SVQ includes 19 items designed to capture visual vertigo with a Cronbach's alpha score of 0.96 52 ; participants are asked to rate the prevalence of vertigo symptoms on a 4-point scale from 0 (not at all) to 4 (very much) in a number of scenarios, e.g. 'riding as a passenger in a car on winding or bumpy roads' 'standing in a lift as it stops' .
Participants were divided into High Neuroticism (HN) and Low Neuroticism (LN) groups using a median split of 24; each group consisted of 16 participants (2 participants fell on the median and were omitted from further analyses). The HN group included 10 females with a mean age of 23.76 years (SD = 3.53), whereas the LN group contained 11 females with a mean age of 24.69 years (SD = 2.83). Between-group t-tests confirmed: (i) no significant differences in age, (ii) no significant differences in scores on the SVQ, i.e., HN group (M = 1. Virtual environment. Participants were required to traverse a grid of 4 × 4 ice blocks from a start platform to a goal platform where an automatic door could be activated (Fig. 1). Each block was approximately 70 × 70 cm. The ice blocks were suspended in the air at a virtual height of 200 m and participants interacted with them via foot movements, achieved by attaching sensors to participants' feet in addition to conventional handheld trackers (Fig. 2c). Foot sensors allowed participants to interact with ice blocks in two ways: (1) a one-footed testing movement (Risk Assessment) (Fig. 2a) or (2) a two-feet movement in which participants committed to standing on the block (Risk Decision) (Fig. 2b).
The grids of ice blocks contained three types of block. If the block was Solid, it would support the weight of the participant and did not change appearance when activated. Crack blocks also supported the weight of the participant but any interaction caused a change of colour from translucent to blue accompanied by a cracking sound effect 500 ms after activation (Fig. 2b). Fall blocks behaved in exactly the same fashion as Crack blocks, but any two-feet activation triggered a shattering sound effect after 500 ms when the block would disintegrate and participants experienced a virtual fall (a video of the VE is provided in the supplementary materials). In the event of a fall, participants were required to return to the start position and repeat their journey across the grid; a gap would appear in the grid to indicate the former position of the Fall block.
Participants were required to traverse three different grids or Levels during the experiment. The level of threat was manipulated by increasing the number of Crack and Fall blocks from Level 1 to Level 3 (Fig. 3). In order to complete each Level, participants activated an automatic door ( Fig. 1) with one of the handheld controllers. At the end of level 1, participants would be instructed to turn and return to the start location for level 2. This process was repeated at the end of level 2.
Virtual reality system. The experiment was conducted within a physical space of 5 × 4 m. Participants wore a tethered HTC Vive head-mounted display (HMD) with two base stations positioned in diagonally-opposite-corners of the space. Each participant held two hand controllers and two trackers were attached to their feet. Hand and feet positions were represented as white luminous outlines in the VE (Fig. 2a, b) www.nature.com/scientificreports/  www.nature.com/scientificreports/ constructed in Unreal Engine 4.21. All assets were purpose built for the study. The VE was rendered on a desktop PC with custom C++ code integrated directly into the Unreal Engine system to capture interactions with blocks and recorded timings. Data representing position of the head, hands and feet were recorded at a rate of 5 Hz and logged to a text file.

Movement. The primary mode of interaction within the VE was achieved via foot movements. By monitor-
ing the position of each foot and interaction with custom trigger volumes in the VE, the frequency of one-footed and two-footed movements were recorded and cross-registered with a position within the VE. Both frequency and timings of all interactions were recorded, e.g., number of one-footed or two-feet interactions with Solid, Crack and Fall blocks. Therefore, we could calculate the frequency of Risk Assessment (one-footed movement) and Risk Decision (two-feet movement) interactions with each block type. Timing data were also logged, e.g., amount of time spent standing on each block.
Event-related psychophysiology. Facial electromyography (fEMG) was recorded at 1000 Hz from the corrugator supercilii and zygomaticus major via the Bionomadix ambulatory psychophysiology system (BIOPAC). fEMG data were processed as follows: (1) bandpass filter applied at 49-51 Hz to remove 50 Hz noise, (2) filtered between 20 and 400 Hz, (3) rectified and smoothed via linear envelope (9 Hz filter), and (4) subjected to root mean square transformation 49,50 . Skin Conductance Level (SCL) was also recorded at 1000 Hz via the Bionomadix system. Data were collected from the index finger and second digit of the non-dominant hand and subsequently filtered with a high pass filter at 0.05 Hz. The level of activation from all psychophysiological data were calculated on an event-related basis. For each one-footed and two-feet interaction with a Solid or Crack block, psychophysiological data were averaged during a pre-movement period of 750 ms and a post-movement period of 1500 ms duration. These data were averaged across all events and baselined (i.e., pre-movement values were subtracted from post-movement values) create Grand Averages for the 0-750 ms and 751-1500 ms periods following each movement for: (1) each category of interaction (Risk Assessment/Risk Decision), and (2) two types of Block (Solid vs. Crack). In order to maximise the number of interactions that contributed to each Grand Average, the level of the VE was omitted from the analyses of psychophysiological data.
Procedure. Participants arrived at the laboratory, they read a Participant Information Sheet and provided informed consent. Participants completed the questionnaires and received written instructions about the task and the VE. The fEMG sensors were attached to the corrugator and zygomaticus sites and SCL sensors taped to the second phalanx of fingers on the non-dominant hand. Participants were subsequently fitted with the Vive Tracker sensors on their feet, the HMD and the hand controllers. The VE for Level 1 was activated and participants were provided with a 30 s countdown while standing on the platform before the grid appeared. Once the grid was available, participants could progress at their own speed. When they reached the task goal (the automatic door), participants received a transition cue to return to the starting position for Level 2 and the process was repeated. At the end of Level 2, the same procedure was followed for Level 3. Upon completion of Level 3, the VR apparatus and psychophysiological sensors were removed and participants were thanked for their time and debriefed.
Hypotheses and statistical analyses. It was hypothesised that movement would be increasingly riskaverse as threat was increased from Level 1 to Level 3, e.g., greater frequency of Risk Assessments, longer time  Psychophysiological data were collected on an event-related basis for each type of interaction. It was hypothesised that highest levels of corrugator and SCL activation plus lower levels of zygomaticus activity would be observed after two-footed interactions with Crack blocks, which contained the greatest potential for a fall. It was also predicted that this pattern would be enhanced for HN individuals compared to those in the LN group. Each psychophysiological grand average was analysed via a 2 (HN vs. LN) × 2 (Risk Assessment vs. Decision) × 2 (Solid vs. Crack Block) × 2 (post-movement periods: 0-750 ms vs. 751-1500 ms) ANOVA.
All statistical analyses were conducted using SPSS v.26. Outliers were defined as any value that deviated from the mean for that cell by 3 or more standard deviations. In the case of within-participants comparisons, sphericity was assessed via Mauchly's Test and the Greenhouse-Geisser adjustment was applied.

Results
Movement. The analyses of participants' movements were based exclusively on interactions with Solid and Crack blocks. Fall blocks were included simply to increase perceptions of threat during Level 2 (L2) and Level 3 (L3) in the VE (Fig. 1). Twenty out of our 34 participants triggered at least one Fall block at either L2 or L3.  (Fig. 4b). In summary, when the threat level of the VE was highest, participants spent longer standing on Solid compared to Crack blocks and this effect was particularly pronounced for HN individuals.   Fig. 6b. In summary, the frequency of Risk Assessments (testing blocks with one foot) increased with the threat level of VE and this movement was associated with higher trait neuroticism at L3. In addition, individuals in the HN group interacted more frequently with Solid blocks, particularly at L1.
Skin conductance level (SCL). SCL data were lost from three participants due to physical artifacts caused by contact between the sensor and the handheld controller. No statistically significant effects were found during analyses of SCL during Risk Assessment. The analyses of SCL activation during Risk Decision revealed only a  The marginal Block effect suggested that stepping onto a Crack block with two-feet caused higher activation of the corrugator muscle compared to the same movement to a Solid block. The interaction between Group x Period is illustrated in Fig. 7a. Post-hoc t-tests revealed that corrugator activation was significantly higher during 751-1500 ms period following a two-feet interaction compared to the earlier period, but only for the HN group  Fig. 7b. Post-hoc t-tests revealed significant between-group differences for both 0-750 ms [t(31) = 1.99, p = 0.051] and 751-1500 ms periods after movement occurred [t(31) = − 2.02, p = 0.052], i.e., greater zygomaticus activation during the 0-750 ms period for the LN group and greater activation for the HN group during the 751 = 1500 ms period (Fig. 7b). The same model was applied to zygomaticus data during Risk Decision, but no significant effects were found.
To summarise the fEMG analyses, activation of the corrugator muscle was associated with Risk Decision interactions, especially to Crack blocks when the risk of a virtual fall was highest. Individuals with high trait neuroticism tended to exhibit greater corrugator activation in the 715-1500 ms period after this interaction. When participants made a Risk Assessment, we found a pattern of individual differences wherein HN individuals showed greater activation of the zygomaticus during the period 751-1500 ms after one-footed movement compared to the LN group, this pattern was reversed during the 0-750 ms period.

Discussion
Our analyses of behavioural data revealed evidence of risk aversion in response to increased threat and trait neuroticism, specifically: • Participants spent longer standing on Solid vs. Crack blocks because the former were deemed to be 'safe' and this effect was pronounced at L3 (Fig. 4a). • The number of Risk Assessments exceeded the frequency of Risk Decisions as participants progressed from L1 to L3 (Fig. 5). This pattern indicated greater caution as one-footed checks were used to identify Crack blocks. • This pattern of risk-averse behaviour was enhanced for participants with higher trait neuroticism: (i) HN individuals spent longer standing on Solid blocks during L3 (Fig. 4b), (ii) participants in the HN group made more frequent Risk Assessments at L3 (Fig. 6a), and (iii) when Solid blocks were available at L1, HN individuals interacted more frequently with those blocks (Fig. 6b). www.nature.com/scientificreports/ A perception of increased threat increased activation of the negativity subsystem 29,30 , which created a compensatory preference for risk-averse behaviour, e.g., one-foot checks, dwelling on Solid blocks.
The influence of threat on psychophysiological measures was assessed on an event-related basis. For these analyses, threat was highest when participants made a two-feet movement to a Crack block, which was visually indistinguishable from a Fall block. We found evidence of increased corrugator activity following a two-feet movement to Crack blocks and our analyses revealed that HN individuals exhibited greater corrugator reactivity following a two feet interaction to any block (Fig. 7a); this pattern may indicate greater negative affect for those individuals in response to uncertainty 34 and anticipation of a virtual fall. It was also predicted that sympathetic activation would increase via elevated SCL during these high-threat interactions 44 , but this hypothesis was not supported by our data.
An interaction between trait neuroticism and period was apparent for zygomaticus activation during Risk Assessment (Fig. 7b). Interpretation of this effect is problematic because at least two lines of inference are possible. If positive affect is synonymous with activation of this muscle 35 , greater zygomaticus reactivity indicates positive anticipation for LN individuals, i.e., that the block would not crack. However, the reversal of this trend during the 751-1500 ms post-movement period (Fig. 7b) could indicate a positive response (e.g., relief) for HN individuals, when the block did not crack. Alternatively, increased zygomaticus activity could be interpreted as a grimace associated with negative affect 51 for both LN and HN individuals. The absence of any differentiation between block type makes it impossible to interpret this effect with greater specificity.
With respect to previous research, our behavioural analyses replicated the association between negative emotional responses to threat and avoidant behaviour previously reported by Biedermann et al 20 . The absence of any significant increase of autonomic activation via SCL in the current study contrasts with earlier research 16,19,22 where virtual height provoked greater SCL. However, the current study differed from existing work in a fundamental way and the lack of replication is unsurprising. The previous studies were designed to compare height with no height 16 or different levels of height 19 or being virtually elevated in an open vs. closed elevator 22 whereas the experience of virtual height was held constant in the current investigation and our primary independent variables was the probability of a fall.
The results of the study can be summarized as follows: (i) room-scale VE allowed us to induce negative emotions that were seamlessly associated with naturalistic behavioural responses, (ii) by cross-referencing data, event-related psychophysiology could index momentary fluctuations following each interaction in the VE, and (iii) this methodology was utilized to achieve a graded activation of the negativity subsystem of the ESM. With respect to the latter, we demonstrated how threat induced compensatory behaviours to mitigate risk and increased negative affect during higher-risk interactions. Both patterns were enhanced for participants with higher trait neuroticism who had an implicit bias towards negative affectivity.
There were a number of areas where the methodology could be improved. A decision to maximise the autonomy of participants within the VE allowed individuals to 'backtrack' over previously activated blocks, which inflated the amount of interaction data derived from those participants; in hindsight, a mechanic should have been introduced to prevent backtracking. We also allowed participants to self-select speed of movement through the VE, which forced us to minimize time windows during the event-based analyses in order to eliminate overlap between successive post-movement and pre-movement periods. In the case of some variables, such as SCL, the 1500 ms post-event period may have blunted the sensitivity of the measure, i.e. SCL latency is 1-3s 52 .
The study demonstrated how room-scale VR can create a controlled environment capable of inducing embodied emotional experiences, which incorporate feedback from body posture as input to the emotional state 53,54 . At a more practical level, our results imply that behavioural data can be utilized to profile individual users of VR, in an implicit fashion similar to existing work utilising data from interactions on social media 55 , which has implications for data privacy as commercial VR systems are used at scale within the general population. www.nature.com/scientificreports/