Application of immersive virtual reality for assessing chronic neglect in individuals with stroke: the immersive virtual road-crossing task

ABSTRACT Background Neglect can be a long-term consequence of chronic stroke that can impede an individual’s ability to perform daily activities, but chronic and discrete forms can be difficult to detect. We developed and evaluated the “immersive virtual road-crossing task” (iVRoad) to identify and quantify discrete neglect symptoms in chronic stroke patients. Method The iVRoad task requires crossing virtual intersections and placing a letter in a mailbox placed either on the left or right. We tested three groups using the HTC Vive Pro Eye: (1) chronic right hemisphere stroke patients with (N = 20) and (2) without (N = 20) chronic left-sided neglect, and (3) age and gender-matched healthy controls (N = 20). We analyzed temporal parameters, errors, and head rotation to identify group-specific patterns, and applied questionnaires to measure self-assessed pedestrian behavior and usability. Results Overall, the task was well-tolerated by all participants with fewer cybersickness-induced symptoms after the VR exposure than before. Reaction time, left-sided errors, and lateral head movements for traffic from left most clearly distinguished between groups. Neglect patients committed more dangerous crossings, but their self-rated pedestrian behavior did not differ from that of stroke patients without neglect. This demonstrates their reduced awareness of the risks in everyday life and highlights the clinical relevance of the task. Conclusions Our findings suggest that a virtual road crossing task, such as iVRoad, has the potential to identify subtle symptoms of neglect by providing virtual scenarios that more closely resemble the demands and challenges of everyday life. iVRoad is an immersive, naturalistic virtual reality task that can measure clinically relevant behavioral variance and identify discrete neglect symptoms.


Introduction
Selectively focusing and dividing our attention are fundamental cognitive processes that are crucial for everyday tasks.For instance, pedestrians must be able to process unexpected and rapidly changing situations when crossing the road.When relevant information is not processed accurately, for example due to distractions, crossing roads can become challenging and potentially dangerous, as seen in individuals with neglect (Corbetta & Shulman, 2011;Heilman et al., 2000;Vallar, 2001).
Neglect is a disabling neurocognitive disorder that frequently occurs after stroke and is characterized by a lack of attention to the contralesional side (Husain, 2008;Karnath, 2015).This systematic shift in the perceived visual horizontal axis toward the unaffected side occurs more frequently and severely after right-hemispheric lesions, thereby affecting the left hemifield (Karnath & Rorden, 2012;Ringman et al., 2004).The lateral orientation bias associated with neglect is not due to deficits in the primary sensory or motor systems (Heilman et al., 1994), but the result of dysfunction in higher level cognitive and attentional processes (Bartolomeo & Chokron, 2002;Karnath, 2015).Furthermore, neglect adversely has been associated with prolonged hospitalization, increased care needs, and reduced functional outcomes (Jehkonen et al., 2006;Kerkhoff & Schenk, 2012).While spontaneous remission can occur in the first post-stroke weeks (Farnè et al., 2004), later recovery is more rare (Nijboer et al., 2013).Given the distinct subtypes and temporal variations (Azouvi et al., 2002;Bowen et al., 1999), neglect is a complex supramodal syndrome (Parton et al., 2004).Clinical neglect assessment tools typically consist of multiple, varying bedside paper-and-pencil tests, such as cancellation, drawing, and line bisection tasks (Azouvi et al., 2002;Moore et al., 2022).However, existing instruments have low sensitivity, particularly for subtle but clinically relevant neglect symptoms in chronic stroke (Barrett et al., 2006;Bonato, 2012;Buxbaum et al., 2012).As a result, discrete forms of neglect often go unrecognized and untreated, even though they may severely affect patients' everyday lives (Chen et al., 2012).Despite their extensive use in clinical practice, conventional tests do not resemble relevant everyday life tasks (verisimilitude) and have limited predictive power for the performance of daily tasks (veridicality) (Kaiser et al., 2022).The discrepancy between test scores and functional performance (Bonato, 2012) can be attributed to various clinical neglect manifestations (Aimola et al., 2012).Reasons for this discrepancy include repeated test administration, (partial) recovery, increased "on-the-spot alertness" (i.e., improvement of sustained attention by situational awareness of the environment and testing situations to respond appropriately) (Bonato, 2012, p. 3), and acquired compensatory strategies (Andres et al., 2019).
The prevalence of neglect following stroke is highly variable, with reported frequencies ranging from 20% to 80% (Esposito et al., 2021).However, neglect symptoms are often under-reported, as demonstrated by test sensitivities from as low as 13% to 82% (Bowen et al., 1999).Therefore, it is hardly surprising that the severity and frequency of neglect in clinical practice and daily life are higher than currently recognized (Buxbaum et al., 2004;Edwards et al., 2006).
While paper-and-pencil tests are two-dimensional, static, and possess low verisimilitude, everyday tasks are typically three-dimensional, dynamic, and interactive.Overcoming the inherent limitations of traditional neglect tests therefore requires tasks that resemble reallife demands and involve a synergy of motor, postural, visual, and cognitive-perceptual skills (Azouvi et al., 2002;Brink et al., 2017;Simpson et al., 2003).Yet, assessing spatial deficits in neglect can be challenging, as an adequate assessment must preclude the possibility of implementing compensatory strategies for the deficits (Bonato, 2012) and elicit lateralized sustained attention and fatigue (Kerkhoff et al., 2020).Other task-related factors like complexity (Aravind & Lamontagne, 2017;Villarreal et al., 2020), varying levels of cognitive demand (Bonato, 2012;Spreij et al., 2020), divided attention (Buxbaum et al., 2008), and stimulus density (Brink et al., 2020;Nijboer & van der Stigchel, 2019) need to be considered to sensitively assess chronic neglect.For a more precise assessment of neglect, it is relevant to incorporate behavioral parameters reflecting everyday behavioral and functional performance.These may include measures of eye and head movements (Karnath et al., 1998) and temporal parameters such as reaction times (Deouell et al., 2005;Rengachary et al., 2009), total time (Jannink et al., 2009), and time constraints (Kwon et al., 2020).
The application of immersive Virtual Reality (VR) has great potential to effectively detect discrete neglect symptoms by presenting tasks that closely resemble real-life scenarios and better target functional performance.In VR, realistic and interactive virtual content with high ecological validity can be presented threedimensionally and dynamically.The vast majority of VR research on neglect has involved either: (1) conventional paper-and-pencil tests converted into immersive tasks (Gupta et al., 2000;Knobel et al., 2020;Tanaka et al., 2005) or non-immersive tasks (Broeren et al., 2007;Fordell et al., 2011) or (2) immersive applications with relevance to everyday life activities.The latter included tasks for navigation (Ogourtsova et al., 2018;Peskine et al., 2011), or detection and searching (Jannink et al., 2009;Kim et al., 2010;Knobel et al., 2021;Ogourtsova et al., 2018;Yasuda et al., 2018).However, previous studies have been limited by small sample sizes and have not included post-stroke patients with subtle neglect symptoms or a clinical control group (Pedroli et al., 2015).The use of head-mounted displays (HMDs) may induce motion sickness, known as cybersickness, manifested as symptoms of nausea, headache, and general discomfort.It is therefore important to have valid self-assessment tools and symptom insights to understand and mitigate the adverse effects of virtual experiences.
Pedestrian accidents at road crossings are a significant global issue, particularly affecting older adults and vulnerable populations with cognitive impairment due to stroke or other neurological disorders (Oxley et al., 2005;Wilmut & Purcell, 2022).Road-crossing scenarios impose high demands on multiple cognitive functions (e.g., attention), visual perceptual processing (e.g., speed, distance perception), and physical abilities (Oxley et al., 2005).This is especially relevant as subjects with neglect are at high risk of falls (Ugur et al., 2000) and frequently collide with contralateral obstacles (Aravind & Lamontagne, 2017).In addition, extensive evidence suggests that healthy elderly pedestrians exhibit riskier crossings, in higher speed traffic, due to poorer distance and speed estimation (Papić et al., 2020;Petzoldt, 2014), longer decision times (Lobjois et al., 2013;Oxley et al., 1999), and longer road-crossing times than younger individuals (Dommes et al., 2014;Holland & Hill, 2010).Taken together, a virtual road-crossing task is highly relevant as the risks of elderly pedestrians with neglect accumulate, resulting in a higher risk of traffic injury.Thus, such a road-crossing task would be highly appropriate for clinical assessment.A virtual road-crossing task, as a diagnostic tool for neglect, has the potential to bridge the gap between neglect task demands and clinical experience.This would allow patient behavior to be examined and quantified in a highly controlled and ecologically valid, yet safe, way.Compensatory strategies, as observed in paper-and-pencil tests, can be overcome given the high relevance of road crossing to everyday life, as well as high cognitive demand, and multimodal, dynamic nature (Kim et al., 2010;Peskine et al., 2011;Wu et al., 2018).
There are various virtual road-crossing tasks documented in the literature.Non-immersive (Broeren et al., 2007;Fordell et al., 2011) and immersive road-crossing tasks have either examined the behavior of children (Morrongiello et al., 2015;Ridene et al., 2015;Simpson et al., 2003), adolescents (Cherix et al., 2020;Clancy et al., 2006;Kaimara et al., 2021) or healthy adults (Anthes et al., 2016;Deb et al., 2017;Feldstein & Dyszak, 2020).However, this leaves a gap in the evaluation of these tasks for neurological patients, particularly those with neglect.Only a few studies have applied virtual roadcrossing tasks for the assessment of neglect.For instance, Mesa-Gresa et al. (2011) used the non-immersive VR Street Crossing Test with five subjects with neglect following acquired brain injury.They showed that neglect patients had a higher number of accidents compared to no neglect patients while the total time taken for the VR task did not differentiate the two groups.In addition, Kim et al. (2010) employed an immersive virtual street scenario with four levels of car speed and one lane to assess neglect in far space.For traffic from the left, neglect patients exhibited significantly higher reaction times and failure rates, and an increased deviation angle to the left compared to no neglect patients.These studies indicate the potential of virtual tasks for neglect detection but suffer from considerable limitations.Specifically, studies in this domain often exhibit issues such as artificial graphics, insufficient validation, and low verisimilitude (Kim et al., 2010), or are restricted by small sample sizes without clinical control groups (Peskine et al., 2011).Although they provide promising results, a more sensitive and virtual road-crossing task with an appropriate study design is needed to assess neglect.
The primary objective of this study was to develop a clinically relevant paradigm involving a virtual roadcrossing task and evaluate its sensitivity for subtle neglect symptoms in a sample of right-hemisphere post-stroke patients with and without neglect, and healthy controls.This is the first step toward the systematic evaluation of novel VR measures to improve neuropsychological assessment of spatial neglect.We aimed to identify relevant behavioral patterns and movement parameters to distinguish between neglecttypical and other behavior.Given the lateral bias in neglect, we examined the performance for different traffic directions.Our secondary objective was to demonstrate feasibility and usability from patients' perspective and acceptability as we applied the VR task to a clinical population.Finally, a typical characteristic of neglect is poor self-assessment and a lack of insight into the syndrome (Grattan et al., 2018).Therefore, we expected that the self-rated behavior in patients with neglect to be less related to performance in iVRoad compared to those without neglect.

Participants
A total of N = 60 participants, all of whom identified as White, divided into three groups of n = 20 each, took part in this study (cf.Table 1): (1) Chronic right-hemisphere post-stroke patients with chronic left-sided spatial neglect as the experimental sample (USN+), (2) chronic right-hemisphere post-stroke patients without neglect as the clinical control group (USN-), and (3) healthy participants as the healthy control group (HC).We used a matched-pair case-control design as cases and controls were selected based on matching medical history (stroke/no stroke, neglect/no neglect), age, and gender.
Participants in the clinical groups were recruited between September 2020 and March 2022 via the databases of the Clinic for Cognitive Neurology, University Hospital Leipzig, Germany, and the Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany (MPI-CBS).Eligible participants received an invitation letter and were contacted by telephone, while current patients were recruited during their hospital stay.We classified clinically eligible participants as members of the USN+ group based on three criteria: (1) a history of right-hemisphere stroke and left hemispatial neglect in the acute phase, (2) a confirmed diagnosis of neglect in the most recent medical report.Additionally, (3) neglect typical patterns, clinical assessments, or clinical observations had to be confirmed by neurologists, neuropsychologists, or orthoptists.Neglect symptoms were classified as discrete if the majority of test results in the paper-and-pencil tests were inconspicuous, but isolated tests and behavioral observations indicate the presence of neglect.Participants were classified as belonging to the USN-group if they had a right-hemisphere stroke and showed no evidence of neglect, as determined by both the latest medical records and observations made during the administration of the study.
Healthy controls were recruited through the MPI-CBS and were included if they had normal or corrected-to-normal vision, and no history of neurological or psychiatric disease.The experiment lasted approximately three hours and all participants received reimbursement for their participation.The study followed the Declaration of Helsinki and was approved by the Ethical Committee at the Medical Faculty of Leipzig University (Ethics code: 117/18-lk, March 5th, 2020) and was preregistered (DOI: 10.17605/OSF.IO/ M9CHU).

Materials and procedure
We used questionnaires, a conventional neuropsychological test battery with various neglect and attention tests, and behavior recorded in the VR task.The quality of the measurements was ensured by training of the experimenters to collect the VR data reliably.Certified neuropsychologists and orthoptists trained the staff in the administration of the neuropsychological tests.

Clinical assessment
The administered neuropsychological test battery evaluated a range of cognitive functions, including global cognition, visuospatial memory, and attention.Neglect assessment tools included the Neglect-Test (NET; Fels & Geissner, i.e., the German version of the Behavioural Inattention Test (BIT; Wilson et al., 1987), the Sensitive Neglect Test (SNT; Reinhart et al., 2016) in both single and dual-task versions, and the Catherine Bergego Scale as self-reported test (CBS; Azouvi, 1996) (Table 2).

Questionnaires
Custom questionnaires were used with regards to the feasibility and usability, using a combination of existing and newly created or adapted items.We assessed cybersickness, user experience, level of presence, motivation, and overall impression of the VR task, as well as the selfassessed pedestrian behavior in real-life situations.These questionnaires were administered on a computer to the examiner and the participants via a study website, both before and after the VR task.

Immersive virtual road crossing task (iVRoad)
A summary of the procedure for the virtual road crossing task is shown in Figure 1.All subjects participated in the study in a seated position.After setting up the HMD and adjusting for the height in the virtual environment,  15) -Significant group differences are asterisked: *indicates p < 0.05, **indicates p < 0.01, ***indicates p < 0.001.a Marks significant differences between USN+ and USN-, b marks significant differences between USN+ and HC.USN+ = Leftsided neglect after right-hemisphere stroke; USN-= No left-sided neglect after right-hemisphere stroke; HC = Healthy controls; CBS = Catherine Bergego Scale (Azouvi, 1996); MoCA = Montreal Cognitive Assessment Test (Nasreddine et al., 2005); VR = Virtual Reality; Years of education: "Low" corresponds to fewer than 10 years of education, including both school and higher education; "Mid" corresponds to 11-14 years; and "High" corresponds to 15 years or more; transportation per week scale: 1 ("never") to 5 ("daily").Technology use scale: 1 ("daily") to 5 ("rarely").
all participants completed a tutorial to familiarize them with the environment, the task (i.e., crossing the road, letter insertion) and the technical equipment (i.e., using the controller and exploring the virtual environment).
Instructions were delivered via headphones and displayed on additional information boards in the virtual environment.
The study took place in a realistic, virtual urban environment, where participants were required to safely cross two parallel, heavily traveled lanes of a road of 300 m in length to each side.The task was to observe the lanes (i.e., watching the cars), cross them safely, post a letter in a mailbox, and return by crossing the street again (see video https://owncloud.gwdg.de/index.php/s/o1RK8KbGlfL1JMn).The experiment consisted of 24 trials, divided into six blocks of four trials each.To increase the level of immersion and minimize the risk of cybersickness, the task was mainly viewed from a firstperson perspective to provide a higher sense of embodiment (Gorisse et al., 2017) until the decision was made to cross the road.When participants pressed the controller button to cross the street, they saw an avatar, representing themselves, crossing the street.This was presented from the third-person perspective to minimize the risk of cybersickness (Evin et al., 2020).The avatar's walking speed of 1.45 m/s, which was the average speed of pedestrians in Germany (Morgenroth, 2008), was not adjustable.We also introduced visual distractors, such as virtual humans or birds in the city scenario, and auditory distractors, such as barking dogs or shattering glass during the road crossing to enhance resemblance to real-life scenarios.

Experimental devices and apparatus
The VR task was carried out in a 4 × 4 m room with a swivel chair positioned in the center.We used the HTC Vive Pro Eye, an HMD that features an integrated eye-tracking system, a field of view of 110 degrees and a resolution of 1440 × 1600 pixels per eye.The HMD was connected to a portable workstation with an Intel i7-9700 processor, 16 GB of RAM, and an Nvidia GeForce RTX 2070 graphics card.To interact with the virtual environment (i.e., decide to cross the road or complete the tutorial), participants had to press a button on the back of a controller operated by their dominant hand.For more information on the task development and technical implementation, see Wagner et al. (2021).

Measured iVRoad outcomes
We manipulated variables related to car speed (30 km/h, 50 km/h), gap size (6.5 s, 7.5 s), traffic direction in both lanes (traffic from left, two-way traffic, traffic from right), and the side of the mailbox (left side, right side).A logical adjustment was made to the street scene (e.g., traffic from left automatically became traffic from right with the same speed upon return).Outcome measures included (1) temporal parameters (total experiment time, reaction time, letter insertion time), (2) error patterns (number of errors, lane-and siderelated errors, error types, and crossing safety), and (3) head rotation (head poses including head tilt angle along the roll axis and both heading position and head turns along the yaw axis).For error types, we measured low-risk, high-risk, and dangerous errors.In addition, participants rated their difficulty in crossing real roads, both since the stroke and in general.For detailed information on outcome measures, see Table S1.
tidyverse for data pre-processing (Wickham et al., 2019), emmeans (Lenth, 2023) for comparisons and contrasts, afex (Singmann et al., 2022) for mixeddesign ANOVAs, and MASS (Venables et al., 2002) for negative binomial regression.Descriptive statistics and a p < .05criterion for significance were applied.Post-hoc analysis used Tukey's HSD, with Cohen's d for effect sizes and 95% confidence intervals (CI).Data were excluded if participants failed to comprehend the task or comply with experimenter instructions as determined by the therapist.A total of 0.69% (i.e., 10 trials) of the data were missing, with three of these trials missing due to technical issues and not at random.

Crossing safety and self-assessed pedestrian behaviour
The effect of Group and Traffic Direction on crossing safety was evaluated using binomial logistic regression.Predicted probabilities and ORs were calculated.
Pearson correlation analyses were used to examine the relationship between self-assessed pedestrian difficulty and number of errors, safe crossing, and high-risk errors.Detecting neglect and group differences Temporal parameters.Mixed-design ANOVAs were used to examine the effects of the independent variable Group on three dependent variables of the temporal parameters: total experiment time, average reaction time, and letter insertion time.Specifically, we employed Group as the between-subjects factor, and Traffic Direction (for average reaction time) or Mailbox Side (for letter insertion time) as the within-subjects factors.
Error pattern.Error analysis involved fitting a Poisson regression model, or a negative binomial regression model in cases of overdispersion, by comparing the mean and the variance.The independent variables for the total number of errors were Group and Traffic Direction.Logistic regression models were used for specific error types and predicting performance probabilities.Binomial and multinomial logistic regression analyses were carried out in an interaction model with the factors Group and Traffic Direction (with USN+ and traffic from left as the reference) to determine their influence on the type of error.Results are presented as Odds Ratios (OR) with 95% CIs, standard errors, and p-values.
Head rotation.We tracked the user's head position on the vertical axis (leftward and rightward lateralization and number of head turns along the yaw axis), and on the longitudinal axis (tilting the head to the left and right shoulder)) before road crossing.Head rotation data was analyzed using mixed-design ANOVAs, with Group as the between-subjects factor, and Traffic Direction as the within-subjects factor.

Applicability and feasibility
We assessed cybersickness symptoms before and after the participants were immersed in iVRoad using the 16-item Simulator Sickness Questionnaire (SSQ; Kennedy et al., 1993) Mixed-design ANOVAs were computed separately for the three SSQ subscores (Nausea, Oculomotor, Disorientation), and the total SSQ score with Group as the between-subjects factor and Measurement Time as the withinsubjects factor.We applied two usability tests: System Usability Scale (SUS; Brooke, 1996), and User Satisfaction Evaluation Questionnaire (USEQ; Gil-Gómez et al., 2017).Presence was assessed using the Igroup Presence Questionnaire (IPQ; Schubert et al., 2001).Group differences were analyzed using linear regression models with Group, the respective item as independent variables and the score as dependent variable.

Detecting neglect and group differences
A comprehensive overview of the results on error patterns, temporal parameters, and head rotation measures in the iVRoad task is presented in Figure 2.

Crossing safety and self-assessed pedestrian behaviour
For successful trials, the USN+ group exhibited the lowest absolute and relative frequency of safe road-crossings (N = 168, 67%) compared to USN-(N = 259, 78%) and HC (N = 322, 81%).Binomial logistic regression with the Group*Traffic Direction interaction to predict safe roadcrossings showed a significant effect only for Group, χ 2 (2, N = 60) = 7.40, p = .02.Predicted probabilities for safe crossings were 68% for USN+, and 85% for both USN-and HC.We found significant differences in the probabilities of safe road-crossing between the groups, with USN+ having 49% and 56% fewer safe crossings than USN-and HC, respectively.For traffic from left, USN+ showed a 62% (SE = 0.15, z = −2.37,p = .047)and 61% (SE = 0.15, z = −2.38,p = .045)lower likelihood of safe crossing compared to USN-and HC.Similar trends were obtained for traffic from right, with ORs of 0.41 for USN+ vs USN-(SE = 0.16, z = −2.31,p = .054),and 0.28 for USN+ vs HC (SE = 0.11, z = −3.34,p = .002).There were no significant group differences in safe crossings for two-way traffic.

Error pattern
Descriptive statistics for the number of errors associated with traffic direction are shown in Table 3.Total number of errors.Negative binomial regression results indicate that Group and Traffic Direction accounted for a significant amount of variance in the outcome, χ 2 (2) = 53.39,p < .001and χ 2 (2) = 41.49,p < .001,respectively.We found a significant interaction between Group and Traffic Direction.χ 2 (4) = 14.00, p = .007.Overall, the USN+ group made 1.50 (p = .005)times more errors than USN-and 3.33 (p < .001)times more than HC.Compared to two-way traffic, the number of errors was 56% lower for traffic from the left (p < .001)and 58% lower for traffic from the right (p < .001).For traffic from the left, USN+ committed more errors compared to USN-(OR = 2.17, p = .003)and HC (OR = 5.85, p < .001).This difference, however, remained significant only between USN+ and HC for traffic from the right (OR = 3.57, p < .001)and two-way traffic (OR = 1.76, p < .01).
Error type.Multinomial logistic regression analysis on error type, including low-risk, high-risk, and dangerous errors showed a significant interaction between Group and Traffic Direction in predicting error types, χ 2 (8) = 36.81,p < .001.Most errors were high-risk errors (N = 324), followed by dangerous errors (N = 98), and low-risk errors (N = 35).Dangerous errors.Dangerous errors did not differ significantly between groups.However, two-way traffic resulted in 57% fewer errors than traffic from left (p = .04)and 69% fewer than traffic from right (p = .002).USN+ made twice as many dangerous errors for traffic from left than USN, but this was not significant (p = .591).For two-way traffic, USN+ made 17 times more dangerous errors than USN-(SE = 13, t = 3.79, p < .004)and 8 times more dangerous errors than HC (SE = 5, t = 3.32, p = .01).
Low-risk errors.HC had very few low-risk errors in all traffic direction conditions (traffic from left: N = 0; traffic from right: N = 0; two-way traffic: N = 2), leading to high ORs compared to both groups, especially for traffic from left (e.g., USN+ vs HC: 799421) and traffic from right (e.g., USN+ vs HC = 543794317).USN+ had a significantly higher likelihood (10.49times) of making low-risk errors than HC (SE = 7.71, t = 3.19, p = .01),and 2.36 times higher than USN-, although significantly (SE = 1.07, t = 1.88, p = .170).The odds of USN+ for low-risk errors for two-way traffic were five times higher than for USN-(SE = 4, t = 2.10, p = .117)and four times higher than for HC (SE = 3, t = 1.68, p = .238),both of which were not significant.
High-risk errors.Regarding high-risk errors, the strongest effect was observed in two-way traffic.USN+ made significantly fewer high-risk errors in this condition compared to USN-(OR = 0, t = −4.79.p < .001)and HC (OR = 0, t = 4.05, p = .002).Overall, USN+ committed 50% fewer high-risk errors than USN-(SE = 0.12, t = −2.87,p = .03).No significant group differences were found in the ORs for traffic from left and right regarding high-risk errors.
Lane-related errors.Following right-hand traffic rules, cars from the left drive in the near lane and cars from the right drive in the far lane.A multinomial logistic regression, with performance as the outcome (success, near-lane error, far-lane error) and Group as the predictor, yielded a significant Group effect, χ 2 (4) = 110.4,p < .001.All groups made more far-lane than near-lane errors; however nearlane errors were predominantly observed in USN+.
Compared to USN-, the ORs for a near-lane error in USN+ was 2.25, and 5.59 compared to HC.Specifically, when an error occurred, the probabilities of an error in the far lane compared to the near-lane were for USN+ at 213% (SE = 0.59, t = 6.05, p = .002),for HC at 494% (SE = 1.89, t = 5.60, p = .003)and for USN-at 667% (SE = 1.89, t = 8.03, p < .001).A significant interaction between Group and Traffic Direction was found in the binomial logistic regression model with near-lane error as the dependent variable, χ 2 (4) = 25.67,p < .001.USN+ had a significantly higher likelihood for near-lane errors for two-way traffic than USN-and HC, with ORs of 21.46 (SE = 18.42, z = 3.57, p = .001)and 14.22 (SE = 12.19, z = 3.01, p = .005),respectively.The predicted probabilities for a near-lane error in two-way traffic were 27% for USN+, 2% for USN, and 3% for HC.

Cybersickness
All participants completed iVRoad without any terminations or reports of discomfort.Mixed-design ANOVAs indicated no VR-induced cybersickness across groups according to the SSQ subscales.On average, all groups reported fewer cybersickness-induced symptoms after the VR exposure than before it, with a significant decrease over time in the SSQ total score (β = 7.85, 95% CI [−0.38, 0.24], t(57) = 3.48, p < .001).

Discussion
The present study was conducted with neurological patients in the chronic phase after stroke.We aimed to develop and evaluate novel measures of an immersive virtual road-crossing paradigm (iVRoad) for the detection of discrete neglect symptoms.In this study, we measured and analyzed several VR behavioral parameters to identify neglect-specific behavioral patterns.Our results revealed that patients with chronic neglect exhibited general differences in temporal measures (i.e., higher total experiment time, reaction time, and letter insertion time).Additionally, there were neglect-typical side-related differences in patterns of successful and erroneous crossings (i.e., higher number of errors overall and in relation to Traffic Direction, specific error types, lane-and side related errors).Furthermore, differences were observed in head rotation measures (i.e., heading position, head tilt angle, and head turns).

Patients with chronic neglect make more and distinct errors in road-crossing
The presence of a chronic spatial neglect can be inferred based on the direction-specific error patterns made during road crossing.Chronic neglect patients committed more errors in general, as well as in neglect-specific error types, which may indicate spatial attentional and temporal deficits.As expected, the groups differed in the side on which errors occurred.While the healthy controls and the no neglect group had a higher incidence of errors with cars coming from the right, the neglect group had a higher proportion of errors on the left.Furthermore, they made fewer safe crossings, again, particularly for traffic from the left.
The traffic direction condition was a significant predictor of the number and type of errors.Neglect patients were more likely to commit errors for traffic from left, and they made more dangerous errors on the near lane (near-lane error) for two-way traffic.Both control groups made more high-risk errors on the far lane, particularly in two-way traffic, and near-lane errors hardly occurred.These results align with previous studies indicating that far-lane errors are most frequent in healthy adults, especially older pedestrians, in two-way traffic (Cavallo et al., 2019;Fontaine & Gourlet, 1997).Dommes et al. (2014) identified the most dangerous scenario for older pedestrians crossing a road with bidirectional traffic is a sufficient gap in the near lane and a short gap in the far lane.The neglect group, however, exhibited a higher incidence of both dangerous and low-risk errors when faced with traffic from the left.This suggests that they do not pay enough attention to the left, as evidenced by the decrease in heading to the left side and increased number of errors for left-sided traffic.Nevertheless, they seem to compensate for their deficit by making more head turns, which results in a late onset of the crossing.In addition, they are unable to calculate their own speed in relation to the approaching cars.These results highlight both general and specific attentional deficits in chronic neglect.

Temporal measures detect neglect-specific behaviour
An overall slower information processing speed was observed in the USN+ group, reflected in slower reaction times, higher total experimental time, and higher letter insertion times, particularly for left-side mailboxes.These results provide support for lateralized as well as non-lateralized attentional deficits in neglect and reflect behavioral delays, particularly in the left spatial hemispace.For instance, the traffic from left and twoway traffic conditions elicited longer reaction times in neglect patients compared to the two other two groups.This may suggest a reduced speed of information processing or, alternatively, compensatory behavior.Notably, patients with neglect exhibited a greater number of head turns than other groups, especially for traffic coming from the left.However, the reaction times within the USN+ group did not differ by traffic direction, demonstrating the potential of iVRoad to detect non-lateralized deficits associated with neglect (Ogourtsova et al., 2018;Villarreal et al., 2021).Letter insertion time also revealed spatial differences between the groups and within the chronic neglect group, as patients with neglect had higher letter insertion times in general, and especially for letters on the left side.In addition, the chronic neglect patients in our study exhibited non-lateralized attentional deficits, as indicated by increased total experimental time and overall reaction times, regardless of traffic direction.Compared to the USN-group, the longer reaction times observed for traffic from left and right suggest that neglect may be associated with slower processing on both sides of space, as previously reported (Buxbaum et al., 2004).This finding also suggests that deficits associated with neglect go beyond the lateralized attentional deficits as they are not restricted to one side (Corbetta & Shulman, 2011).These findings have important implications for the clinical assessment of chronic neglect, highlighting the need for a more comprehensive assessment of both spatial and non-spatial selective attention using temporal measures.
Patients with chronic neglect committed more lowrisk errors, indicating delayed responses in road-crossing initiation.For low-risk errors, the time between two approaching cars was sufficiently long for healthy controls to allow the road to be crossed safely.Thus, lowrisk errors in neglect subjects indicate that they started too late to cross the road in time.These results align with prior studies suggesting temporal measures as sensitive indicators of neglect (Schendel & Robertson, 2002).Reaction times have been used to measure attention in various tasks (e.g., computerized tasks: TAP, Starry Night test; VR task: Kim et al., 2010) and total experiment time in VR to quantify neglect-specific behavior (Mesa-Gresa et al., 2011;Weiss et al., 2003).These findings collectively support the use of temporal measures for identifying subtle or compensated neglect symptoms.

VR head movement measures are sensitive neglect indicators
VR monitoring of head movements revealed valuable insights into neglect-typical behavior.Head rotation measures, both along yaw and roll axes, and the number of head turns before crossing were sensitive to neglectspecific behavior.Our results showed differences in head movement parameters between the groups depending on traffic direction.Specifically, the neglect group had less leftward lateralization along the yaw axis than the no neglect and healthy groups for traffic from the left.This difference was also significant for traffic from right.Moreover, the neglect group tilted their heads further to the right shoulder than the other groups, especially for traffic from right.They also exhibited a higher number of head turns for left/right traffic and traffic from left, underscoring the importance of head movements.Given that participants in the study were mainly chronic patients, the higher number of head turns may be an indicator of compensatory behavior.
The neglect group did not exhibit longer or more rightward lateralization along the yaw axis compared to the no neglect and healthy groups.However, their head movement was less stable with unidirectional traffic.This aligns to the chronic phase of neglect symptoms, as patients with acute neglect still show a marked deviation of head position to the right (Karnath et al., 1998).In contrast, the neglect group exhibited significantly less leftward lateralization than the other groups for traffic from left.Although the pronounced rightward head rotation seen in the acute phase was no longer evident, our results suggest that patients with chronic neglect still exhibit distinct patterns of head movement.Specifically, they demonstrate more head turns along the yaw axis and tilt their head further to the right shoulder along the roll axis, particularly in varied traffic directions.

iVRoad has the potential to be used clinically
The iVRoad task introduces a novel approach to behavioral assessment using VR technology, providing objective measures such as head rotation and error patterns in a safe clinical environment.This technology requires minimal space and therapist involvement, allowing for controlled and automated data collection.The increasing integration of VR into both households and clinical settings strengthens the need for the adoption of clinically validated VR tasks.In addition, the iVRoad task significantly enhances the ecological validity of cognitive assessments by providing realistic and dynamic simulations.
All groups completed the VR task comfortably.They easily learned to use the controller and in line with Huygelier et al. (2022), post-VR cybersickness scores were lower than before, suggesting that the VR exposure did not negatively affect patients' well-being.Symptoms occurring prior to the VR task suggest that those typically associated with cybersickness (Brandt & Dieterich, 2017) are rather unspecific and may generally occur in right hemisphere stroke patients.In addition, neglect has been associated with a decreased awareness of symptoms (Jehkonen et al., 2001), and VR's attentionengaging impact may have contributed to a reduction in the symptoms reported after the task.High usability scores and motivation combined with willingness to use VR for clinical purposes demonstrate its potential in neuropsychological settings.Most participants showed interest in cognitive training using VR, underscoring the feasibility of immersive VR for chronic right-hemispheric post-stroke patients with and without neglect.

Selection of difficulty parameters
Choosing appropriate difficulty parameters was crucial for detecting neglect-specific behavior.The VR task was designed to have moderate difficulty, ensuring that errors were made by all groups without an excessive number of errors overall.Two-way traffic led to more errors in all groups, with USN+ showing more near-lane errors and dangerous crossings.In addition, their head tilt angle was greater to the right.The attentional-perceptual deficits associated with neglect were particularly evident for two-way traffic with longer reaction times, and more errors.Manipulating the traffic direction also revealed neglect-specific behavior in head movements, including heading position, head tilt angle, and head turns.The letter insertion task differentiated between groups, highlighting its potential usefulness not only for storytelling but also in applications beyond a clinical setting.
These findings illustrate the importance of task complexity, stimulus density, attention allocation, and dual tasks for the detection of neglect.Andres et al. (2019) showed that dual tasks were more sensitive in detecting compensated neglect symptoms compared to standard neglect tests, such as the cancellation task.Similarly, Azouvi et al. (2002) reported behavioral assessments that evaluate functional performance in daily life situations were more sensitive in assessing neglect deficits compared to paper-and-pencil tests.This aligns with our findings that behavioral assessments in VR were good predictors to identify neglect.

Clinical implications and study limitations
Assessing the driving ability of individuals post-stroke is highly relevant.Neglect symptoms are often overlooked in chronic assessments.Our findings reveal dangerous behaviors, such as increased errors and slower reaction times in pedestrian scenarios, despite similar self-assessments compared to no neglect patients.This lack of awareness is typical in neglect.Increased head turns suggest attempted compensation but fail to prevent errors, especially with inadequate leftward checks for left-approaching traffic.This may increase the risk of errors, as patients tend to start crossing the road late and fail to accurately assess their speed in comparison to the speed of approaching cars.
Future studies should focus on developing advanced methods for measuring neglect severity and subtypes with iVRoad.One potential approach would be the use of adaptive testing or data-driven techniques to quickly obtain results and provide individually tailored cognitive rehabilitation programs with clinical immersive VR tools.This approach would lead to faster and more accurate test results, as well as improved monitoring of follow-up care of neglect symptoms and a reduction in testing time.
A limitation of the study is that the avatar's walking speed in the virtual environment was not adjustable, potentially leading to unrealistic experiences for some participants.This fixed speed was necessary to accommodate participants with hemiparesis and to minimize the risk of cybersickness.Furthermore, we used a self-reported CBS score, which might be problematic due to anosognosia observed in neglect (Chen & Toglia, 2019).We acknowledge that group assignment in relation to the USN+ group is a limitation of this study.Given the low sensitivity of the neglect tests for mild and compensated neglect symptoms, the neglect classification was based on clinical testing and clinical staff evaluation.Despite these limitations, the VR task could still be conducted in a clinical setting and used to examine cognitive functions that might be challenging to assess in real world scenarios due to mobility or gait limitations.Furthermore, it offers the potential to assess and monitor the presence of neglect symptoms from early stages using a task that mirrors everyday situations.

Conclusion
Our study provides evidence that iVRoad is a viable tool for detecting symptoms of neglect in patients with chronic right-hemispheric strokes.The use of spatiotemporal experimental manipulations, such as varying traffic directions, combined with temporal parameters, error patterns, and head movements everyday life situations was particularly effective in detecting chronic left-sided neglect.Reaction times, left-sided errors, and lateral head movements for traffic from left were particularly sensitive in identifying the presence of neglect.A first step toward an improved assessment of neglect, particularly of mild or wellcompensated neglect symptoms, has been made using iVRoad as part of a comprehensive behavioral evaluation for neglect.Furthermore, our results indicate that the use of a road-crossing task was feasible, highly usable, and enjoyable in chronic right-hemispheric stroke patients with and without neglect, as well as healthy participants.Furthermore, iVRoad was highly feasible from patients' perspective and applicable without any adverse effects related to cybersickness interfering with task completion.These results suggest that VR-based methods can improve our detection of discrete neglect syndromes that may otherwise be overlooked.

Figure 1 .
Figure 1.Overview of the immersive virtual road crossing task.The procedure consisted of three parts: (I) instruction and tutorial to familiarize the participants with the system and the task, (II) crossing the roads with three variations of traffic direction, and (III) inserting the letter into the mailbox on either the left or the right side.

Figure 2 .
Figure 2. (a) Neglect patients showed significantly longer times in experiment, reaction, and letter insertion for left-sided mailboxes than control groups.(b) Average number of errors and safe crossings for patients with neglect, patients without neglect and healthy controls.(c) Patients with neglect showed more rightward lateralization along the yaw axis, particularly for traffic from left.Error types and group means; safe crossings inverted.1: Total errors.2: Errors by traffic direction.3: Error type.4: Lane-related errors.5: Siderelated errors.6: Successful crossings and quality.B:.Negative values represent leftward lateralization, while positive values represent rightward lateralization along the yaw axis.

Table 1 .
Participant demographic and clinical data.

Table 2 .
Results of neglect paper-and-pencil tests for each of the three study groups.

Table 3 .
Descriptive statistics for the number of errors associated with traffic direction and group and the results of the group-wise contrast analyses of the negative binomial regression.