We recruited stroke patients over 2 years time and only achieved 2 patients per-protocol and 6 intention-to-treat patients. The recruitment rate was thus too low and therefore the study protocol as planned was not feasible.

Hemispatial neglect is a post-stroke condition in which patients fail to detect stimuli presented on the side of space opposite to the damaged brain hemisphere (contralesional space). To date, there is no established effective treatment for this condition. A virtual reality (VR) behavioral training for the attention deficits characteristic of patients with hemispatial neglect was developed. Patients are stimulated in the visual and auditory modality to orient towards the contralesional side and are rewarded for detecting targets on this side in this training. In the current study the researchers aim to answer two main questions: 1) how feasible is a VR game-based intervention in stroke patients? and 2) what is the efficacy of the virtual reality game-based intervention in reducing the attention deficits characteristic of hemispatial neglect? To answer these questions a randomized partially double-blind placebo-controlled crossover study will be conducted. Two within-subject conditions will be compared: in the active condition patients will play a VR game in which multisensory stimulation is progressively presented in the neglected region (the location where previously presented targets were missed by the patient) and in the placebo condition patients will play a VR game in which the stimulation is presented in the center of of the VR environment. Neglect symptoms will be measured on a two-daily basis to establish the trend of symptom recovery through time. The hypothesis states that symptoms will recover more quickly when patients receive the active version of the VR intervention compared to the placebo version of the VR intervention.

SAMPLE SIZE ESTIMATION:

Power analysis was performed with the SIMR package in R which estimates power for generalized linear mixed models using Monte Carlo simulations. The main analysis will compare the evolution through time of the primary outcome variable between the two within-subject conditions placebo and active intervention. Power was estimated as a function of the number of patients who complete the entire study protocol and as a function of the number of assessment moments per patient. In addition, the power analysis was run under the assumption that the measurement error (residual variance) would be equal to 0.20 SDs. The latter implies that the outcome variable must have a reliability of at least .80. The power analysis revealed that 8 patients need to complete the entire study protocol (per-protocol sample size) - when the study protocol involves a 1-day in-between assessment schedule - to detect a moderate effect size (SD = 0.5) with a type I error rate of 1% and a power of 80%. Thus, for each counterbalancing group a minimum of 4 patients is needed. Assuming that 50% of all patients allocated to a counterbalancing group drop-out at some time point during the study, a total of 16 patients will be recruited to obtain a large enough per-protocol sample size.

MISSING DATA HANDLING:

Missing data can occur when patients do not take part in one or more visits throughout the study protocol (non-monotonous missing data) or when patients drop-out from the study and there is no data available of a patient after drop-out (monotonous missing data). The frequency of occurrence of these two types of missing data will be reported. If inconsistent data occurs on an individual level this will not be considered to be missing data. Out-of-range results for most behavioral outcomes are not likely to occur because computerized assessment tasks guarantee accurate data acquisition. For the behavioral observation scale the inter-rater reliability will be evaluated as a quality check. If the inter-rater reliability across all assessments made in the context of the study is lower than .70 this measure will be reported as insufficiently reliable to be used as a meaningful outcome variable. Only eye tracking data that was sufficiently accurately measured will be considered to be used as an outcome measure. Meaning that, if eye tracker calibration is not good to excellent according to the software delivered with the eye tracker after 5 repeated calibrations the eye tracking data for that assessment will be considered as missing data.

STATISTICAL ANALYSIS:

1. MAIN ANALYSIS: The data will be analyzed using Bayesian mixed models in R. Mixed models are the recommended approach to combine data of single cases and are increasingly acknowledged as a more powerful data analysis approach for clinical trials compared to classic ANCOVAs since mixed models can accurately model time-unstructured data. A Bayesian approach to analyze data is preferred above a classic null hypothesis significance testing because the Bayesian approach allows to quantify the strength of evidence in favor of the null hypothesis. The latter is a valuable attribute in the context of clinical trials as these studies often require proof for no difference between groups on covariates that can be assumed to affect response to treatment.

   The main analysis of interest will compare the effect of the within-subject conditions placebo and active intervention. The model to estimate this effect will include the main effect of time since start of intervention condition, intervention and the counterbalancing group. In addition, the pairwise and three-way interactions of these predictors will be included. A random intercept and random slope for time will be included in the model. This model will be used to predict the primary outcome variable and the secondary outcome variables.

   In addition, the association between the different outcome variables will be reported as a means to estimate to what extent treatment effects may have affected 1 specific outcome or to what extent symptom evolution across different outcome variables was associated.

2. EXPLORATORY ANALYSES: In addition to these analyses, the experience of patients with the VR game based intervention will also be reported. The vocal responses made by patients during gameplay will be rated by two independent raters as expressions of negative or positive emotions. The number of negative and the number of positive expressions relative to the total number of expressions will be compared to each other. If the proportion of positive expressions is higher than the proportion of negative expressions this is taken as evidence suggesting that the patients had a positive experience with the game and vice versa. In addition, since not all patients will spontaneously make vocal responses during gameplay, the mean score of patients on the questionnaire that gauges their experience with the VR-game based intervention will be reported. Given the exploratory nature of this part of the study descriptive statistics will be reported, but no statistical analysis on this outcome variable will be performed. The results of the safety checklist will also be reported. These data are valuable since it can inform other researchers on whether VR is safe to use within the stroke population. These data will be reported in the form of descriptive statistics. All exploratory analyses will be performed on the intention-to-treat sample.

3. SIGNFICANCE LEVEL: The Bayes Factors will be interpreted according to the following interpretation rule: a Bayes Factor of larger than 3.2 suggests substantial evidence in favor of the alternative model, a Bayes Factor larger than 10 suggests strong evidence in favor of the alternative model and a Bayes Factor larger than 100 is decisive for the alternative model. All effects will be evaluated against a threshold of a Bayes Factor of 10. Bayes Factors that are in between 1/10 and 10 will be interpreted as inconclusive evidence. Evaluating effects at a threshold of a Bayes Factor of 10 is comparable to the approach of evaluating effects at a significance level of .01. The primary outcome variable and 5 secondary outcome variables should lead to a maximum Type I error rate of 6% in a worst-case scenario where all 6 outcome variables are completely uncorrelated. This type I error rate is obtained through the formula: 100 [1- (1- α)^k ] where α stands for the significance level and k stands for the number of independent measures.

Inclusion Criteria:

• They are above 18 years.
• They have had a stroke.

Exclusion Criteria:

• They or their legal representative are unable to provide informed consent.
• They have a severe comorbid psychiatric (E.g. psychotic symptoms) disorder.
• They have a premorbid neurodegenerative disease (E.g. Alzheimer's dementia, vascular dementia).
• They have severe written language comprehension deficits.
• They have a medical implant, such as a cochlear implant or a pacemaker.
• They have a severe visual or auditory impairment that cannot be corrected for by wearing glasses or a hearing aid while wearing the Oculus Rift headset.
• They are unable to concentrate on a task for more than 15 minutes or are unable to complete a task according to simple task instructions.
• They have a history of epileptic seizures.
• They do not show signs of a spatial asymmetry in performance on a battery of screening tasks.
• The expected discharge of patients is in a period shorter than 7 weeks.

Data of participants that does not contain any identifiable information and that supports the reported analysis in a publication will be made publicly available to other researchers upon request.

The data will be made available to other researchers upon request.

The principal investigator will evaluate a request for sharing of individual participant data. Access criteria to the coded IPD are that the person who requests access has a profession that is associated with professional confidentiality.

Patients are not explicitly explained that a placebo and active version of the intervention will be compared, which makes it more likely for patients not to be aware of the treatment conditions. However they may notice a difference between the two interventions when they switch over from the first to the second intervention condition. In addition, the clinician who will administer the intervention to the patient on a daily basis cannot be blinded to the specific intervention that is administered to the patient, because the clinician will remain present during the intervention to guide the patient through the intervention. However, the clinicians that will perform the evaluation of symptoms using the tasks that are most sensitive to observer bias will be blinded to the treatment condition that is currently applied to the patient to avoid that the measurement of outcome is affected by observer bias.