Pupil diameter (PD) is another eye measurement which can be used to detect workload and mental fatigue. As mental fatigue increases the pupil diameter decreases with respect to baseline measurements [20] . While pupil diameter is also sensitive to changes in arousal and fatigue, it is also sensitive to changes in mental workload. Gonca Gokce Menekse Dalveren and colleagues conducted a study to measure changes in mental workload in surgical residents during surgery where participants were subjected to a computer-based simulation of surgical task. This study found that pupil diameter grows in direct proportion with mental workload [21] . Bastian Pfleging and colleagues also used pupil diameter to develop a model able to predict mental workload under different task load demands. They were able to accurately predict workload under varying mental demands and lightning condition with 75% accuracy [22] . Thus, during the transition state between high mental workload and mental fatigue, pupil diameter is expected to decrease.

The electroencephalogram is a non-invasive way to measure the electrical activity originating from the brain from a set of electrodes spaced on the scalp. EEG is a common way to assess and monitor mental workload and fatigue because of the fluctuation in EEG waveforms, delta (0.5 - 4 Hz), theta (4 - 8 Hz), alpha (8 - 12 Hz), beta (12 - 18 Hz), in various cortical areas [23] . Figure 1 shows a scalp topography of electroencephalogram activity before (beginning) and after (end) driving simulation. Blue indicates a decrease in power spectra and red an increase in power spectra for the specified waveband, from Zhao et al. 2012 [24] .

Transition state between increasing workload and mental fatigue generally results in an overall increase in delta, theta and alpha frequency band, and a decrease in beta frequency band as we can see on the figure.

Justine R. Magnuson and colleagues (2021) investigated the development and recovery of task-induced mental fatigue. They collected EEG signals and subjective assessment of participants while performing a 60 min N-back test followed by a 60 min post-task resting time. The authors observed that mental fatigue was induced after 30 - 45 min of the N-back test, from both objective and subjective measures. They also observed a 60 min recovery for some alpha and theta bands to baseline levels during resting time. However, complete recovery was not achieved after 60 min resting time [29] .

VR-based relaxation therapy has received a lot of attention over the last few years. Among them, the virtual train therapy [30] is a VR relaxation therapy in which participants are immersed in a moving train and travel virtually looking through the windows to the landscape (Figure 3). The virtual train immersive therapy has shown to be effective in promoting relaxation, apathy, sleep, and reducing negative emotion such as stress and anxiety [30] . Hence, the cognitive benefits of a short virtual train therapy motivate the interest of exploring its effects on mental fatigue symptoms.

We generate mental fatigue among participants through a simulated VR environment of a work office where participants are prompted to perform various set of cognitive tasks. There are two types of cognitive tasks set: one with distractors and one without distractors. Distractors are fake answers or wrong hints that will be displayed in the user’s virtual visual field to distract them and provoke fatigue. During these tasks, fatigue is measured through 3 indicators: pupil dilation, workload, and task engagement. The exercises chosen to generate mental fatigue are aligned with the active fatigue framework discussed in Section 3: we generate mental fatigue through cognitive tasks which aim to exert a high cognitive load on the participants.

At first, participants will perform in both set of tasks consecutively (25 min). Then the relaxation period will begin (10 min). Finally, the participant will answer different questions from the set of cognitive tasks without distractors (10 min).

The set of cognitive tasks with distractors consists of mental arithmetic tasks, anagram tasks and backward digits span (BDS) tasks. The user must perform these tasks within the time allocated for each. For each cognitive tasks, fake answers in the form of “hints” are displayed to distract the users. The arithmetic tasks (Figure 2) consist of a series of addition and subtraction of decimal numbers presented in the virtual environment that must be solved by the user. An example of a mental arithmetic task is an equation of the form “24.54 − 12.89 + 2.13 + 11.72 − 7.08 − 3.23” in which the user must use the virtual keyboard to summit the correct answer to this equation, 15.19 in this example.

In the anagram task, the user is presented with a set of letters, and must rearrange these letters to form the appropriate dictionary word. For example, a user is presented with the letters “R”, “E”, “D”, “R”, “U”, “M”, and must use his virtual keyboard to type the correct word “MURDER”.

In the backward digits span, the user is presented with an ordered sequence of numbers that appears on the user’s virtual environment for a short period of time. The user is asked to memorize the sequence during that period. A few moments after the sequence is removed from the user’s screen, the participant is asked to recall the sequence in the reverse order of presentation using their virtual keyboard. For instance, if the user is presented with the sequence “9-7-4-2-5-9-3”, he must memorize the sequence during the allowed time window and enter the reverse sequence order, “3-9-5-2-4-7-9”.

Then another set of cognitive tasks without distractor is presented to the participant. The goal here is to show that the eventual fatigue has been previously generated and is no more due to the distractors. This set has different types of exercises of a 5 min duration to measure cognitive and memory performance. These tests continue to generate mental fatigue as they require concentration, attention, memory, and other cognitive resources. This set is composed of an attention, a naming exercise and three different memory tests to evaluate contextual/visual memory, working memory and short-term memory. These exercises will allow cognitive performance comparison between post-fatigue and post-recuperation states.

After the mental fatigue generation events, participants are exposed to the “Travelling Therapy” VR environment for 10 - 15 minutes to reduce their negative emotions, mental fatigue and increase their concentration [30] . This environment (Figure 3) projects the users into a virtual train (360-degree environment), where they are sitting and can turn their head to look through the windows or observe events inside the train. The windows reveal a natural or relaxing landscape which can consist of forests with animals, mountains, snow mountains, simple roads, lakes, etc. In the train, the user can also see persons and/or pets interacting with each other. It is known that exposure to natural elements and landscape aid in the recovery of attentional fatigue and experiments realized with this virtual train has proved a reduction of negative emotion and an increase of memory and cognitive function [30] . We aim to analyze if exposure to this virtual natural landscape can produce similar effects on mental fatigue recovery.

31 participants (15 female and 16 male) aged between 19 and 29 years old were invited to a room at Beam Me Up office, partner of the project, 5925 Monkland Ave, H4A1G7, Montréal, to complete the different steps of the experiment and a few real-time and offline outcome measures.

During the experiment, some electrodes loss their signal due to displacement caused by movement of the head and weak signal caused by hair density on the scalp. We selected EEG channels in a way that for the majority of participant, selected, at least 1 of 2 electrodes defining a scalp region remained intact. Regions are defined as follows:

1) (fp1, fp2) prefrontal

2) (f4, f8) frontal right

3) (f7, f3) frontal left

4) (t4, t6) temporal right

5) (t3, t5) temporal left

6) (c4, c3) central

7) (p3, p4) parietal

8) (o1, o2) occipital

Selection of participant: Participant missing 1 or more regions from the selected scalp region were discarded from the dataset. Thus, we had to discard the central electrodes (c4, c5) from the dataset for all participants. Moreover, we had to discard 7 participants (4 female and 3 male) due to a too large number of failed electrodes during the experiment. Finally, 3 more participants (1 female, 2 male) were additionally discarded from the dataset due to failure of the EEG headset during the experiment or because they were unable to finish the experiment due to discomfort/headache. The final dataset was thus composed of 12398 EEG power spectral density segments of prefrontal, frontal right, frontal left, temporal right, temporal left, parietal, and occipital electrodes from the 21 participants (10 female and 11 male).

The EEG signal was band-pass filtered with a fourth-order Butterworth filter (high-pass filter cut-off frequency: 1 Hz, low-pass filter cut-off frequency: 30 Hz). The signal was then passed onto a wavelet denoising filter to remove signal noise, then the power spectral density was computed from the result using the Welch method. Outliers were removed using the interquartile range independently for each participant dataset. From the result of the Welch calculation to retrieve power spectral density, the absolute power of theta, alpha, beta and delta at each electrode was calculated. Relative powers (W/Hz) were used to compute workload and task engagement using the formulas presented in

Relative power of θ = (power of θ)/(power of θ + power of α + power of β + power of ∆)

Relative power of α = (power of α)/(power of θ + power of α + power of β + power of ∆)

Relative power of β = (power of β)/(power of θ + power of α + power of β + power of ∆)

Relative power of ∆ = (power of ∆)/(power of θ + power of α + power of β + power of ∆)

Band power of two electrodes elements of the same region were averaged when both electrode data was available. If one of the two electrodes was identified as “railed”, only the available was taken to describe the target region.

Labels have been established based on variation of the pupil diameter, which is a well-known physiological indicator of mental fatigue and mental workload: pupil increases with respect to baseline when subjects are under high mental charge and decreases when fatigue rises. Baseline range is calculated at the first 60 epoch for each participant. 10 epoch moving segments by 1 epoch increments are assigned to the label 0 if the mean of the segment falls between the baseline range, 1 if the mean falls above baseline range, and 2 if the mean falls below the baseline range.

Labels were assigned with respect to pupil size variation with respect to baseline, to identify segments of increasing workload and fatigue. Assigning fatigue labels according to eye measurements is a methodology that was employed by Ren Ziwu and colleagues [10] to classify mental fatigue using an RBF neural network. The correlation between pupil diameter and workload/fatigue progression was verified against a well-known EEG indicator of workload and fatigue: Theta F/Alpha P (theta bands of frontal electrodes over alpha bands of parietal electrodes). Thus, increases in the theta frontal to alpha parietal ratio indicate an increase in workload. An increase of this index followed by a decrease indicates transition states from high mental workload to increasing mental fatigue levels [28] . Figure 4 shows the measure of such a ratio during cognitive exercise for participant #2. We see, in dashed yellow, an increase of both indexes followed by a decrease in dashed red which indicates transition states from high mental workload to increasing mental fatigue levels.

The correlation value between the smoothed and normalized pupil diameter and thetaF/alphaP ratio over all participants is 0.54, which is a strong relation for the nature of the data (see Figure 4). The distribution of the labels during the first half and the second half of the cognitive tasks part of the experiment shows a significant decrease in absence of fatigue segments and slow fatigue progression proportion. Moreover, we noted a significant increase of rapid fatigue segment proportion during the second half of cognitive tasks compared to the first half. To compare the different fatigue segments, we used the Wilcoxon non-parametric test. This test was chosen (instead of paired t-test) because the distribution of fatigue segments did not follow the normality assumption from parametric tests (Table 1). Results show a significant decrease in non-fatigue and slow fatigue signs, and significant increase in rapid fatigue progression signs.

The workload and task engagement EEG score was computed from the formulas presented in 4.2.1 and 4.2.2. To provide evidence of their evolution we compare the measures of the first half of tasks with the measures of the second half. Hence, we observe a significant decrease in workload score during the second half of the cognitive tasks compared to the first half (p < 0.05) (Figure 5 and Figure 6). However, we did not observe a significant decrease (p < 0.05) in task engagement for the majority of participant. This phenomenon can be explained by the nature of the fatigue generated by the cognitive tasks: active fatigue. Active fatigue (fatigue induced by high workload condition) results in a slower task disengagement overtime than passive fatigue (fatigue induced by low workload condition) [19] . Thus, a longer period of cognitive tasks would have been needed to see a change in task engagement. Figure 5 shows that the mean ThetaF/AlphaP index decreases in the second half as a result of fatigue. No significant change in task engagement.

Another phenomenon that might explain the stability of the task engagement observed in this experiment is the diversity of the exercise presented to the participants. Thus, it is possible that participants did not disengage in the presented task simply because the rapid rotation of different exercises did not allow them to get bored by them, but rather intrigued. This motivational influence on task-engagement was also observed by Jesper F. Hopstaken and colleagues, who observed that change in reward during mentally fatigue task was able to restore task-engagement levels and performance score, while participants still reported to be highly fatigued [31] . Thus, change in task engagement is likely to be influenced by a complex relation of trade-off/reward system rather than depletion of a finite reserve of cognitive resources.

Figure 6 shows the progression of workload. This index is known to decrease as mental fatigue increases. In dashed yellow, an increase of theta/alpha score followed by a decrease, shown in dashed red, indicates transition states from high mental workload to increasing mental fatigue levels.

Our aim was to classify states of mental fatigue segment during a period where participants were asked to perform in various cognitive tasks. All 8 electrodes cerebral regions with their transformation (presented in 6.3), except for central electrode region, were used in the feature matrix for classification. The balanced accuracy as opposed to standard accuracy was chosen as one of the parameters to evaluate model performance, as the proportion of labels across the dataset was unbalanced. Various machine learning (ML) algorithm candidates were trained and tested in order to select the ML algorithm which best fits our data with respect to evaluation metric “balanced accuracy” and “f1” measures (Table 2). Among the 8 different classifiers candidates, RBF SVM showed a better performance with respect to balance accuracy and f1 evaluation metrics.

After parameter tuning achieved through grid search algorithm, the balanced accuracy of the RBF SVM on the test data was 95%.

Our aim was to verify the effect of a short (10 - 15 min) VR relaxation therapy on the participant fatigue levels. During the experiment, the set of cognitive tasks without distractors were presented before and after the relaxation therapy to compare fatigue levels through pupil dilation.

A Wilcoxon test performed on mean pupil diameter across participants, for pre and post relaxation period, favored rejection of the alternative hypothesis: a difference in mean pupil diameter across participants eye segments for pre and post relaxation period (Figure 7). We used the Wilcoxon test for the same reasons indicated in Section 7.1.

While these results do not invalidate the use of virtual therapy to alleviate mental fatigue symptoms, they indicate that time allocated for relaxation in this experiment might have not been sufficient to observe benefits in this recuperation technique. Thus, since standard recovery after mentally fatiguing tasks can take as long as 60 min post-task [29] , further research would be needed to compare the recuperation of fatigue symptoms with the aid of VR relaxation therapy and standard no activity recovering.