Ethics statement

The research adhered to the tenets of the Declaration of Helsinki. This study was approved by the Medical Ethics Committee of the Erasmus Medical Centre, Rotterdam and the Human Research Ethics Committee of the Delft University of Technology. All individuals gave their written informed consent.

Participants

Twenty-five glaucoma patients with Primary Open Angle Glaucoma (POAG) and 14 control participants without glaucoma were recruited from the Rotterdam Ophthalmic Institute (ROI) from an ongoing longitudinal follow-up study into glaucoma imaging.

Participants were measured with standard automated perimetry (SAP) by means of the Humphrey Field Analyzer II (Carl Zeiss Meditec, model number 750), with the 24-2 Swedish Interactive Threshold Algorithm (SITA; Software Version: 4.2.2). The exclusion criteria for all participants were: (1) having a best-corrected visual acuity worse than 6/12 on a Snellen Chart in at least one eye (2), having a refractive error outside the -10.0 to +5.0 diopters range (3), having had cataract surgery in the previous 12 months before the experiment (4), having had previous refractive or vitreoretinal surgery (5), presenting evidence of diabetic retinopathy, macular edema, or other vitreoretinal disease (6), having had previous keratoplastic surgery. There were no age cut-offs during recruitment. We attempted to invite patients and controls of corresponding ages, with the aim to keep the mean age of both groups approximately equal.

Glaucoma patients had a confirmed diagnosis based on optic nerve damage and a visual field loss according to the following measures: (1) Glaucoma Hemifield Test (GHT) result outside normal limits, and (2) 3 or more adjacent points depressed at p < 0.05 or 1 or more points at p < 0.01 in the total deviation plot. Patients with an eye disease other than glaucoma that might affect the visual field (including greater than mild cataract) or with secondary glaucoma were not included. In order to test various types of visual field loss, the patients were selected based on their scotoma pattern as follows: (1) defects in the left eye only (n = 4), (2) defects in the right eye only (n = 7), (3) defects in alternated regions (defects in the upper field of one of the eyes and the lower field of the fellow eye) (n = 5), (4) defects in the upper visual field of both eyes (n = 6), and (5) defects in the lower visual field of both eyes (n = 3). The control participants had a visual field within normal limits in the GHT and no history of ocular disease.

Procedures and driving instructions

Prior to the simulator test, participants filled out a 37-item multiple-choice questionnaire in Dutch language. This questionnaire consisted of 8 general items (e.g., age, gender, medication use, estimated health, worry about sight), 5 items about difficulties with daily activities, 10 items about driving history and self-rated driving performance (e.g., mileage, number of accidents, number of fines, driving skills), 12 items about violations and errors while driving, 1 item about preferred driving modality, and 1 item about experience with cruise control. The items were based on the Driving Habits Questionnaire [25], the National Eye Institute-Visual Functioning Questionnaire (NEI-VFQ) Driving Subscale [26], and the Manchester Driver Behaviour Questionnaire (DBQ) [27,28].

Next, the participants performed reaction time tests on a desktop computer, and received a video instruction of 3.5 minutes. The video described general indications for controlling the car and the tasks per session.

Participants then seated inside the driving cabin, and put on the seat belt in order to maintain the body in a relatively constant position with respect to the screen. Next, participants carried out a series of head movements and eye movements to calibrate the eye-tracker. Participants then drove a training session, followed by an optional break, and the driving test consisting of four sessions. The sessions, including the training session, lasted 278 s each. After each session, participants filled out the NASA-TLX questionnaire for measuring workload.

The environment for all sessions consisted of a two-lane highway without on- and off-ramps. The highway contained several slight bends and a 270 degree left curve of 300 m radius. The highway had a total length of 7534 m and a lane width of 3.6 m. No other moving vehicles shared the road with the participant.

The participants were asked to drive in the right lane of the two-lane highway. The control of the car consisted of steering only; the participants did not use the pedals or gear lever. The car accelerated automatically until reaching a constant speed of 100 km/h (reached 17 s after the car was started).

In session 1, the road was empty. In session 2, the road was also empty, and the participants performed a letter detection task as secondary task. Participants had to verbalize equally sized (5 cm height) green colored (RGB 16%, 100%, 16%) letters appearing on the projected environment. The letters automatically disappeared 4 s after their appearance, and intervals between the letters varied between 3 and 6 s. The first letter appeared 6 s after the session was started. Thirty-two letters per session were projected at 25 different preprogrammed coordinates on the screen.

Based on the positions of the participants’ eyes with respect to the screen as measured during the experiment, we estimated the horizontal and vertical peripheral eccentricities of the projected letters. The mean horizontal angle across participants between a line through the leftmost projected letter and the origin of gaze (i.e., the point between both eyes) and a line perpendicular to the screen through the origin of gaze was -6.4 degrees (SD = 1.4). For the rightmost projected letter this horizontal angle was 36.1 (SD = 2.3) degrees. The vertical mean angles of the lowest and highest projected letters were estimated at -2.9 (SD = 2.3) and 16.8 (SD = 2.7) degrees, respectively. Note that the nonzero standard deviations arise from individual differences in seating position and eye height.

In session 3, the participants had to evade nine static obstacles by changing to the left lane. The obstacles were positioned in the center of the right lane, and were three bicycles, three pairs of bicycles, and three cars. The participants encountered the obstacles at the following times in seconds: 19.7, 49.2, 61.7, 83.4, 121.8, 135.1, 226.0, 236.8, and 267.9 (corresponding distances in meters: 350, 1171, 1521, 2125, 3193, 3563, 6095, 6395, and 7262).

In session 4, the participants had to both perform the letter task and evade the same nine obstacles as in session 3. The times and positions in which the letters and obstacles appeared in session 4 were identical to those in sessions 2 and 3, respectively. The actual letter characters were different from session 2. In the training session, the participants had to evade five static obstacles located at the positions corresponding to the 2nd, 4th, 6th, 8th, and 9th obstacles of sessions 3 and 4. The distance between the car and obstacle at which obstacles became perceivable (due to the finite resolution of the LCD projector) was estimated at 135 m.

Demographic and driving habit data

The following items were selected from the 37-item questionnaire: (1) age, (2) gender, (3) worry about sight (1 = never, 7 = always), (4) medication use, (5) number of collisions involving damage in the past 36 months, as a driver, (6) number of kilometers driven in the past 12 months, (7) driving frequency in the past 12 months (1 = every day, 7 = never), and (8) cruise control experience (no or yes). These items were considered relevant for assessing the (non)equivalence of the glaucoma group versus the control group.

Visual field measures

The monocular visual fields were merged into an integrated visual field (IVF) by determining the best monocular sensitivity for 52 corresponding positions around visual fixation. Furthermore, a measure of sensitivity loss was calculated for the monocular (better eye and worse eye) and integrated visual fields, based on a method reported by Szlyk et al. [9]. Sensitivity measures represented the percentage of 13 points (i.e., 52/4) falling below a 25 dB sensitivity threshold for the quadrants of the visual field (upper left, upper right, lower left, and lower right). The mean deviation (MD) per eye was recorded as well for each participant.

Reaction time

Participants performed the Deary-Liewald reaction time task, consisting of a simple reaction time (SRT) and a choice reaction time (CRT) task [29]. Reaction time is associated with health and general cognitive ability [29], and studies suggest that similar neurodegenerative mechanisms may underlie both glaucoma and cognitive decline [8,30,31]. The reaction time tasks were included to verify whether glaucoma and control groups had equal ability on this elementary cognitive task.

Driving simulator

A fixed-base interactive driving simulator (Green Dino, Wageningen, The Netherlands) was used. The simulator cabin was equipped with the following components: steering wheel, ignition key, gear lever, single seat, and pedals. The feedback of the steering wheel was provided by a passive spring system. Surround sound was used to simulate wind, tires, and engine noise. The virtual environment, including dashboard and rear-view mirror, was projected onto the central screen (1.09 m X 0.77 m) by means of an LCD projector (NEC VT676) providing an image of 1024 x 768 pixels. With the exception of projected side mirrors, the lateral screen projections were not used in order to reduce driving simulator discomfort [32]. The total field of view (FOV) perceived by the participant was approximately 70 degrees horizontally and 50 degrees vertically, depending on the distance between the seat and the screen.

Driving performance measures

The steering angle and vehicle position were recorded with a sampling frequency of 60 Hz. High frequency noise from the steering wheel angle was lowpass filtered (Butterworth, 2nd order) with two cutoff frequencies: 0.5 Hz and 2 Hz. The maximum frequency with which a human can steer is about 1 to 2 Hz [33]. Hence, 2 Hz is an appropriate cutoff frequency to remove sensor noise, whereas a cutoff frequency of 0.5 Hz filters out high-frequency steering inputs (e.g., fierce steering around obstacles) and focuses on the driver’s steering oscillations.

The following measures were calculated per session:

1. • Steering activity (SA) (deg/s) was defined as the average speed of the steering wheel angle [34]. This metric was defined in two different ways: based on the steering angle filtered with high (SA-high; 2 Hz) and low (SA-low; 0.5 Hz) cutoff frequencies. SA-high and SA-low were calculated for obstacle periods (defined as sections starting 150 m before and ending 200 m after obstacles) and outside obstacle periods, regardless of whether the obstacles were actually present in that session. Steering activity can be seen as a measure of driving style, where low steering activity indicates smooth steering, and a high steering activity describes active steering.
2. • Standard deviation of lateral position (SDLP) (m) is a commonly used measure, describing the vehicle’s swerving on the road. A high SDLP value indicates imprecise lane keeping [35,36].
3. • Longitudinal distance to obstacle (LongDtO) (m), a measure of reaction with respect to the static obstacles, with a higher value indicating that the participant reacted earlier. LongDtO was defined as the longitudinal distance between the center of the obstacle and center of participant’s vehicle when the latter crossed the road center (i.e., the line between the two lanes). The mean of the nine obstacles was calculated.
4. • Lateral position to obstacle (LatDtO) (m), the lateral distance between the center of the obstacle and the center of the participant’s vehicle when passing the obstacle. The mean of the nine obstacles was calculated.
5. • A collision was defined as a lateral distance between the center of the participant’s vehicle and the center of an obstacle of less than 1.6 m (contact implied from the car’s dimensions) while driving alongside the obstacle. Note that physical impact would never occur because the participants’ car could move unimpeded through the obstacles.

Letter task measures

The projector (EPSON EPN30) used for the letter stimuli was mounted on top of the projector used for projecting the virtual environment. Sessions 2 and 4 were recorded with a microphone that was unobtrusively fixed to the simulator’s cabin. Times of voice events were extracted automatically from the auditory signal. The reaction time was defined as the elapsed time between the letter appearance and the voice event. A letter was considered as detected if the participant called out any sound within the time in which the letter was projected until one second after its disappearance. Accordingly, the number of misses, and the average letter reaction time (LRT) per session were obtained. Spearman correlations between the letter measures (number of misses and reaction time) and the sensitivity-loss percentages were calculated for the glaucoma group.

Eye-scanning measures

A remote eye tracker (Smart Eye 5.9, Sweden) was used to measure eye gaze during driving. The eye tracker consisted of three Sony XC-HR50 cameras equipped with two infrared illuminators mounted below the virtual scenery of the driving simulator. Data was recorded with a sampling frequency of 60 Hz and was lowpass filtered (Butterworth, 2nd order) with a 20 Hz cutoff frequency.

The following eye-scanning measures were calculated per session:

1. • Percentage of time that gaze was directed at the screen’s top and bottom, as a measure of visual attention (see Figure 1 for an illustration).
2. • Number of fixations per second, a measure of the visual search strategy of participants [23]. Fixations were defined according to a dispersion-based method with a threshold of 3 degrees and a minimum fixation duration of 150 m [37,38].
3. • Mean saccade amplitude (mm), a measure of visual scanning, calculated as the mean distance between subsequent fixations [23,39].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Illustration of top and bottom regions (inside dashed rectangles) used for calculating the eye-scanning measures.

https://doi.org/10.1371/journal.pone.0077294.g001

Self-reported workload

A Dutch version of the NASA-TLX [40], a common questionnaire in driving simulator research (e.g., [41,42]), was used to verify whether glaucoma patients and control participants experienced similar workload. The NASA TLX includes the following six items: mental demand, physical demand, temporal demand, performance, effort, and frustration. Scores were marked on a 21-tick bar ranging from 1 = very low to 21 = very high (1 = perfect and 21 = failure for the performance item). A total score was calculated by averaging the six items and expressing the results on a scale from 0% (lowest rating on all items) to 100% (highest rating on all items).

Statistical tests

Group comparisons (glaucoma patients vs. control participants) were conducted by means of a two-sample t-test assuming that the two samples came from normal distributions with unknown and unequal variances (also known as Welch’s t-test). Differences between sessions were evaluated by means of a paired t-test. Associations between variables were evaluated using the Spearman correlation coefficient. Analyses were two-tailed and the α value (i.e., the false positive rate) was set at 0.05.

No α correction for multiple testing was made, because our sample size was relatively small. Fielder et al. [43] explained that “it remains a statistical fact that measures that decrease α will often increase β”.Mudge et al. [44] argued that researchers should aim to minimize the combined probability (or cost) of Type I errors (i.e., false positives) and Type II errors (i.e., false negatives). Mudge et al. showed that when the sample size is smaller, a more liberal α value should be set to minimize the total probability of error. We argued that setting the α value below 0.05 would disproportionally increase β, the false negative rate.

Excluded data

The data from two glaucoma patients and two control participants was excluded from the analyses. These participants experienced simulator discomfort during the training session and were at that point withdrawn from the experiment. Accordingly, data from 35 participants were used for further analyses.

Eye-scanning recordings may be inaccurate because of the system’s inability to detect facial features or pupils. This problem can be aggravated by reflections and physical obstructions of a participant’s glasses. The data from two participants were excluded after visual inspection of the video recordings from the eye-tracking system. Their data was removed because the eye-tracker measured erratic gaze patterns contradicting pupil movements. Data samples corresponding to 0.5 s before and after blinks, missing data values, and data segments of less than 1 s were removed. Eye-tracking data of participants was completely excluded, if one or more sessions had more than 60% removed data. In total, the data from 8 participants (7 patients, 1 control) was excluded from further eye-scanning analyses.

Descriptive statistics

Table 1 shows descriptive statistics for both groups. There was no statistically significant difference for mileage, driving frequency, SRT, and CRT between the glaucoma group and the control group. Participants from the control group reported to be more experienced with cruise control, and were less likely to wear glasses during the experiment than glaucoma patients. Three participants (one patient and two control participants) had participated in a pilot test 7 months prior to this experiment. Two of the glaucoma patients reported to be under the influence of ophthalmic medication that may affect driving.

Driving performance of glaucoma group vs. control group

The results from the driving measures averaged over sessions are presented in Table 2. The glaucoma group displayed significantly higher steering activity than the control group on the following three metrics: SA-high (obstacle periods), SA-low (non-obstacle periods), and SA-low (obstacle periods). There were no statistically significant differences between the two groups for SA-high (non-obstacle periods), SDLP, LongDtO, and LatDtO.

Figure 2 illustrates the difference between low and high steering activity. For illustrative purposes, the participant with the minimum and the participant with maximum value of SA-high in session 4 were selected. The corresponding SA-high values for these two participants (one control participant, and one glaucoma patient, respectively) were 1.28 deg/s and 2.70 deg/s for the non-obstacle periods, and 1.54 deg/s and 6.91 deg/s for the obstacle periods. Figure 3 illustrates the lateral position for the same two participants in session 4. The corresponding SDLP values for the non-obstacle periods were 0.73 m for the participant with low steering activity and 0.46 m for the participant with high steering activity. It can be seen in Figure 3 that the person with low steering activity took longer to go back to the right lane after avoiding the obstacle, had larger lane center errors, and had smaller lateral margins around obstacles, than the participant with high steering activity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Steering angle versus time of two selected participants in session 4.

One selected participant had very active steering, whereas the other had very smooth steering. The steering angle was filtered with a cutoff frequency of 2 Hz. Positive = steering to left, negative = steering to right. The dotted vertical lines indicate when the participant passed one of the nine static obstacles.

https://doi.org/10.1371/journal.pone.0077294.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Lateral position versus time of two selected participants in session 4.

One selected participant had very active steering, whereas the other had very smooth steering. Positive = lateral position to left with respect to right-lane center, negative = lateral position to right with respect to right-lane center. The dotted vertical lines indicate when the participant passed one of the nine static obstacles.

https://doi.org/10.1371/journal.pone.0077294.g003

The analyses indicated that there were no collisions for any of the participants. Within the group of 23 glaucoma patients, none of the correlations between the percentage of depressed IVF points and the driving performance measures listed in Table 2 reached statistical significance (p > 0.2 for all correlations).

Letter task performance of glaucoma group vs. control group

The glaucoma group missed significantly more letters and had a significantly longer letter reaction time than the control group (Table 3). Within the group of 23 glaucoma patients, people with a larger percentage of depressed IVF points had longer reaction times (ρ = 0.62, p = 0.001) and a higher number of letter misses (ρ = 0.67, p = 5.0*10^-4).

Eye-scanning of glaucoma group vs. control group

Table 4 shows that participants from both groups hardly looked at the top and bottom regions of the screen. Table 4 further shows that there were no statistically significant differences between the glaucoma group and the control group for the eye-scanning measures. Figure 4 illustrates the median percentage of time per session that gaze was directed at the upper and lower regions of the screen. There were no statistically significant differences between the two groups for any session.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Median of percentage of time that gaze was directed at the top and bottom regions of the screen, per session (L = letter task, O = obstacle avoidance task).

https://doi.org/10.1371/journal.pone.0077294.g004

An additional exploratory analysis showed that participants with a larger percentage of IVF points depressed below the 25 dB sensitivity threshold tended to gaze less to the top (ρ = -0.41, p = 0.033) and bottom parts (ρ = -0.35, p = 0.076) of the screen in sessions 1 to 4 combined (n = 27). Similarly, participants with a lower mean deviation (dB) were less likely to gaze to the top and bottom parts of the screen (ρ = 0.41, p = 0.033 for the top part; ρ = 0.43, p = 0.025 for the bottom part; MD of both eyes averaged).

Figure 5 illustrates the eye-scanning behavior of a typical participant during session 2. The participant directed his gaze mostly to the focus of expansion (i.e., the point on the horizon from which, when being in forward motion, the optic flow seems to emanate), directed attention to the letters shortly after they became visible, and immediately directed attention back to the center region. This participant did not display noticeable active visual search in between letter projections; the percentage of session time that this participant spent looking toward the top and bottom regions was 0.45% and 0.00%, respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 5. Distance between gaze and focus of expansion (FOE; static point on the screen) versus time, for one typical participant during session 2 (from the control group).

The line is red when driving on a curved road segment; the line is blue when driving on a straight road segment. Distances from the projected letters to the FOE are indicated by horizontal black lines. Distances from the FOE to the rear-view mirror, speedometer, and side mirrors are 450 mm, 300 mm, and 600 mm, respectively.

https://doi.org/10.1371/journal.pone.0077294.g005

Self-reported workload of glaucoma group vs. control group

There was no statistically significant difference between the control group and the glaucoma group for the mean self-reported workload across the four sessions (p = 0.894). The mean workload score for the control group and glaucoma group was 25% (SD = 18%) and 24% (SD = 15%), respectively. Figure 6 shows the mean workload reported by glaucoma patients and control participants per session. For the control group, there was a non-significant decrease of workload for session 1 versus session 4 (p = 0.233), whereas for the patient group the workload percentage significantly increased (p = 1.3*10^-4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 6. Mean self-reported workload per session (L = letter task, O = obstacle avoidance task).

https://doi.org/10.1371/journal.pone.0077294.g006

Correlations between sensitivity loss and letter task performance

The correlation matrix between sensitivity loss and letter measures is reported in Table 5. Quadrants from the upper visual field were strongly correlated with letter misses and letter reaction times (LRT). The quadrants in the lower visual field yielded relatively weak correlations with the letter measures. The IVF and better-eye VFs showed stronger correlations with letter performance than the worse-eye VF.

Figures 7 and 8 show the positions on the simulator’s screen where letters were prone to be missed by participants and positions where participants showed higher reaction times, respectively. Most missed letters were those projected in the peripheral regions whereas the most successfully verbalized letters were those confined to the road region.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 7. Mean number of letter misses for each projected letter position (red = patients, black = control participants); the radius of the circle linearly corresponds to the mean number of letter misses.

Patients’ mean number of missed letters ranged between 0% (on the road) and 78% (rear-view mirror).

https://doi.org/10.1371/journal.pone.0077294.g007

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 8. Mean letter reaction time for each projected letter position (red = patients, black = control participants); the radius of the circle linearly corresponds to the mean reaction time.

Patients’ mean reaction times ranged between 0.98 (on the road) and 2.48 s (rear-view mirror).

https://doi.org/10.1371/journal.pone.0077294.g008