Saccades during visual search in macular degeneration

Abstract

Macular degeneration (MD) compromises both high-acuity vision and eye movements when the foveal regions of both eyes are affected. Individuals with MD adapt to central field loss by adopting a preferred retinal locus (PRL) for fixation. Here, we investigate how individuals with bilateral MD use eye movements to search for targets in a visual scene under realistic binocular viewing conditions. Five individuals with binocular scotomata, 3 individuals with monocular scotomata and 6 age-matched controls participated in our study. We first extensively mapped the binocular scotoma with an eyetracker, while fixation was carefully monitored (Vullings & Verghese, 2021). Participants then completed a visual search task where 0, 1, or 2 Gaussian blobs were distributed randomly across a natural scene. Participants were given 10 s to actively search the display and report the number of blobs. An analysis of saccade characteristics showed that individuals with binocular scotomata made more saccades in the direction of their scotoma than controls for the same directions. Saccades in the direction of the scotoma were typically of small amplitude, and did not fully uncover the region previously hidden by the scotoma. Rather than make more saccades to explore this hidden region, participants frequently made saccades back toward newly uncovered regions. Backward saccades likely serve a similar purpose to regressive saccades exhibited during reading in MD, by inspecting previously covered regions near the direction of gaze. Our analysis suggests that the higher prevalence of backward saccades in individuals with binocular scotomata might be related to the PRL being adjacent to the scotoma.

1. INTRODUCTION

In healthy vision, the fovea plays a critical role for high-acuity vision and for being the oculomotor reference. When foveal vision in both eyes is compromised by diseases such as macular degeneration, tasks that require both high-acuity and eye movements are impacted. We are particularly interested in the adaptations to eye movements that occur in real-world tasks, particularly how eye movements compensate for the information that is hidden by the scotoma. So, we chose to examine eye movements in visual search, while individuals viewed the display binocularly.

Previous studies have investigated visual search in individuals with central field loss. They have most often been interested in latency as a performance measure (Mackeben & Fletcher, 2011; Wiecek et al, 2011), or estimating general saccade characteristics such as amplitude, dwell time, and scan path (Boucard et al 2015; Thibault et al 2015, 2018). Two previous studies have specifically tried to examine saccade characteristics with respect to the scotoma, but only a coarse estimate of the binocular scotoma was used (Van der Stigchel et al 2013, Janssen & Verghese, 2016). As we are specifically interested in how each AMD participant’s eye movements compensate for the information hidden by the binocular scotoma, we needed a precise characterization of their binocular scotoma, while fixation was monitored.

Given that the size of the binocular scotoma depends on the overlap of the two monocular scotomata (Arditi, 1988; Satgunam et al, 2012), which in turn depends on convergence (and viewing distance), we wanted to measure the scotoma under the same viewing conditions as the visual search task. Although previous studies have measured the binocular scotoma, their methods either used specialized displays for scotoma measurement (e.g., custom dichoptic visual display as in Woods et al, 2010), specific assumptions to superimpose monocular scotomata (Ghahghaei and Walker, 2015), or did not have a precise estimate of the binocular scotoma (Van der Stigchel et al; Janssen & Verghese, 2016; Shanidze & Verghese, 2019). Sullivan & Walker (2015) had a more precise map, but the binocular scotoma was measured using a grid of points centered on the non-functional fovea — they used a large wagon-wheel stimulus with a spoke-like arrangement in the periphery to help guide the fovea to the center of the display. Thus the measurements were made with respect to the old fovea and not the preferred retinal locus (PRL). To obtain a detailed map of the binocular scotoma relative to the PRL, we used an iterative coarse-to-fine strategy; first mapping the scotoma coarsely and then manually selecting points near the scotoma boundary to delineate the extent of the scotoma (Vullings & Verghese, 2021), while fixation was carefully monitored. The visual search images were later displayed on the same screen under similar viewing conditions, using the same eye tracker.

We used a slightly modified version of the visual search task used by Janssen & Verghese (2016), where observers actively searched for a variable number of Gaussian blobs (0 to 2) superimposed on real-world images. In that study, we only had a rough estimate of the location of the binocular scotoma and therefore our analysis of saccades with respect to the scotoma was crude (classified as either toward or away from the binocular scotoma). In the present study we were able to analyze saccades in exquisite detail to determine both the direction and magnitude of saccades with respect to the scotoma, and whether they adapted to recover information hidden by the scotoma. Specifically, we analyzed saccades to determine if their distribution was consistent with maximally uncovering the scotoma, or whether they simply followed the controls’ distribution of saccade.

Our data show that individuals with central field loss (CFL) due to macular degeneration tend to direct their saccades towards regions obscured by their individual scotoma, much more than control observers did to comparable regions. These saccades tended to be smaller in amplitude than saccades of control observers and did not completely uncover the region hidden by the scotoma. Rather than making sequential saccades to fully uncover this region, individuals with CFL tended to look back at the newly uncovered region. This tendency for backward saccades is pronounced in individuals with larger binocular scotomata. Thus, backward saccades appear much more common among individuals with central scotomata during visual search and probably serve a similar purpose to regressive saccades exhibited during reading (Bullimore & Bailey, 1995) to inspect previously covered regions near the direction of gaze. Our analysis suggests that the higher prevalence of backward saccades in individuals with binocular scotomata might be related to the PRL being adjacent to the scotoma.

2. MATERIAL AND METHODS

2.1. Participants

Seven adults with macular degeneration (57–87 years old, 4 females) and six age-matched controls (61–74 years old, 5 females) participated in this research. Six MD participants had age-related macular degeneration in one or both eyes and one had Stargardt’s disease. Five participants had a binocular scotoma (B) and three had non-overlapping monocular scotomata (M), including one participant whose scotoma became binocular during the testing period (participant M1 is the same individual as participant B2). Control participants (C) had normal vision, or vision that could be corrected to normal. As binocular MD participants chose not to wear their spectacle corrections, we had all observers (monocular MD and controls) not wear their spectacles for the stimulus presented at a viewing distance of 1m. Binocular visual acuity without correction is reported for all participants in Table 1. All participants with a binocular scotoma were referred to us by the low-vision rehabilitation practice of Dr. Donald Fletcher at the Pacific Vision Foundation. All experimental procedures were approved by the Institutional Review Board of the Smith-Kettlewell Eye Research Institute and followed the tenets of the Declaration of Helsinki. All participants gave informed written consent after an explanation of the nature of the study, and received monetary compensation for their participation. Participants’ characteristics are summarized in Table 1.

Table 1.

Participant Characteristics

ID	SEX	AGE	BIN. ACUITY (LOGMAR)	DOM. PRL [X°,Y°], 95% BCEA (DEG2)	NON-DOM. PRL [X°,Y°], 95% BCEA (DEG2)
PATIENTS
B1	F	76	0.3	[−0.7, −0.8], 0.1	[−1.0, 0.4], 1.65
B2	F	80	0.42	[0.0, −5.5], 0.37	[0.0, −7.5], 5.09
B3	F	87	0.5	[−4.5, −1.5], 0.71	[−7.0, −2.0], 0.91
B4	M	57	1.1	[−1.9, −5.5], 0.71	[−2.9, −4.2], 0.86
B5	M	77	1.18	[−8.8, −23.2], 0.36	[−7.1, −25.9], 1.58
M1	F	78	0.0	N/A 0.31	[−6.0, −5.7], 4.49
M2	M	75	0.1	N/A 0.11	[−2.7, −6.4], 2.21
M3	M	79	0.18	N/A 0.03	N/A 0.05
CONTROLS
C1	F	65	0.44	N/A 0.06	N/A 0.34
C2	F	74	0.26	N/A 0.04	N/A 0.05
C3	M	61	0.6	N/A 0.14	N/A 0.14
C4	F	72	0.2	N/A 0.15	N/A 0.17
C5	F	64	0.2	N/A 0.05	N/A 0.07
C6	F	65	0.24	N/A 0.15	N/A 0.27

Prior to testing, all participants were screened using a standard battery of tests to measure residual functional vision. Monocular scotomata and optic disk location were mapped in each eye using microperimetry with unattenuated 0-dB dots (dot luminance, 125 cd/m²; Weber contrast, 12.5) with a scanning laser ophthalmoscope (OCT/SLO; Optos, Marlborough, MA, USA), which probes a field size of 29.7°. A custom arrangement of points was used to determine the scotoma profile. Fixation stability was also measured by using a 10-second fixation target and was characterized by a bivariate contour ellipse area (BCEA; Steinman, 1965) that included 68% of fixations. When possible, the foveal pit location was estimated from optical coherence tomography. If the foveal pit from one eye had been located but we could not locate the foveal pit in the other eye due to advanced AMD, we assumed similar eccentricity (we did this for one participant, B5).

Table 1 lists the gender, age, and relevant visual characteristics of our participants. The PRL eccentricity was obtained by using the center of mass of fixational eye position during a 10 s interval (by fitting a 95% BCEA to the data) and calculating its distance in degrees from the anatomical foveal pit obtained in the OCT. Fixation stability corresponds to the area of this BCEA. We report uncorrected binocular acuity as participants did not wear their optical correction during the experiment to view the screen at 1 m. Note: PRL eccentricity is not provided for M3, as this participant did not have a central scotoma, and had foveal sparing. Similarly, PRL eccentricity does not apply to the unaffected eye of monocular MD participants and controls who had foveal fixation.

2.2. Apparatus

Stimuli were generated using the Psychophysics Toolbox (Brainard, 1997; Pelli, 1997) for MATLAB (MathWorks, Natick, MA) and displayed on a large projection screen (40.4° × 30.3°), illuminated from behind by a Mitsubishi XD490U DLP projector. Participants were seated on a height-adjustable chair in a darkened room facing the center of the screen at a viewing distance of 1m, while their heads were restrained using a chin and forehead rest, to minimize movement. Viewing was binocular, but only the position of the better/dominant eye was monitored in the eye tracker. Eye movements were measured continuously with an infrared video-based eye tracking system (Eyelink, SR Research Ltd., Ontario, Canada), sampled at 1000 Hz. Eyelink data were transferred, stored, and analyzed via programs written in MATLAB running on a Windows computer.

2.3. Procedure

2.3.1. Binocular scotoma mapping

In a first step, we used the eyetracker to map the static binocular scotoma with respect to the fixational PRL according to a manual coarse-to-fine recursive method (Vullings & Verghese, 2021). To summarize the method we used previously, we measured the binocular scotoma while participants viewed a tangent screen and fixated a central marker. Throughout the mapping of the binocular scotoma, we monitored the eye position of the dominant eye with an Eyelink 1000 eye tracker. The points were mapped with high-contrast dots (with a luminance of 284cd/m² on a background of 79 cd/m², yielding a Weber fraction of ~2.6), so unseen locations represent absolute scotomata. The dots were flashed briefly (200 ms) on the screen and participants pressed a button when they saw a flash. In each block of trials we probed 35 locations, with each location tested twice. In the first block of trials, we started with a regular 7 X 5 location grid that spanned the screen the 36 X 29° screen, and had a coarse sampling of 5° between test locations. In subsequent blocks we manually refined the tested locations, sampling more densely where dots were missed. Across an average of 6 blocks, we were able map the scotoma in detail by implementing this iterative coarse-to-fine strategy. We assumed that binocular scotoma participants used the same PRL for fixating and making saccades (see Section 3.3 that discusses previous saccade and pursuit data for these five participants that justifies this assumption; also see White & Bedell (1990)).

2.3.2. Visual search

The visual search task was adapted from Janssen & Verghese (2016), in which participants had to look for and report the number of Gaussian blob targets superimposed in a natural scene. The scene consisted of an image selected pseudo-randomly from a database of 80 images of outdoor and indoor scenes. Participants sat at a distance of 1 m from a display that subtended 36.17 × 29.12°. Zero to two blobs (two-dimensional Gaussians with a spatial standard deviation of 0.5°) were superimposed on this scene. The peak luminance of the gray blob was adjusted on each trial to match the average gray level on the scene. At the beginning of each trial, participants looked at the fixation cross displayed at the center of the screen and pressed the space-key to start the trial. Then, the scene was displayed on the full screen for a duration of 10 s, and subjects were instructed to search the scene actively with eye movements. At the end of the trial, participants either verbally reported the blobs to the experimenter (preferred method for those with a scotoma), or entered the number on a keypad. All participants completed 80 search trials, except participant B4, who only completed 60 trials.

2.3.3. Acquisition and Data analysis

Eye-tracker calibration of the dominant eye of all participants was attempted with 5-point calibration grid on the Eyelink 1000. For some of our binocular MD participants, if we could not record stable eye position with the 5-point grid, we used a 3-point grid and repeated the calibration and validation steps, until we got consistent values. The eyetracker determined a fixation locus for static fixation, and we report the position of this locus during active visual search. Our previous study (Vullings & Verghese, 2021), which measured the binocular scotoma and PRL in the eye tracker, suggests that this fixation locus is the binocular PRL. In that study, the binocular scotoma appeared consistent with the superposition of the two eyes’ monocular scotomata measured in the SLO (and aligned on the respective foveae), and the binocular PRL measured in the eye tracker appeared consistent with the monocular dominant eye PRL measured in the SLO. The fixation stability of the binocular PRL during binocular scotoma mapping is shown in Figure 2Buj6zs5. Four of our five binocular scotoma participants were also participants in the Vullings and Verghese (2021) study.

Figure 2.

A. Accuracy at the visual search task for all participants (orange: Controls, blue: monocular MDs, green: Binocular MDs). The horizontal dashed line corresponds to 33%, the probability of guessing the answer by chance. The solid line represents the average within a participant group and the dotted lines represent the bootstrapped 95% confidence intervals (CI). B. The individual plots display the size of the scotoma for our binocular scotoma participants. Their fixation stability is estimated by a kernel density plot of fixation distribution during binocular scotoma mapping. The monocular scotoma participants and controls have no measurable field defect (binocular scotoma).

Eye movements were recorded throughout each trial. For online saccade detection, we used the Eyelink saccade detector to identify saccade onset and offset, using 30°/s velocity and 8,000°/s² acceleration thresholds. For offline analyses, a human observer validated each saccade manually, to check for glissadic eye movements or noisy eye signals (including blinks) that were sometimes spuriously detected as saccades by the Eyelink’s algorithm; saccades with duration longer than 100 ms were automatically excluded. On average, we kept 91.3 ± 1.9%, 86.3 ± 6.2% and 93.5 ± 1.6% of saccades for participants B1-B5, M1-M3 and C1-C6, respectively.

We do not have complete location data for the position of the target on the screen from our earliest observers, whom we could not rerun because of the pandemic. In particular, we are missing part of the trial-by-trial target location data for the binocular MD participants B4 and B5 whose individual scotomata were large enough to obscure the search target. So, while it may have been potentially interesting to examine whether a saccade was made to inspect a recently “uncovered” target after it had been previously obscured by the scotoma, it is not clear that such an analysis would be conclusive about when the MD participant “saw” the target, for the following reason. The target was relatively large (~2°), and most of the display was visible even for the individual with the largest scotoma (B5 had at most one-quarter of the display occluded at any time). Furthermore, the trials lasted 10s during which eye movements traversed the display, which means the target was likely visible at multiple points during the trial, so it is not possible to determine when exactly the observer saw the target, as they only responded at the end of the experiment.

3. RESULTS

This section first describes the accuracy of finding the targets in the trial, and then goes into a detailed analysis of saccades for each individual MD observer. The saccade analysis includes general saccade characteristics for each individual such as saccade direction and amplitude as well as more in-depth comparison of the spatial distribution of saccades to potential saccade adaptations. Specifically, we examine whether saccades adapt to uncover the region hidden by the scotoma, or whether they continue to follow the spatial saccade distribution of controls. We also examine the temporal sequence of eye movements with respect to the binocular scotoma to determine whether MD observers use a sequence of saccades to uncover the scotoma and whether they tend to inspect a region that was recently uncovered by a previous saccade. For both the spatial and sequential analysis of saccades we compare the eye movement patterns of the binocular MD participants to that of controls.

3.1. Accuracy:

We measured the accuracy as the fraction of trials where the number of blobs was correctly reported. Because there were three choices (0, 1 or 2 blobs), 33% correct represents chance performance. Figure 2 plots the data for individuals with a binocular scotoma (B), for those with monocular scotomata without a field defect (M), and for control participants (C). Individuals with monocular scotomata perform comparably to controls, while the performance of individuals with a binocular scotoma varied over a wide range, and seemed to be particularly impacted for the individual with the largest scotoma (Figure 2b). The solid horizontal line plots the average accuracy for each group and the dashed lines plot the corresponding 95% confidence intervals, estimated by bootstrapping the proportion correct with replacement. As can be seen, there is no significant difference across groups in their accuracy for detecting the blobs.

3.2. Saccade Amplitude and Direction:

We also measured the saccade amplitude distribution during all visual search trials and estimated the median saccade amplitude and 95% confidence intervals for the control, binocular and monocular scotoma groups. The median saccade amplitude for controls during visual search is around 5.36° (95% CI: lower bound 5.27, upper bound 5.44) (Figure 3), similar to the average amplitude reported for visual search tasks in natural images (Tatler, Baddely, Vincent, 2005). Individuals with monocular scotomata have a similar median saccade amplitude to controls (5.23°; 95% CI: lower bound 5.11, upper bound 5.34). However, the median amplitude for individuals with binocular scotomata is significantly smaller at 3.76° (95% CI: lower bound 3.67, upper bound 3.85). Figure 3b examines the cause for difference in saccade size for binocular MD participants, by binning saccades into those that went toward or away from the scotoma. It is clear that saccades towards the scotoma are particularly small with median amplitude 2.69°(95% CI: lower bound 2.61, upper bound 2.76), similar to the report of Vander Stigchel et al (2013), At the same time, saccades in non-scotoma directions have a median amplitude of 5.25°(95% Confidence intervals: lower bound 5.10, upper bound 5.37)). which is not significantly different from the median amplitude of control saccades.

Figure 3.

The median saccade amplitude for the control (orange), monocular (cyan) and binocular (green) scotoma groups, that includes saccades in all directions. For the binocular saccade group, saccades are further broken down by saccades away from the scotoma (gray) and those toward the scotoma (gold) as shown in the schematic. Error bars indicated 95% confidence intervals of the median.

In addition to the amplitude, we analyzed the direction distribution of saccades of our MD participants and compared them to that of controls. Figure 4 shows a polar histogram of the saccade frequency in various directions for our MD participants (green and cyan plots indicate distributions for those with Binocular and Monocular scotomata, respectively). The orange contour marks the median saccade frequency for controls in corresponding directions. Compared to controls, MD observers with binocular scotomata (shown in green) seem to have a skewed direction distribution with more saccades in certain directions and fewer saccades in other directions. Interestingly, the directions with more saccades appear to be in the direction of the scotoma for individuals with binocular scotomata (with values exceeding the 95% confidence intervals of control frequencies). At the same time, individuals with monocular scotomata but no binocular field defect (shown in cyan) have saccade direction distributions largely indistinguishable from controls (values within the 95% confidence intervals of control frequencies). One interesting case is M1, who first participated in our studies when she had non-overlapping monocular scotomata in the two eyes. In the span of a year, she developed wet macular degeneration in her dominant eye, resulting in a significant binocular scotoma (her data are shown as B2). From a comparison of the direction profiles of M1 and B2, it is clear that the same individual’s saccade direction distribution changes after the development of the binocular scotoma and becomes skewed in the direction of the scotoma.

Figure 4.

A polar histogram of saccade directions for our 5 MD participants with Binocular scotomata (green) and 3 MD participants with monocular scotomata (cyan). B2 is the same individuals as M1, following wet AMD in one eye, creating a binocular scotoma. The superimposed gray regions indicate the size of the scotoma for the binocular MD participants, relative to a PRL at the origin. The orange contour marks the average saccade direction for control participants. The dark and light gray arcs indicate direction where the saccade frequency is significantly higher or lower than the 95% confidence limits of the frequency for controls. The saccade distribution appears to be skewed toward the scotoma for these participants.

Overall, it appears that individuals with a binocular scotoma make more frequent saccades toward their scotoma (Figure 4) that are smaller in amplitude (Figure 3b). Comparing saccades toward and away from the scotoma suggest that there are some qualitative differences. However, to determine the true extent of these differences, and whether it is possible that MD saccade behavior has actually adapted to uncover information hidden by the scotoma, we need to consider both the amplitude and direction of saccades with respect to the scotoma (Section 3.3), and to examine sequences of saccades over time (Section 3.5).

3.3. Saccade Distribution:

All the analyses from here on consider the complete distribution of saccade amplitude and direction, with respect to the previous fixation. The data are with reference to the eye position while viewing the fixation point during eye-tracker calibration, which we assume is the fovea for controls and the static PRL of the dominant eye for individuals with macular degeneration individuals (during binocular viewing). It is possible that participants may have used multiple PRLs during active search during binocular viewing. However, we think it is unlikely for our 5 binocular-scotoma participants as they used a single retinal locus when they tracked a step-ramp target (Rashbass, 1961) in a smooth pursuit task with their dominant eye. These measurements were made in a Rodenstock SLO where we could visualize the retinal locus used during both the saccade and pursuit phases of the task (Shanidze et al, 2016; Safi et al, 2020). Figure 5 plots the saccade distributions of our 6 control observers on the left and 5 Binocular MD participants on the right. Given that the monocular MD participants have similar accuracy as well as similar saccade amplitude and direction distribution to controls, we do not discuss their saccades any further. The saccade distributions in Figure 5 for the binocular MD participants show the same general trend as the distributions for controls, with a greater tendency for saccades to go in the horizontal direction, compared to vertical. To compare the control and MD distributions in Figure 5, we fit each individual’s distribution with a 2-dimensional Gaussian and documented the center (x, y location) of each distribution, as well as the horizontal and vertical spread (spatial standard deviation). Center locations and standard deviations were compared between control and MD groups, using an unpaired t-test, assuming unequal variance. There was no difference in the x-location of the center (mean ± standard error: MD 0.51 ± 0.55 vs control 0.04 ± 0.26, p = 0.23), and a barely significant difference in the y-location of the center (MD 0.28 ± 0.22 vs Control 0.80 ± 0.14, p = 0.041). However, the spread of the distributions was significantly smaller for MDs in the horizontal (MD: 3.02 ± 0.75 vs Control:6.23 ± 0.37, p < 0.009) and vertical (MD: 1.56 ± 0.21 vs Control:3.3 ± 0.25, p < 0.0005) dimensions. The smaller spread of the saccade distributions for MDs can be seen in Figure 5 and is likely due to the smaller amplitude of saccades toward the scotoma as shown in Figure 3.

Figure 5.

A. Saccade distribution with respect to the fovea for control participants. B. Saccade distribution with respect to the PRL for binocular MD participants. The closed contour outlines the extent of the scotoma with respect to the PRL for each participant. Gold points mark saccades that were directed to locations within the scotoma and gray points mark saccades to visible points.

To visualize the location of the scotoma with respect to the PRL, we have superimposed scotoma outlines on the saccade distributions of each MD participant. For all participants, gray points indicate saccades to visible points (all points for controls), while gold points indicate saccades that were directed to the region hidden by the scotoma (for MD participants). Figure 4 has already demonstrated that MD participants tend to make more saccades in the general direction of their scotoma. Figure 5 goes further and shows that all binocular MD participants, no matter how small their scotoma, make saccades to regions that were within the scotoma before the saccade occurred. As these saccades could not have been driven by the presence of visual information, these data suggest that directing saccades toward the scotoma may be an adaptive strategy by binocular MD participants to uncover information hidden by the scotoma. We will examine this conjecture further in the section on Backward Saccades and in the Discussion.

3.4. Uncovering hidden information

Given that participants with binocular MD preferentially direct their saccades towards their scotomata (Figures 4 and 5), we further analyzed these saccades to determine whether their amplitude and direction (1) maximized the information uncovered, (2) followed the control distribution of saccades, or (3) were a hybrid between these two patterns.

Uncover model:

This model considered whether saccades are specifically directed toward the scotoma to uncover the region previously hidden by the scotoma and estimated how much of the scotoma was unmasked by a saccade of a particular direction and amplitude. Figure 6A shows a sample saccade that uncovers 60% of the area of the scotoma. Here we consider only saccades to the scotoma, although we are aware that saccades in non-scotoma directions can also unmask the region hidden by the scotoma. Figure 6C plots this information for one MD observer with a binocular scotoma as the proportion of the scotoma uncovered (summary data for all observers are presented in Table 2). For ease of comparison to other models we have divided the distribution into tertiles where red represents the top third of saccades that uncovered the most area, blue represents the middle third, and green represents the lowest third. As can be seen, the blue, green and red points are at increasing distance from the PRL, marking greater proportions of uncovered information.

Figure 6.

A. Schematic showing the starting position of the PRL and scotoma (light grey), the position of these after a saccade toward the scotoma (dark grey) and the fraction of the original scotoma region that is uncovered by the saccade. B. Distribution of control saccades with one observer’s scotoma profile superimposed to show what MD saccades within would look like if they followed the same distribution as controls. C. Predictions of the “Uncover” Model showing the proportion of the scotoma uncovered by saccades to various points. D. Predicted map of the Control Model where MD observers’ saccades follow the distribution estimated from control observers. E. Predicted map of a model that is hybrid between the Uncover and Control Models. F. Observer B4’s data for saccades that are directed toward the scotoma. Maps and data in Panels C through F are represented as tertiles. See text for explanation.

Table 2:

Model Comparisons: Kullback-Leibler divergence (D_KL) between the spatial distribution of saccades toward the scotoma and the three models considered in Figure 6, for each binocular MD participant. Also listed are the 99% confidence limits (lower and upper bounds) of a random model to determine if the D_KL values are significant. Significant values are indicated in bold italic.

	UNCOVER	CONTROL	HYBRID	LOWER BOUND	UPPER BOUND
B1	0.27	0.26	0.26	0.22	1.25
B2	0.69	0.44	0.55	0.52	1.48
B3	1.14	0.65	0.98	0.76	1.62
B4	0.94	0.25	0.6	0.54	0.99
B5	0.86	0.27	0.49	0.54	1

Control Distribution:

The “control” model predicts that saccades toward the scotoma simply have the same amplitude and direction distribution as control saccades, obtained from the data of our 6 control participants. The control saccade distribution in Figure 6B was generated by fitting a bivariate Gaussian to the aggregate control data and displaying the probability at locations where actual saccades occurred. Figure 6B also shows one MD observer’s binocular scotoma superimposed over the control distribution of saccades. In Figure 6D (as well in Figures 6E and 6F), the top third most-probable locations are in red, the middle third in green, and the lowest third in blue. (The probabilities of saccades within the scotoma have been normalized so that they sum to 1.). It is clear that small-magnitude saccades are the most frequent in the control distribution.

Hybrid model:

The third model predicts that saccades in binocular MD are a hybrid between making saccades to maximize the area uncovered and following the control distribution of saccades. Each dot in Figure 6E represents the point-by-point multiplication of the values in Figures 6C and 6D, followed by normalization so that the probabilities within the scotoma sum to 1.

Comparison to Data: Before we compare models to data we need to state a linking assumption for the uncover model. The uncover model is in terms of area uncovered, whereas the other models refer to the probability of saccades. The linking assumption for the uncover model is that it favors saccades that uncover the largest area covered by the scotoma, making them more likely, while it disfavors those that uncover the smallest area, making them less likely. This assumption allows us to relate area uncovered to probability and thus allows a comparison to other models and data. The fraction of area uncovered has been normalized, so that the sum across all points is 1.

Figure 6F is the actual probability that the MD observer’s saccade went to a particular location within the scotoma. A visual inspection of the predictions of the models and data for this particular observer suggest that the data are most similar to the control distribution—smaller and medium saccades (red and green) are more likely than large saccades (blue). On the other hand, the data appear to be least consistent with the uncover model, which predicts that larger saccades will maximally uncover the scotoma. To quantify the similarity between the data and the models we calculated the Kullbach-Leibler Divergence of each model from the data using the following expression

$$D_{KL}=\sum _{i=1}^{n}P\left(i\right)log\frac{P\left(i\right)}{Q\left(i\right)}$$

where P refers to the data, Q is the model (Krakov & Feitelson, 2013), and the index i references all the locations within the scotoma. The model that best describes the data is the one with the lowest divergence. Table 2 lists the D_KL values of each of the 3 models for our 5 participants with binocular scotomata. We also compared the data to a random model where each point within the scotoma was assigned a random probability, and calculated the corresponding D_KL value, after normalizing the total probability to 1. One hundred thousand iterations of the random model allowed us to estimate the 99% confidence intervals (Listed in Table 2 as the lower and upper bound). Comparison to this random model indicates whether the divergence of the uncover, control and hybrid models is significantly different from random. It is clear that for individuals with larger scotomata (B2 to B5), the divergence values are the lowest for the “control” distribution and are well below the lower confidence limit of the random model, suggesting that their saccade distribution within the scotoma was most similar to that of controls. The “uncover model” had the greatest divergence from the actual saccade data of our binocular scotoma observers, indicating that actual saccades did not optimally uncover the scotoma. The “hybrid” model had intermediate divergence, consistent with it being a combination of the control model (low divergence) and the uncover model (high divergence). While the observed distribution of saccades looks most like the control distribution, with a higher probability of smaller saccades, divergence values are non-zero, indicating that there are differences. The difference is likely related to the finding that saccades toward the scotoma have smaller amplitudes in MDs than in controls.

3.5. Sequential Saccades to Scotoma.

From the data showing that saccades into the scotoma were of smaller magnitude (Fig 4b), and the data comparing the “uncover” model to the data (Figure 6a and 6d), it is clear that single saccades are typically too small to maximally uncover the region hidden by the scotoma. To investigate whether the scotoma was uncovered over a sequence of multiple saccades, we specifically looked at the temporal sequence of individual MD participants’ saccades. Each time the PRL moves, the scotoma moves with it, and if the PRL makes small amplitude movements towards the scotoma with each saccade, it is possible that multiple saccades are directed to sequentially uncover the region originally obscured by the scotoma. Figure 7 plots the count of sequential saccades towards the scotoma (gold) compared to sequential saccades that stay in the visible region (gray), for our 5 observers with binocular scotomata. The first saccade has a significant tendency to go toward the scotoma (consistent with the data in Figure 4), but the count of subsequent saccades towards the scotoma (gold) progressively diminishes in comparison to the count for a sequence of saccades that stay in the visible region (gray). So, it does not appear that the region hidden by the scotoma is uncovered in time by a sequence of saccades towards the scotoma. (B3 is an exception and shows a similar count of sequential saccades to covered and visible regions). Of course, saccades made in any direction, including already visible areas, can uncover hidden parts of the scotoma, but it is not clear how much MD participants are aware of the scotoma region uncovered by a saccade to an already visible region (see Discussion section on the PRL being the locus of attention).

Figure 7.

Frequency of consecutive saccades that either stay in the visible region (gray), or are directed toward the region covered by the scotoma (gold). The count of consecutive saccades that stay in either of these regions progressively decreases, but consecutive saccades to covered regions typically occur less frequently than consecutive saccades to visible regions.

3.6. Backward Saccades:

So, what happens after the first saccade towards the scotoma? Figure 7 indicates that after a first saccade toward the scotoma, further saccades into the scotoma are unlikely. Thus, subsequent saccades are more likely to target the region that is no longer in the scotoma. These include saccades to regions that were previously visible, as well as saccades to regions that were just uncovered by the preceding saccade. This latter type of saccade might be an adaptation to inspect newly uncovered information and is illustrated in the schematic on the left of Figure 8A, where a saccade toward the scotoma (gold arrow) is followed by a saccade to a region that was just uncovered by the preceding saccade (black arrow). Figure 8B shows the proportion of saccades that went to the recently uncovered region. Data from our 5 MD observers show that as the size of the binocular scotoma increases (top to bottom in the first column of Figure 8B), a greater proportion of saccades go towards the region just uncovered by the previous saccade toward the scotoma. We call these backward saccades as they target regions previously covered by the scotoma.

Figure 8.

A. Schematic showing a saccade toward the scotoma on the left (gold arrow) and a saccade away from the scotoma on the right (gray arrow). In both cases, the subsequent saccade is toward the region uncovered by the preceding saccade (black arrow). B. Comparison of backward saccades when they are preceded by saccades toward the scotoma or by saccades to visible regions. The first two columns compare backward saccades to recently uncovered regions following a saccade toward the scotoma. Green bars show the data for individual binocular MD participants and orange bars show the proportion of such saccades occurring among controls with simulated scotomata (see text for details). The last two columns compare backward saccades to uncovered regions following a saccade to an already visible region for MD participants (green) and simulated controls (orange). Errors bars are standard errors of the mean.

To determine if the trend for backward saccades is specific to MD participants, we ran a simulation to determine whether controls also show a sequence of saccades that fit the pattern found in MD observers. We took each control participant’s saccade distribution and superimposed post hoc an individual MD participant’s scotoma on that distribution. Then we calculated what proportion of the control’s saccades go toward the simulated scotoma (gold arrow in Figure 8A) and are followed by backward saccades to visit the recently uncovered (simulated) region (black arrow). We repeated this calculation with the saccade distribution of all six control participants and then calculated the average probability of these “backward” saccades to obtain an estimate of how likely controls simulated with a particular MD participant’s scotoma were to generate backward saccades. We then repeated the analysis for each AMD participant, simulating their scotoma in the control group. The results of the simulation are shown as the orange bars in the second column of Figure 8B. It is clear that every MD participant (green bar), no matter the size of their scotoma, makes a greater proportion of backward saccades than do controls with the same simulated scotoma (orange bars represent mean proportion across 6 controls; error bars are the standard error of the mean).

Recall that a saccade in any direction will uncover some part the scotoma. Is the tendency for backward saccades to inspect recently uncovered regions specific to the case when they are preceded by saccades toward the scotoma, or do they occur even when saccades target already visible regions? To address this question, we examined whether MD observers made backward saccades to inspect the recently uncovered region, when the first saccade went to a previously visible region (gray arrow in schematic on the right of Figure 8A). The third column of Figure 8B shows the tendency of MD participants to look back toward the recently uncovered region after making a saccade to a visible region and the fourth column of Figure 8B shows the result of simulating MD scotomata in controls just as we did before, looking at this probability after superimposing a simulated scotoma post hoc on a control’s saccade distribution. After a saccade to a visible region, four out of 5 MD participants make similar proportions of saccades to the recently uncovered region as simulated controls, suggesting that in this case the sequence of saccades for MD participants is no different from controls. In the Discussion, we consider the reasons why individuals who use a consistent PRL, as our participants do, have a higher proportion of backward saccades following a saccade to the scotoma than a saccade to an already visible region, and whether this is a true adaptation to gather information in the task, or a pattern resulting from using a PRL that abuts the scotoma as an oculomotor reference.

4. DISCUSSION

In this study we examined eye movements during visual search in participants with macular degeneration. We were particularly interested in the pattern of eye movements with respect to the scotoma and whether there was an adaptation to uncover information hidden by the scotoma. Our study showed that observers with binocular scotomata tend to make more frequent, small- amplitude saccades toward the scotoma, whereas their saccades to regions outside the scotoma have the same amplitude as that of controls. We also found that individuals with MD have a characteristic saccade sequence where saccades toward the scotoma are often followed by backward saccades to regions just uncovered by the preceding saccade. In this section, we compare our results to other studies that have examined eye movements during visual search, and discuss whether backward saccades are truly an adaptation to inspect recently uncovered information or whether they are a related to increased attention to objects near the PRL.

4.1. Relation to other studies of visual search in the presence of a central scotoma.

Previous studies have shown that visual search performance is impacted in the presence of a central scotoma and is associated with increased search time (Kyuk et al, 2005; Mackeben & Fletcher, 2011; Wiecek et al, 2011; McIlreavy et al, 2012; Boucard et al 2015; Taylor et al, 2017; Thibault et al 2016, 2018, 2020). The display duration of 10 s in our study was typically long enough for participants to scan the entire display with eye movements and to find the targets (except for B5), so the latency to find the target, and the number of fixations is not as relevant as in studies that measured the time to find the (single) target. Previous studies have also reported decreased accuracy for visual search in the presence of a central scotoma (Thibault et al 2016, 2018, 2020), consistent with the lower search accuracy that we observed, particularly for the individual with the largest scotomata.

Furthermore, consistent with other studies that examined saccade characteristics during visual search (Taylor et al, 2017; Van der Stigchel, 2013), we find that average saccade amplitude is smaller in individuals with MD. This lower saccade amplitude is primarily due to small-amplitude saccades directed toward the scotoma, consistent with the trend reported for 2 of 4 participants in the Van der Stigchel et al (2013) study. In our study, saccades away from the scotoma had a similar amplitude to that for controls. Two previous studies tried to examine saccade characteristics with respect to a coarse estimate of the binocular scotoma (Van der Stigchel et al 2013, Janssen & Verghese, 2016). The novel contribution of this study is the measurement of eye movements with respect to a more precise estimate of the binocular scotoma, during a visual search task. On an individual basis, we are able to show that each of our MD participants with a binocular scotoma tends to direct eye movements toward their scotoma during visual search. These saccades are smaller in amplitude than control saccades and are certainly not large enough to uncover the region hidden by the scotoma in a single saccade, particularly for those individuals with large scotomata. After making a saccade toward the scotoma, individuals with CFL tend to look back at the region uncovered by the previous saccade, rather than executing sequential saccades to completely uncover the scotoma. This behavior is reminiscent of backward saccades in reading.

Limitations: Before we discuss the implications of backward saccades, we consider the limitations of our study. We measured eye movements in an eye tracker during binocular viewing and thus do not know whether MD participants used an established binocular PRL with a specified relation to the binocular scotoma. It is possible that participants used multiple PRLs during active search during binocular viewing. However, we think it is unlikely for our 5 binocular-scotoma participants for two reasons: First, they used a single retinal locus when they participated in a smooth-pursuit experiment and tracked a step-ramp target (Rashbass, 1961) monocularly, with their dominant eye. These measurements were made in a Rodenstock SLO where we could visualize the retinal locus used during both the saccade and pursuit phases of the task (Shanidze et al, 2016; Safi et al, 2020). Second, their dominant eye PRL (in the SLO) is consistent with their binocular PRL (in the Eyelink eyetracker) as shown by our measurements in these same individuals (Vullings & Verghese, 2021). Earlier studies (Kabanarou et al, 2006; Tarita-Nistor et al, 2012), also suggest that the binocular PRL also aligns with the dominant-eye PRL. Furthermore, for static binocular fixation, there is no evidence of multiple PRLs for our participants (see Figure 2b).

Another possibility is that backward saccades are an artifact of the increased fixation instability associated with eccentric viewing. An inspection of Figure 8 indicates that the probability of backward saccades increases with scotoma size, not necessarily with fixation stability as participants B1 to B4 have increasing binocular scotoma size and an increasing tendency for backward saccades, yet have comparable binocular fixation stability (Figure 2b). B5 has clearly larger fixation instability, as well as larger scotoma size. Thus, there is not a clear relationship between fixation stability and backward saccades. More importantly, a superposition of fixation stability on the amplitude distribution of saccades (Supplementary Materials, Figure S1) shows that fixation stability is much smaller than the amplitude of saccades. This is true for all 5 participants with a binocular scotoma, arguing against fixation stability as the explanation for backward saccades.

4.2. Backward/Regressive Saccades

Cummings et al (1985) and Bullimore and Bailey (1995) were among the first to demonstrate an increased incidence of regressive saccades in reading with central field loss. In the Bullimore & Bailey (1995) study, participants were reading English text and they recorded eye movements to determine how eye movements progressed across the text. Individuals with central field loss had a much lower forward saccade ratio (proportion of forward saccades relative to total saccades) compared to controls. Although scotomata were coarsely mapped for some participants, eye movements were not characterized with respect to the scotoma and “regressive” saccades were simply classified as those that went backward along the line of text. Calabrèse et al (2014) performed a careful analysis of why the letters per forward saccade is lower for individuals with central field loss and attributed an increased number of fixations as the main contributing factor. This finding suggests that some of these “excess” fixations are not in the forward direction.

A study that investigated visual search in the presence of an artificial central scotoma (David et al, 2011) measured the relative angle between successive saccades and reported that for scotoma diameters of 7° and greater, there was an increased incidence of saccades that went in the opposite direction of the preceding saccade. Cornelissen et al (2005) also found an increased incidence of “return” saccades with increasing scotoma diameter, during visual search with an artificial central scotoma. These results are consistent with our findings for individuals with real scotomata. One important addition is that unlike previous studies, we were able to measure saccades with respect to the scotoma and to show that backward saccades were much more prevalent following saccades towards the scotoma.

At this point, it is useful to consider the reasons why backward saccades occur more frequently following saccades toward the scotoma. Previously we have suggested that they may be directed towards recently uncovered information. But as we have mentioned before, information is uncovered by saccades in any direction. Why then do “backward” saccades frequently follow saccades toward the scotoma, and not saccades to already visible regions? We think a potential explanation relates to the distance of the uncovered information from the PRL in these cases. When an observer makes a saccade toward their scotoma, the uncovered region is very close to the PRL (schematic on the left in Figure 9A), but when they look away from the scotoma, the span of the scotoma separates the PRL from the uncovered region (schematic on the right in Figure 9A). Therefore, a larger-amplitude saccade is needed to reach the uncovered region following a saccade away from the scotoma compared to a saccade toward the scotoma, especially for individuals with larger scotomata (B3, B4, B5). Moreover, the span of the scotoma for these 3 individuals is larger than the median saccade amplitude of ~5° reported for visual search in natural scene backgrounds (Tatler et al, 2006; data in Figure 4). Thus, only the smallest saccades away from the scotoma are likely to be followed by backward saccades. Figure 9b plots the amplitude of a preceding saccade when it was followed by a saccade into a recently uncovered region, for each of our 5 MD observers. For preceding saccades of a given amplitude, saccades toward the scotoma (gold) were more frequent than saccades away from the scotoma (gray). This supports the conjecture that scotoma span contributes to the smaller incidence of backward saccades following saccades away from the scotoma.

Figure 9.

A. The dashed cross and the dashed outline shows the original position of the PRL and scotoma. The gold and gray arrows and corresponding scotoma profiles mark PRL and scotoma locations after saccades toward and away from the scotoma, respectively. The black arrow shows a sample saccade to the recently uncovered information in both cases. It is clear that there is newly uncovered information in both cases, but this information is close to the PRL (black cross) after a saccade towards a scotoma, but is further from the PRL after a saccade away from the scotoma. B. Histograms of the distribution of first saccade amplitudes that were followed by a saccade to a recently uncovered region. Yellow and gray depict preceding saccades that went toward and away from the scotoma (regions of overlap appear darker yellow). Alongside the individual histograms are the scotoma profiles for each observer.

4.3. The PRL as an oculomotor reference:

The increased incidence of backward saccades in participants with an established PRL compared to controls suggests that attention might begin to be associated with the PRL such that participants are more aware of information that it is uncovered near the PRL. This suggests that the PRL may begin to inherit some of the properties associated with the usual oculomotor reference—the fovea. Classic studies have demonstrated an increase in sensitivity at the location of an upcoming saccade target, just prior to a saccade, showing a close coupling between the allocation of attention to the execution of a saccade in space and time (Kowler et al, 1995; Deubel & Schneider, 1996). Each of our MD participants had an established binocular PRL, so it is possible that this extra-foveal oculomotor reference (White & Bedell, 1990) is beginning to be associated with attention. If true, this might explain why participants become more aware of information that is uncovered close to the PRL.

On the other hand, it can be argued that because peripheral vision is poorer than foveal vision and there are multiple regions with residual vision that have similar acuity, the PRL does not have a privileged status. In the context of our task, the target Gaussian blob was large enough (spatial standard deviation of 0.5°, or a subtense of about 2°) to be seen with peripheral vision, when it was not obscured by the scotoma. Thus, the PRL of MD observers might not necessarily have conferred an acuity advantage for the task. However, if the locus of spatial and temporal attention is associated with the PRL, it might explain why eye movements are drawn to recently uncovered information in the vicinity of the PRL, even when it does not confer an acuity advantage such as looking with the old fovea. Thus, backward saccades toward recently uncovered information may occur in part because of the proximity of newly uncovered information to the PRL, and not solely as a strategy to recover missing information as in the case of reading in MD (Bullimore & Bailey, 1995).

The goal of the study was to determine the oculomotor characteristics of the PRL when MD observers used eye movements to find an unknown number of targets in a visual search display. Our results show two clear patterns. First, the PRL is directed toward the region occupied by the scotoma much more in MD than the fovea is directed to comparable (intact) regions of the visual field in controls. Second, the PRL is directed to recently uncovered regions in its vicinity, a pattern suggestive of attention being associated with the PRL.

Figure 1:

Timeline of a trial. Trial start required the observer to fixate the central marker and press a button. A full-screen image was presented for 10 s, during which observers searched for Gaussian blobs while eye movements were recorded. This particular trial shows 2 blobs (marked by circular outlines for purposes of illustration only; the outlines did not appear in the display). After the trial ended, participants reported the number of blobs.