Retinal phenotyping of variants of Alzheimer's disease using ultra‐widefield retinal images

Abstract Background Posterior cortical atrophy (PCA) is the most common atypical variant of Alzheimer's disease (AD). Changes associated with PCA in the brain affect the visual cortex, but little is known about retinal changes in PCA. In this study, we explored retinal phenotypic variations in typical AD (tAD) and PCA. Methods Retinal phenotyping was carried out on ultra‐widefield (UWF) images of 69 control, 24 tAD, and 25 PCA participants. Results Individuals with tAD (odds ratio [OR] = 2.76 [confidence interval (CI):1.24 to 6.10], P = .012) and PCA (OR = 3.40 [CI:1.25 to 9.22], P = .016) were more likely phenotyped as hard drusen. tAD (OR = 0.34 [CI:0.12 to 0.92], P = .035) were less likely to have soft drusen compared to control. Almost 3‐fold increase in reticular pseudodrusen formation in tAD (OR = 2.93 [CI:1.10 to 7.76], P = .030) compared to control was estimated. Discussion Studying the peripheral retina may contribute to a better understanding of differences in retinal phenotypes of different AD variants.

drusen under the retinal pigment epithelium (sub-RPE) in the peripheral retina in AD, 7 and laboratory observations also reported peripheral retinal changes in the neurosensory retina. 8 Deposit formation between the RPE and the retina (sub-retinal) called reticular pseudodrusen (RPD) has been associated with outer retinal atrophy. 9 Peripheral reticular pigmentary degeneration (PRPD), another retinal imaging feature, is associated with compromised systemic circulation and choroidal vascular insufficiency. 10 These, and sub-RPE deposits, are features that can be readily identified on UWF images. 11,12 We propose that monitoring these retinal changes on UWF images could help improve patient stratification in AD. In this study, we examined patients with tAD and PCA using UWF images and compared these to controls to identify retinal phenotypic differences.

METHODS
We enrolled 29  Tang-Wai et al. 15 proposed clinical criteria based on available information at baseline visit and expert retrospective clinical review. In a subset of participants (PCA = 18; tAD = 14), the clinical assessment was also confirmed by biomarker evidence using positron emission tomography (PET) and cerebrospinal fluid (CSF), and it fulfilled the amyloid PET and CSF AD profile criteria. 16 Ethical approval was provided by the National Research Ethics Service Committee London Queen Square; all participants provided written informed consent.
Color UWF images were acquired using the Optomap P200Tx SLO (Optos Plc) without pupil dilation. After quality control (QC), images

HIGHLIGHTS
• First ultra-widefield retinal imaging study that shows improved patient stratification • Both tAD and PCA associated to increased peripheral hard drusen formation • tAD associated to decreased peripheral soft drusen formation • tAD associated to increased peripheral reticular pseudo- deemed acceptable for grading were stereographically projected to compensate for distortions due to the retinal curvature using the Optos Projection Tool. 17 After projection, images were graded for hard and soft drusen, RPD, PRPD, geographic atrophy (GA), pigment epithelial detachment (PED), and choroidal neovascularization (CNV),

F I G U R E 1
Representative ultra-widefield images used for grading. Ultra-widefield composite (A1 and A2) and red-free (A3) image of a mixture of hard (white arrowheads) and soft (black arrowheads) drusen. The area outlined on A1 on the nasal periphery is visible on A2 and A3. Ultra-widefield composite (B1 and B2) and red-free (B3) image showing RPD (reticular pseudodrusen) at the superior hemisphere (black arrowheads). The area outlined on B1 is visible on B2 and B3. Ultra-widefield composite (C1 and C2) and green-free (C3) image showing PRPD (peripheral reticular pigmentary degeneration; black arrowheads). The area outlined on C1 in the superonasal periphery can be seen on C2 and C3. Drusen and RPD were often assessed using the red-free images, while PRPD was often evaluated using the green-free images as these give better contrast between the pathology and the rest of the image retinal phenotypes that are readily detectable on UWF images. 12 Sub-RPE deposits, appearing as yellowish patches on color images and gray patches on red-free images, were graded as hard (< 125 μm in size) or as soft drusen (> 125 μm) 18  . PRPD is also referred to as peripheral reticular degeneration (PRD). 12 Images with questionable pathological changes were adjudicated by a retinal specialist (T.P.). Two independent graders carried out the grading (N.Q. and L.C.) masked to the participants' case-control status.
The high spatial resolution grading ( Figure S1 in supporting information) was carried out after overlaying a grid of squares (that will be referred to as Manchester Grid [MaG]) in which the area of all squares (754 squares per image) equaled the size of the optic disc, using the Manchester grid tool (Optos, version r6076). 20 During grading, each square was assessed for the presence (1) or absence (0) of the different pathological features outlined above. A square was defined as ungradable if more than 50% of the square area was impossible to assess.
After data extraction and summation, heatmaps were generated for visual assessment for the regional distribution of pathologies in the back of the eye ( Figure 2). The results from patients with tAD or PCA were subtracted from the prevalence in control. These differential plots are shown in Figure 3. Blue indicates features higher in control, while red indicates features higher in patients.  Figure S1. After conversion, heatmaps for all grading categories were plotted for distribution ( Figure 4) and differences ( Figure 5).
For direct comparison with our previous study, 7 we carried out hierarchical phenotyping as well: Those with hard drusen only were assigned to the hard drusen phenotype; those with both hard and soft drusen were designated as soft drusen phenotype.

Data analysis
All data analysis was conducted using SPPS (version 26.0; SPSS Inc.).
When assessing differences in study characteristics, the Chi-square test was used for categorical variables and one-way analysis of variance for continuous variables. General estimating equation (GEE) enabled data from both eyes to be included in the binary logistic regression analysis, 22 which was used to assess the relationship between retinal pathologies and diagnosis, with control as a reference group. The size of the ungradable area was recorded for each

Prevalence of pathologies
The difference in ungradable area size on UWF images between the three groups did not reach statistical significance (P = .068; Table 1 Four eyes in the control, two eyes in the tAD, and one eye in the PCA group showed no detectable pathological changes on our UWF images (   (Table 2).
There was no GA, PED, or CNV detected in any of the eyes in this study (Table 2).

Distribution of pathological features using the MaG
Sub-RPE deposits were distributed with a preference toward zone 5, especially at the superonasal (SN) and temporal quadrants (Figure 2A-C). While hard drusen were the most frequent pathological feature, after hierarchical phenotyping, the distribution of the harddrusen-only phenotype was detected in the nasal periphery, primarily in zone 5 ( Figure 2B). Similarly, soft drusen were most prevalent in the nasal far-periphery in zone 5 ( Figure 2C).
RPDs were distributed mainly on the superior retinal quadrants in zones 4 and 5 ( Figure 2D). PRPD featured mainly at the nasal retinal far-periphery (zone 5; Figure 2E), although in PCA, there was a distinct ring appearance of PRPD in zone 5. When all pathologies were combined, we found that retinal pathologies were present throughout the retina with an enrichment in the far nasal periphery ( Figure 2F).
Kappa statistics showed moderate agreement between graders for all the graded pathologies, with PRPD showing the highest (κ = .573) and hard drusen the lowest (κ = .440) agreement between the two graders.
To identify disease-specific changes, tAD and PCA patients' prevalence values were subtracted from those of control ( Figure 3). The prevalence of hard-drusen-only phenotype was higher in patients than control ( Figure 3A). However, the prevalence of soft drusen was lower in patients than in control, especially in the nasal quadrants ( Figure 3B).
RPD prevalence was higher in tAD in the superior quadrants, while a mixed picture can be seen in PCA ( Figure 3C). PRPD was higher in PCA, especially in the temporal quadrants, while it appeared lower in tAD, especially in the nasal quadrants ( Figure 3D).

Distribution of pathological features using the MoG
Similar to our previous findings, the prevalence of pathologies was higher in the peripheral retina, especially in the nasal quadrants (Figure 4). To better visualize the differences between patients and control, we subtracted the prevalence values from those in control and plotted the differences in Figure 5. Hard drusen had the highest prevalence in Zone 5 in the SN quadrant in tAD ( Figure 5A). The prevalence of peripheral soft drusen was lower in the nasal quadrants, both in tAD and PCA ( Figure 5B). In tAD, PDR was higher in the superior quadrant with little difference observed in PCA ( Figure 5C). In contrast, the prevalence of PRPD was higher in zone 5 in PCA but lower in tAD ( Figure 5D).
Using GEE analysis, we found that individuals with tAD were twice as likely to have hard drusen phenotype in zone 5 compared to the   Figure S3).  Figure S3). There was no significant difference detected in RPD between PCA and control (Table 4, Figure S3).
While on the visual representations there appeared to be an increased prevalence of PRPD in PCA, but not in tAD compared to control ( Figure 5D), the difference did not reach statistical significance in any of the zones or quadrants (Table 4, Figure S3).
The same results were obtained when the model was or was not adjusted for ungradable area size, apart from soft drusen in the IN quadrant of zone 5, when tAD was compared to control. In this case, a significant difference was only detected in the adjusted final GEE analysis.

DISCUSSION
We have previously shown that peripheral retinal hard drusen were associated with AD. 7 This study expanded on this finding. We analyzed UWF images using a higher resolution grading grid to identify potential "hot spots" for pathology. This more detailed grading then was transformed into the low-resolution grid we used previously. 7 We included retinal phenotypes that were not considered earlier, and it appears that RPD and PRPD could become distinct phenotypes to stratify patients with AD.
Undoubtedly, the higher resolution MaG grading grid allowed us to generate more nuanced maps of pathological changes on UWF images (Figures 2 and 3). However, developing these maps was time consuming and resource intensive and may not be feasible in a clinical setting unless automated grading algorithms are developed.
In the current study, patients were stratified into typical AD and PCA, based on consensus criteria for PCA 2 and NIA-AA criteria for tAD. 13 Despite dividing patients into variants, we found that both patients with tAD and PCA had a higher prevalence of hard druse-only phenotype in the far periphery, especially in the SN quadrant, supporting our previous findings on a general AD population. 7 We also found a lower prevalence of soft drusen phenotype at the retinal far-periphery in tAD, but not in PCA, compared to control. A trend for lower soft drusen prevalence was also present in the general AD population. 7 Drusen is considered a hallmark for age-related macular degeneration (AMD), and lipids and apolipoprotein E (apoE) have been associated with drusen formation and progression. [24][25][26] Although the exact role of apoE in AMD is still under intense investigation, 27 We believe that this is the first study to assess RPD for AD. RPD (also known as SDD) was first recognized by Arnold et al. in post mortem tissues with AMD. 33 Later, it was shown that RPD was associated with AMD in vivo and conferred an increased risk to progress to late AMD. 34 In population-based studies and cohorts of AMD patients, RPD prevalence showed a considerable variation. [35][36][37][38] In our study, the prevalence of RPD in control and PCA were ≈16% but significantly higher in patients with tAD (37.2%). This finding appears to associate RPDlike pathology with tAD, especially in the superior hemisphere in the peripheral retina. The molecular mechanism of RPD formation is not yet fully elucidated. 39 It is clear that RPDs are different from sub-RPE deposits, 39 but how the molecular composition of RPD might be linked with tAD and what the difference in retinal phenotype between tAD and PCA means is yet to be fully explored. RPD can be detected using OCT, which shows an accumulation of debris internal to the retinal pigment epithelium in the macula. 39 The detection of peripheral and far peripheral changes with current OCT cameras, however, is challenging.
A small cross-sectional study identified retinal features that are positive for curcumin and most commonly present in the superior periphery of the retina in patients with AD. 8 After assessing these lesions using OCT, the authors could not rule out the possibility that the curcumin signal originates from RPD. 8 The recent development of a UWF-guided swept-source OCT, which enables OCT imaging at any part of the retina, giving a more comprehensive analysis of the retinal periphery, 40  Our study's strengths are that our cohort only included wellcharacterized tAD and PCA patients, and these were compared to a sizeable control population. In addition, we graded all eyes with sufficient quality and used GEE analysis that appropriately accounts for the correlation between the two eyes, 22 providing a more comprehensive analysis. Also, we used a high-resolution grading grid to refine our phenotypes.
In summary, our study highlighted the need for better patient stratification through more precise phenotypic characterization, which could include UWF imaging of the retina. These could help the success of future dementia trials to reduce disease heterogeneity. We should look beyond the macula to better understand the link between retinal and brain pathologies.

ACKNOWLEDGMENTS
We are deeply indebted to the participants for their generous dona- Alimera; Oxurion; Roche.

CONFLICTS OF INTEREST
LC was supported by an unrestricted PhD studentship from Optos plc.
LC is currently employed by Optos plc.