Quantitative analysis of optical coherence tomography imaging in patients with different severities of hydroxychloroquine toxicity

Abstract

Purpose

To determine the diagnostic validity of quantitative measures derived from optical coherence tomography (OCT) images in their ability to discriminate between cohorts of eyes unaffected by hydroxychloroquine (HCQ) and those with a range of toxicity severities, including mild toxicity.

Methods

Prospective, single-centre, case–control study conducted between August 2010 and May 2017. Participants were exposed to HCQ for at least 5 years (mean±SD =14±7.2 years) and classified into affected and unaffected cohorts based on the American Academy of Ophthalmology’s 2016 recommendations. For affected eyes, severity (groups 1–4) was assigned based on the extent of ellipsoid zone loss. For all eyes, spectral domain-OCT scans were analysed quantitatively to compute inner retinal thickness (IRT), outer retinal thickness (ORT), and minimum signal intensity (MI) and compared across toxicity groups.

Results

Of the 85 participants (mean age 59±12 years, 93% female), 30 had retinal toxicity. Significant differences in ORT and MI were observed between each affected severity group and unaffected eyes. Significant differences in IRT were observed for groups 3–4 but not groups 1–2. ORT and MI were each able to discriminate between unaffected and group 1 eyes with the highest discrimination at the inner subfields (areas under the curve, AUC=0.96 for ORT and AUC=0.93 for MI).

Conclusions

Quantitative analysis of OCT scans revealed significant differences between eyes with and without toxicity in two different measures. Each individual metric could discriminate between the unaffected and the lowest severity category, suggesting their potential utility in screening for HCQ toxicity in patients at risk.

INTRODUCTION

Hydroxychloroquine (HCQ) is used widely in rheumatology and plays an integral role in the treatment of autoimmune diseases, including systemic lupus erythematosus and rheumatoid arthritis, as well as other inflammatory and dermatologic disorders.^{1 2} Moreover, its indications are expanding, since it is now used as an adjuvant therapy to treat certain cancers and even explored for treatment or prophylaxis against some viral infections.^{3 4} A limiting side effect is its potential to cause a retinopathy by altering the retinal pigment epithelial (RPE) cells’ ability to process the phagocytosed outer segments of photoreceptors.^5-8 The retinopathy is not only irreversible but can even worsen after cessation of HCQ,^8-10 which highlights the need for surveillance in detecting retinal toxicity at the earliest possible time point.

In this context, the development of automated, objective, and sensitive tests to detect the earliest signs of retinal toxicity is invaluable to the safe administration of the drug.² Currently, the American Academy of Ophthalmology (AAO) recommends several subjective and objective tests in screening for HCQ retinopathy and highlights the importance of spectral domain optical coherence tomography (SD-OCT).^{1 11-13} It recommends qualitative inspection of the SD-OCT B-scans for evidence of paracentral disruption of the ellipsoid zone (EZ) reflectivity band, as this correlates with a number of functional measures of HCQ-induced retinopathy.^{1 8 14 15} However, qualitative analysis of SD-OCT scans has two main drawbacks: first, it is subjective and highly reliant on expert evaluation by retinal specialists, potentially resulting in suboptimal reproducibility; second, it may not exploit the full three-dimensional wealth of information available from OCT scans, potentially leading to decreased sensitivity.

By contrast, the exploration of quantitative SD-OCT metrics has the potential to identify structural changes that are clinically relevant but beyond what is readily apparent with qualitative inspection. Retinal thickness alterations have been explored in patients with HCQ retinopathy and our previous work also introduced SD-OCT MI measurements as an independent method to identify outer retinal changes in HCQ toxicity.¹⁶ However, for these metrics to be clinically useful, they need to demonstrate success in discriminating between eyes with no toxicity and those with a range of toxicity severities, ideally including those with very early toxicity.

In this study, we assessed how quantitative OCT metrics correlate with the severity of HCQ retinopathy. The determination of toxicity followed the AAO’s updated 2016 recommendation on screening tests and relied on combination of findings from subjective testing (perimetry) and objective testing (OCT and multifocal electroretinography (mfERG)).¹ Those eyes with toxicity were stratified into severity groups, based on evaluation of foveal OCT B-scans. In order to facilitate the development of objective screening tests for toxicity, the sensitivity of OCT-derived retinal thickness and image intensity-based quantitative metrics to identify toxicity were compared among the unaffected group and affected groups with different severities, with special attention to the group with the lowest severity (ie, mild).

METHODS

Study participants

This prospective, case–control study was conducted at the National Eye Institute (National Institutes of Health, Bethesda, Maryland, USA).The inclusion criteria comprised a current or previous history of HCQ treatment for a total duration exceeding 5 years and the absence of concomitant retinal disorders (eg, diabetic retinopathy, retinal vein occlusion, age-related macular degeneration, or Stargardt’s disease). The design of the study was a case–control study aimed at studying a relatively rare outcome and not designed to reflect the prevalence of the disease. Information on patient demographics and clinical characteristics including body weight and height, duration and cumulative dose of HCQ therapy, and diagnostic indications for HCQ treatment were obtained by medical chart evaluation. Previous analyses have reported on subsets of this cohort.^{16 17}

Study procedures

All participants underwent a comprehensive ocular examination, including best-corrected visual acuity testing using the Early Treatment Diabetic Retinopathy Study (ETDRS) protocol, slit-lamp examination, and dilated fundus examination. In addition, all participants underwent mfERG, automated 10–2 Humphrey perimetry, and multimodal retinal imaging including SD-OCT. Testing was performed in both eyes of all participants.^{16 17} mfERG testing was performed according to the International Society for Clinical Electrophysiology of Vision guidelines,¹⁸ based on the 61-hexagon stimulus pattern of the VERIS Clinic system (Electro-Diagnostic Imaging, Redwood, California, USA). The SD-OCT images were acquired using the Heidelberg Spectralis HRA-OCT system (Spectralis; Heidelberg Engineering, Heidelberg, Germany) using two acquisition protocols. The first protocol consisted of two line scans through the fovea (one horizontal and one vertical, length 30°, averaging 100 frames) for each eye; these were used for severity grouping. The second protocol comprised a macular cube scan (121 horizontal B-scans with60 μm spacing, comprising 30° horizontally and 25° vertically) for each eye; these were used to calculate quantitative metrics.

Determination and severity scale grading of retinal toxicity

The determination of toxicity was based on the AAO’s 2016 recommendations that include a combination of at least one objective test (OCT or mfERG) confirming an abnormality on a subjective test (automated perimetry).¹ The Humphrey visual field tests were interpreted as positive if there was a focal deficit, defined as the presence of at least three contiguous abnormal points on the pattern deviation map (p<2%) or a full ring scotoma.¹⁹ OCT B scans were interpreted as positive if there was evidence of photoreceptor loss and EZ disruption.¹ The mfERG tests were interpreted as positive in the presence of either an increased ring 1:ring 2 ratio greater than 2.6 or a reduced central absolute amplitude less than 35 (nanovolts/degree).^{2 16 17 20}

Eyes with toxicity were classified into four severity groups, based on the foveal SD-OCT line scans, as follows (figure 1): Group 1 consisted of scans with either no or <100 μm loss of the EZ band, group 2 with EZ loss ≥100 μm but <1 mm, group 3 with EZ loss ≥1 mm and group 4 with a foveal island of ≤500 μm intact EZ. EZ loss was defined as complete absence of the EZ band, such that areas of EZ band attenuation without complete absence (as occurs in some cases of early severity) did not qualify. For grading, the foveal horizontal and vertical OCT line scans were presented in a random order to three independent masked graders (two retinal specialists, CC and TK, and one trained grader, RM), with any discordant scores resolved by consensus after joint review. The scores for the horizontal scans were utilised for all subsequent analyses, since these are used more commonly in clinical practice. In participants with different severity scoring between the two eyes, the more severe group was used to describe the participant.

Figure 1.

System for grading severity of spectral domain optical coherence tomography (SD-OCT) foveal line scans, with a representative example for each severity group. (A) Group 1 scan with no to <100 μm loss of EZ on either side of the fovea. The open arrows indicate areas of abnormality without complete EZ disruption. (B) Group 2 scan with EZ loss ranging from 100 μm to 1 mm on either side of fovea. The solid arrows denote EZ loss. (C) Group 3 scan with EZ loss >1 mm on either side of the fovea but with >500 μm foveal island of EZ remaining. (D) Group 4 scan with <500 μm of the foveal island of EZ remaining. EZ, ellipsoid zone.

Quantitative metric computation from OCT

SD-OCT volume data were exported from the Heidelberg archives and automatically segmented using OCT-Explorer (V.3.8, University of Iowa, Iowa, USA) to generate three contours: (1) inner limiting membrane (ILM), (2) outer plexiform layer (OPL; outer boundary of the inner nuclear layer) and (3) RPE (outer boundary). These contours were manually inspected for accuracy and adjusted (7/169 eyes) when necessary.

Mean retinal layer thickness and minimum signal intensity (MI) were quantified for each of the following layers: (1) ILM to RPE (total retina), (2) OPL to RPE (outer retina) and (3) ILM to OPL (inner retina). Prior to computing the MI measurements, the image intensities were normalised by performing two image preprocessing steps. First, an edge-preserved smoothing operation was applied using a bilateral filter to decrease noise. Second, the image intensities were multiplied by a scaling factor such that the average intensity of the inner retina was constant across the images. The MI measurements of the total retina (between the ILM and RPE) were then calculated along each A-scan within the coordinates of the segmented contours. In a clinical setting, these metrics can be computed automatically when the SD-OCT images and the contours are available without requiring any human effort or time.

Mean measurements were calculated for each ETDRS subfield, labelled as follows: (1) foveal, (2) inner inferior, (3) inner temporal, (4) inner superior, (5) inner nasal, (6) outer inferior, (7) outer temporal, (8) outer superior and (9) outer nasal.

En face projection images from OCT

The SD-OCT volume scans were evaluated with automated algorithms. The approaches investigated included segmentation of the retinal layers with designation of the inner retina and outer retina and an application of MI algorithm. The implementation of these algorithms is shown in one eye with group 1 retinopathy in figure 2. The foveal OCT scan reveals mild parafoveal EZ loss with overlying ONL thinning (figure 2A,C). In figure 2A, the contours of the segmentations are shown dividing the neural retina into inner retina (red contour to green contour) and outer retina (green contour to orange contour) components. The decreased outer retinal thickness (ORT) in this area manifests as a darker (thinner) paracentral ring in the 2D en face thickness projection image (figure 2B). Figure 2C illustrates the application of the MI algorithm which detects the lowest intensity value indicated in yellow occurring in the ONL. The parafoveal area demonstrates an increase in this intensity measure in the parafoveal region visible as a lighter ring on the 2D en face MI projection image (figure 2D).

Figure 2.

An example of the implementation of segmentation and minimum intensity algorithms in an eye with group 1 retinopathy. (A) OCT B-scan of a group 1 eye with segmentations for ILM (red), inner boundary of the OPL (ie, INL-OPL boundary, green), RPE (orange). White arrows indicate areas with decreasing outer retina thickness. (B) Projection image of outer retina thickness measures in millimetres (INL-OPL to OB_RPE) across the macula. Darker parafoveal ring indicates area of thinning (C) OCT B-scan of the same group 1 eye with yellow points indicating the location of the pixels with the minimum intensity detected between the ILM and outer boundary of the RPE. (D) Projection image of the minimum intensity values across the macula. Brighter parafoveal ring indicates area of increased minimum intensity. ILM, inner limiting membrane; INL, inner nuclear layer; OCT, optical coherence tomography; OPL, outer plexiform layer; RPE, retinal pigment epithelial.

Statistical analyses

OCT quantitative measurements (ORT, inner retinal thickness (IRT) and (total retina) MI) were compared between the five groups (ie, unaffected and the four severity groups) at each ETDRS subfield, using a one-way analysis of variance (GraphPad Prism V.8.3.0 (538)). Receiver operating characteristic (ROC) analysis was used to assess the ability of each individual metric to discriminate between the unaffected and affected groups including the least severe affected group (ie, group 1).

Random forest (RF) analysis was performed to evaluate the discriminability between different groups when feature sets are were used in combination to build a binary classification model. First, each metric’s ability (ORT, IRT, MI) was assessed using the values of the nine ETDRS subfields as predictor variables. Second, all features were used in combination in a model constituting 27 predictor variables in total. Models were trained separately to distinguish affected and unaffected groups as well as affected-group 1 and unaffected group. Each binary classification model was evaluated in a 10-fold cross-validation setting with the parameter maximum depth-of-trees=3.

RESULTS

Of the 85 participants, 55 participants had no sign of toxicity in either eye whereas 30 participants had evidence of toxicity. One eye of an affected participant was excluded due to poor quality OCT scans. Patients with toxicity had toxicity determinations in both eyes except for 4 patients where one eye had positive mfERG testing but visual field defects less than three contiguous points on the pattern deviation. The 59 eyes with HCQ retinopathy were divided into four severity groups, according to prespecified definitions, as follows: 16/59 (27%) group 1, 8/59 (14%) group 2, 24/59 (41%) group 3, and 11/59 (25%) group 4.

The functional testing, consisting of visual acuity and visual field data, are shown separately in table 1 according to the presence and severity of HCQ retinopathy. Eyes with group 3 and group 4 retinopathy had lower visual acuity (p<0.0001) than unaffected eyes. Visual field 10–2 mean deviation was significantly different between unaffected eyes and all toxicity severities (p<0.0001).

Table 1.

Eye-specific findings of unaffected, stage 1, stage 2, stage 3 and stage 4 eyes

	Unaffected	Stage 1	Stage 2	Stage 3	Stage 4
N	110	16	8	24	11
Average visual acuity, letters (Snellen equivalent)	85.6 (20/20)	84.6 (20/20)	85.9 (20/20)	80.7* (20/25)	67.4* (20/50)
Average Humphrey visual field mean deviation, dB	−0.8†	−4.0*	−3.6*	−12.5*	−20.9*

^*Indicates p<0.05 when compared with unaffected group.

^†Two eyes did not have HVF data.

Humphrey visual field, HVF.

The characteristics of the study population are shown separately in table 1 according to the presence and severity of HCQ retinopathy. Participant age was significantly higher in those with group 4 toxicity than in unaffected participants. Daily HCQ dose was numerically but not statistically higher in participants with toxicity group 4 than in unaffected participants (table 2).

Table 2.

Demographics of unaffected, stage 1, stage 2, stage 3 and stage 4 participants

Demographics	Unaffected, N=55	Stage 1, N=7	Stage 2, N=4	Stage 3, N=12	Stage 4, N=7	P value
Age, years, mean±SD	56.5±11.6	60.7±16.3	54.0±16.8	64.7±8.9	69.6±5.1*	0.02
Patients older than 60 years (%)	23 (41.8)	3 (42.9)	1 (25.0)	8 (66.7)	7 (100.0)*	0.03
Sex, female (%)	51 (92.7)	6 (85.7)	4 (100.0)	11 (91.7)	6 (100.0)	0.83
Asian ethnicity (%)	4 (7.3)	0 (0.0)	1 (25.0)	1 (8.3)	2 (28.6)	0.27
Length of treatment, years, mean±SD	14.5±7.4	16.7±9.4	12.2±3.9	12.0±5.4	16.7±7.7	0.77
Height, cm†	169.9±9.8	161.2±7.3	156.8±4.0	162.9±10.9	160.7±3.8	0.78
Weight, kg†	72.9±16.4	77.5±42.8	68.0±24.2	71.8±22.4	53.2±8.4*	0.04
Daily dose, mg/kg real body weight/d †	5.1±1.6	4.8±1.8	5.5±2.1	6.0±1.5	6.5±1.8	0.31
Indication for HCQ use, n (%)
Lupus	38 (69.1)	7 (100.0)	4 (100.0)	6 (50.0)	3 (42.9)	0.06
Rheumatoid arthritis	16 (29.1)	0 (0.0)	0 (0.0)	4 (33.3)	3 (42.9)	0.26
Sjögren’s syndrome	1 (1.8)	0 (0.0)	0 (0.0)	2 (13.3)	0 (0.0)	0.13

^*P<0.05 in comparison to unaffected group.

^†One patient from groups 1, 3, 4 missing height and weight data

HCQ, hydroxychloroquine.

Quantitative analyses

The quantitative SD-OCT metrics were compared between the unaffected and all affected eyes at each ETDRS subfield. Eyes in the affected group demonstrated thinner ORT in all ETDRS subfields (p<0.001). MI values were also significantly different and increased in affected eyes compared with unaffected eyes in all ETDRS subfields (p<0.001). IRT demonstrated thinner inner retina in the affected group in all ETDRS subfields except ETDRS1 (p<0.001 ETDRS2-7, ETDRS9, p=0.02 ETDRS8).

As the goal of using OCT metrics for screening is to evaluate the performance across a range of toxicity severities, the OCT metrics were evaluated by comparing unaffected eyes to each toxicity severity group (figure 3). ORT was significantly lower in each severity group 1–4, compared with the unaffected group, for all ETDRS subfields (p<0.01 for all except ETDRS 8 group 1 p=0.013). Inner retinal thickness was significantly lower (p<0.01) in all subfields in group 4 eyes and for most (ETDRS subfields 2–7 and 9) in group 3 eyes.

Figure 3.

Quantitative optical coherence tomography measurements in unaffected (black), group 1 (purple), group 2 (orange), group 3 (green) and group 4 (blue) eyes according to ETDRS subfield. (A) Comparison of outer retina thickness (p<0.05, +p<0.01, *p<0.0001). (B) Comparison of inner retinal thickness (+p<0.05, *p<0.001). (C) Comparison of minimum intensity (o p<0.05, +p<0.01, *p<0.0001). ETDRS, Early Treatment Diabetic Retinopathy Study.

The MI values were generally the lowest in the unaffected group and tended to increase with increasing severity. Comparisons revealed significant increases of MI in all severities of toxicity in the inner ETDRS subfields (2-5). Groups 1, 3 and 4 were also significantly greater (p<0.01) than unaffected eyes in subfields ETDRS 6, 7 and 9. The central 1 mm (ETDRS1) and outer subfields ETDRS8 demonstrated significant MI increases in groups 3 and 4 compared with unaffected eyes.

The ability of SD-OCT retinal thickness and MI metrics to discriminate between the mildest severity group (group 1) and unaffected eyes was investigated via ROC by comparing the areas under the curve (AUC). The AUC values and their respective 95% CIs are summarised in table 3. ORT demonstrated excellent (AUC >0.93) discrimination between group 1 and unaffected eyes, when considering the inner subfields ETDRS 2–5, with the highest AUC (0.97) found in ETDRS2 (inner inferior subfield). MI also provided very good (AUC >0.88) discrimination, when considering the inner subfields ETDRS 2–5, with the highest AUC (0.94) in ETDRS5 (inner nasal subfield).

Table 3.

Performance of the ROC curve for unaffected versus group 1 eyes

ETDRS subfield	AUC	95% CI
ORT
ETDRS 2	0.97	0.92 to 1.00
ETDRS 3	0.95	0.90 to 1.00
ETDRS 4	0.93	0.87 to 0.98
ETDRS 5	0.94	0.90 to 0.99
MI
ETDRS 2	0.89	0.80 to 0.97
ETDRS 3	0.89	0.82 to 0.96
ETDRS 4	0.88	0.81 to 0.95
ETDRS 5	0.94	0.89 to 0.98

AUC, area under the curve; ETDRS, Early Treatment Diabetic Retinopathy Study; MI, minimum intensity; ORT, outer retinal thickness; ROC, receiver operating characteristic.

Table 4 shows the performance of RF analysis evaluated in terms of AUC when different feature sets were used in combination to classify binary disease status (ie, affected vs unaffected). In this setting, all metrics exhibited excellent discriminability (AUC >0.88) when classifying affected and unaffected groups. Slightly inferior performance of RF analysis can be attributed to the cross validation setting where the RF model was evaluated in a masked test data set, whereas the previous ROC analysis was performed using all the data points to construct the model. When classification was restricted to group 1 versus unaffected, both the ORT and the combined models exhibited AUC=0.83. Since there were only 11 data points representing group 1, random splits of train/test may not have comprised enough training samples to learn generalised models, and thus could explain the inferior performance in the test sets. It was also interesting to note that for both types of classification, the combined model (with 27 predictor variables) did not improve performance beyond the model using only 9 ORT predictor variables. This indicates the variability of the outcome is captured entirely by ORT in our data set.

Table 4.

Comparison of AUC values for different models in random forest analysis

RF model	IRT	ORT	MI	Combined
Affected/unaffected	0.88	0.95	0.90	0.95
Group1/unaffected	0.64	0.83	0.66	0.83

AUC, area under the curve; IRT, inner retinal thickness; MI, minimum intensity; ORT, outer retinal thickness; RF, random forest.

DISCUSSION

In this work, we demonstrated that quantitative SD-OCT metrics can distinguish between eyes without toxicity and those with a range of severities of toxicity, including the most mildly affected. We identified two separate metrics, ORT and MI, that can each individually distinguish between unaffected eyes and those with very mild toxicity and thus have the potential to detect retinopathy. The two metrics quantify different aspects of the retinal OCT image. ORT measures the distance between reflectivity bands, while MI measures the change in intensity of the ONL as the degeneration causes collapse of this layer. Identifying abnormalities using two different quantitative OCT metrics could enhance the confidence of the diagnosis while capturing different characteristics of the underlying pathologic changes. The robustness of detecting toxicity could thus be improved when the two metrics are used in combination, providing independent, objective additions to the current clinical decision-making process.

An interesting finding in this study was that outer retinal thinning occurs in eyes with the mildest toxicity, while both outer and inner retinal thinning is found in eyes with severe toxicity. While outer retina changes are the focus in HCQ toxicity^{10 21 22} as highlighted by the AAO screening recommendations, studies evaluating on inner retina changes have reported varying results. Some studies have reported on a lack of inner retina changes in toxicity,¹⁰ whereas others have found evidence of inner retina thinning in cases of toxicity.^{21 22} Our analyses, which investigate changes across severity groups, support that the predominant changes associated with toxicity occur in the outer retina. They also demonstrate that eyes with mild toxicity do not incur inner retina changes but that eyes with severe toxicity do demonstrate inner retina thinning. While potential errors in segmentation are a limitation to the conclusions, this data suggests that outer retinal cell layers are involved earlier in the degeneration process, with inner retinal sequelae occurring in more severe toxicity. Spatially, inferior, temporal, and nasal ETDRS subfields had superior discrimination from unaffected eyes when compared with the superior inner and superior outer ETDRS subfields. More detailed and refined detection and quantification of changes in retinal layers/substructures beyond the segmentation of three layers performed here (ie, ILM, RPE and OPL) could further improve the results reported in this paper. For example, detecting and quantifying the changes and disruptions in the EZ layer visible in SD-OCT could reveal parameters relevant to toxicity. Other studies have supported this avenue of analysis with quantitative OCT segmentation studies providing evidence of ONL thickness and HFL-EZ thickness measures showing promise as useful biomarkers of toxicity.²²

The study was conducted using a prospective study design, resulting in a well characterised data set, where the screening tests performed for the determination of toxicity comprised not just some but all of those listed by the AAO in 2016,¹ including automated perimetry, SD-OCT and mfERG (with objective interpretation).^{16 20} The unaffected group consisted of participants with the same underlying systemic diseases, which makes the comparisons in this study more balanced and meaningful (eg, controlled for any systemic disease effects). Potential limitations to the study include its single-site nature and cross-sectional design. The study population included only a small number of patients of Asian ethnicity (four participants in the unaffected cohort and four in the affected cohort) who are more likely to have pericentral changes.¹ Inclusion of data from additional sites is needed to further validate the metrics and demonstrate generalisability. Additional study of the reliability of OCT automated measures between different OCT scans, operators and sessions would also be useful additions in future investigations. While much can be gained from cross-sectional analysis, with severity scales representing a parallel to disease progression, the critical dataset that could demonstrate the utility of these algorithms would be in the setting of a prospective longitudinal study, where the development of toxicity and an identification of a crossing point, from unaffected to affected, could be observed.

OCT metric analyses in these eyes demonstrated that both ORT and MI could provide discrimination in most subfields. However, due to the limited data set used in this study, further exploration with additional data could validate these findings in different populations. While RF analysis also confirmed the utility of both these metrics in distinguishing affected and unaffected patients, further studies involving larger number of patients are required to assess the interplay between these features when detecting early changes induced by toxicity. Such studies could validate an automatic, quantitative, OCT-based decision support tool to assist clinicians in accurately identifying patients with toxicity. This could consist of the development of a platform for estimating the risk of toxicity after SD-OCT acquisition to be implemented in multiple clinical settings. As the field moves increasingly toward reliance on SD-OCT scans for screening, these metrics may provide clinicians with additional information as they monitor patients taking HCQ for the development of retinopathy, which may or may not be evident on qualitative inspection of an OCT scan. Arming clinicians with tools to aid in the discrimination of OCT imaging would enable interpretations that save vision (ie, are sensitive) and enable the drug to be continued safely (ie, are specific). Moreover, these metrics could be integrated into the development of advanced clinical decision support systems that could combine several metrics along with patient demographic data to assess the overall risk of disease progression. Additional study of the reliability of OCT automated measures between different OCT scans, operators and sessions would also be useful additions in future investigations.

As an analogy, in the field of glaucoma, relevant OCT metrics have gradually been developed over the last 15 years and subsequently integrated into analytic software that aids the physician in assessing optic nerve and ganglion cell layer anatomy.²³ Following many studies to establish normative values and progression data, these OCT metrics are now highly valuable and widely used by clinicians screening patients for glaucoma.^{24 25} Similarly, perhaps following further studies using large, longitudinal datasets, it may be possible to develop and validate OCT-based metrics to aid in screening for HCQ toxicity, which currently remains an unmet need with great public health importance.

Data availability statement

Data are available on reasonable request. All data relevant to the study are included in the article or uploaded as online supplemental information. Deidentified participant data that support the findings of this study are available from the corresponding author, CC at cukrasc@nei.nih.gov, on reasonable request and abiding by institutional sharing guidelines.