CAPTCHA as a Visual Performance Metric in Active Macular Disease

Purpose CAPTCHA (completely automated public turing test to tell computers and humans apart) was designed as a spam prevention test. In patients with visual impairment, completion of this task has been assumed to be difficult; but to date, no study has proven this to be true. As visual function is not well measured by Snellen visual acuity (VA) alone, we theorized that CAPTCHA performance may provide additional information on macular disease-related visual dysfunction. Methods This was designed as a pilot study. Active disease was defined as the presence of either intraretinal fluid (IRF) or subretinal fluid (SRF) on spectral-domain optical coherence tomography. CAPTCHA performance was tested using 10 prompts. In addition, near and distance VA, contrast sensitivity, and reading speed were measured. Visual acuity matched pseudophakic patients were used as controls. Primary outcome measures were average edit distance and percent of correct responses. Results 70 patients were recruited: 33 with active macular disease and 37 control subjects. Contrast sensitivity was found to be significantly different in both the IRF (p < 0.01) and SRF groups (p < 0.01). No significant difference was found comparing the odds ratio of average edit distance of active disease (IRF, SRF) vs. control (OR 1.09 (0.62, 1.90), 1.10 (0.58, 2.05), p=0.77,  0.77) or percent correct responses of active disease vs. control (OR 0.98 (0.96, 1.01), 1.09 (0.58, 2.05), p=0.22, 0.51) in CAPTCHA testing. The goodness of fit using logistic regression analysis for the dependent variables of either IRF or SRF did not improve accounting for average edit distance (p=0.49,  p=0.27) or percent correct (p=0.89,  p=0.61). Conclusions Distance VA and contrast sensitivity are positively correlated with the presence of IRF and SRF in active macular disease. CAPTCHA performance did not appear to be a significant predictor of either IRF or SRF in our pilot study.


Introduction
Computer automated public turing test to tell computers and humans apart (CAPTCHA) has gained popularity in recent years as an electronic visual challenge-response test of human authenticity (Figure 1). e test involves a user correctly entering all alphanumeric symbols displayed in a text prompt. Although CAPTCHA testing is useful in thwarting potential hackers and spamming, passing it can prove challenging to those with visual impairment [1,2]. Responders with diseases affecting central vision, in particular the macula, may have even more difficulty completing these tasks. To this end, quantifying patients' level of performance on CAPTCHA testing may provide a noninvasive, sensitive, and easy way to perform method of potentially monitoring active macular disease.
To evaluate macular disease, CAPTCHA performance may provide an advantage over other visual acuity metrics methods in assessing central visual acuity. Traditionally, Snellen visual acuity is the standard metric for assessing vision and treatment outcomes [3][4][5][6][7][8][9]. Nonetheless, Snellen testing often fails to address aspects of important visual impairment to patients for daily functioning and quality of life. Elliott et al. reported on a series of thirty-three patients with a history of visually significant cataract in at least one eye. e authors measured each patient's visual acuity, brightness acuity, and binocular contrast sensitivity and correlated these findings to a questionnaire of vision-related quality of life [10]. Binocular contrast sensitivity correlated more strongly to visual quality of life than Snellen visual acuity alone. Similarly, Hazel et al. found that low-contrast visual acuity, contrast sensitivity, and binocular reading speed had the strongest correlation with visual quality of life in 28 patients with active macular disease from either presumed ocular histoplasmosis syndrome (POHS) or exudative macular degeneration [11].
Snellen visual acuity also fails to address visual symptoms in individuals with noncenter involving macular disease. Quantifying a patient's CAPTCHA performance may provide a functional endpoint that is noninvasive, inexpensive, and easy to perform compared to structural makers of disease activity such as optical coherence tomography (OCT) and optical coherence tomography angiogram (OCTA). To our knowledge, CAPTCHA performance has not been studied in patients with visual disability, and this is the first study to examine CAPTCHA performance in patients with active macular disease [12][13][14][15][16][17][18][19][20]. Our study hypothesized that patients with macular disease, defined as active diabetic macular edema (DME) or exudative age-related macular degeneration (AMD), would perform poorly on a series of CAPTCHA prompts compared to visually matched pseudophakic controls.

Materials and Methods
e study was a prospective cohort study. Patients were recruited from the University of Washington Eye Institute, Seattle, WA. Institutional review board approval was obtained, and the study followed the World Medical Association Declaration of Helsinki Ethical Principles for medical research involving human subjects. Inclusion criteria for the macular disease cohort were defined as having ICD-10 code diagnoses of diabetic macular edema (E11.311) or exudative age-related macular degeneration (H35.32), OCT confirmed intraretinal fluid (IRF) or subretinal fluid (SRF), and English as the primary spoken language. Exclusion criteria for the macular disease cohort were defined as patients with concomitant pathology affecting the central visual axis (e.g., visually significant cataract, central corneal scar, and amblyopia), best near corrected visual acuity worse than 20/50, and the presence of both IRF and SRF. Control patients were pseudophakic, spoke English as their primary language, and had a bestcorrected visual acuity of at least 20/20 with no other diseases affecting the central visual axis. Only one eye from each patient was tested. If both eyes met inclusion and exclusion criteria, a coin flip was used to determine the testing eye.
All patients were asked to complete a series of visual tasks. Control patients were blurred to a predetermined visual acuity (20/20-20/50) using positive power spherical lenses in an attempt to create a 1 : 1 match of visual acuity to the macular disease cohort. Distance acuity was tested using a standard Snellen chart while near distance was tested using a Rosenbaum near point acuity card at 14 inches. CAPTCHA testing was performed at the same near distance, 14 inches, as the Rosenbaum near point test by using an iPAD Air (Model A1566, Apple Inc. Cupertino, CA) tablet. A series of 10 CAPTCHA images were presented to each patient in a randomized order. Typed responses were recorded in a separate database. Contrast sensitivity (logcontrast) was tested using a Pelli-Robson chart. Reading speed in seconds was tested at near distance using a 156word paragraph. e primary outcome measure was average edit distance. In computational science, edit distance is defined as the minimum number of operations needed to transform one string of letters into another. e average edit distance of each subject's responses to the ten randomized CAPTCHA prompts was calculated. e secondary outcome measure was percent correct. For example, if a patient answered all of one CAPTCHA prompt exactly as displayed, this was marked as correct, and any deviation from a prompt was marked incorrect.
us, if a patient answered five CAPTCHA prompts correctly out of ten, the value of this outcome variable would be 50%.

Statistical Methods.
All statistical analyses were completed using SPSS version 24 (IBM corporation, New York, United States) and independently verified by a biostatistician. Logistic regression analysis was used to determine the goodness of fit (Nagelkerke R 2 ) of a base model consisting of the dependent variable (IRF or SRF) and covariates (visual acuity, reading speed, and contrast sensitivity). Average edit distance and percent correct were independently added to this base model, and chi-square analysis was used to determine significance in model fit.

Results
Seventy patients were recruited for this study (Table 1).
irty-three patients had active macular disease comprising either subretinal fluid or intraretinal fluid.
ere was no statistical difference in age (p � 0.552), gender (p � 0.824), or race (p � 0.214) between the active disease and control groups. When near visual acuity was stratified into 5 groups (20/20, 20/25, 20/30, 20/40, and 20/50), no significant difference was found between active disease and control cohorts. e active disease cohort was further stratified into patients with IRF (n � 20) and SRF (n � 13). On univariate analysis, there was no significant difference in average edit distance between either the IRF group or SRF group vs. control (Table 2). Likewise, no significant difference was found in percent correct between either the IRF group or SRF group vs. control. As demonstrated in Figure 2, no distinct trend was seen in comparing the IRF group or SRF group vs. control for average edit distance, percent correct, or reading speed. However, contrast sensitivity performance was noted to significantly differ between both IRF vs. control (p � 0.016) and SRF vs. control (p � 0.024).
Logistic regression analysis is shown for both IRF and SRF cohorts (Table 3). e base model (A or B) assumes either IRF (A) or SRF (B) as the dependent variable with covariates of visual acuity, reading speed, and contrast sensitivity. e −2 log likelihood, which can be interpreted as a pseudo R 2 for this model, was calculated. For both the IRF and SRF base models, covariates of average edit distance and percent correct were added to determine if this improved the goodness of fit of the model. In neither IRF (x 2 � 0.07, p � 0.79) nor SRF (x 2 � 0.15, p � 0.70), did the addition of average edit distance improve the goodness of fit of the model. A similar result was found when the secondary outcome variable of percent correct was added to the model: IRF (x 2 � 0.68, p � 0.410) and SRF (x 2 � 0.02, p � 0.88).

Discussion
is was a prospective cohort study to determine whether macular intraretinal or subretinal fluid correlated with CAPTCHA performance. e "goodness of fit" of the logistic regression model predicting either IRF or SRF did not improve when average edit distance was added to the base model. Similarly, no improvement was seen in comparing No statistically significant difference was found between active disease and control cohort for any baseline characteristics. BCVA: best-corrected visual acuity, DME: diabetic macular edema, and AMD: age-related macular degeneration.  [21]. Both the IRF and SRF cohorts performed significantly worse on contrast testing compared to pseudophakic visual acuity-matched controls. CAPTCHA images contain a series of variably separated optotypes that create a challenging visual prompt to the examiner. Although CAPTCHA images do not fit a standardized visual testing method, it mostly resembles ascertaining a subject's minimum legible acuity. Both Snellen distance and Rosenbaum near point testing also assess a subject's minimum legible acuity by differentiating increasingly smaller optotypes to determine a final visual acuity [22]. Unlike these tests, however, CAPTCHA does not employ strong serifs in its font, making its text more difficult to discern. Furthermore, given the variation in CAPTCHA letter size within each image, it is difficult to determine the equivalent visual acuity with Snellen testing. However, the majority of letters in each test image were between 10-and 14-point font or 20/70 and 20/100 Snellen equivalent.
In addition, all CAPTCHA prompts in our study employed high contrast, black print on white background. A tablet computer was chosen to administer the CAPTCHA prompts to most realistically simulate real-life testing conditions and facilitate a standardized near acuity. However, the increased background luminance during testing has been shown to improve a subject's visual acuity and small letter contrast sensitivity [23,24]. us, the heightened background luminance of the tablet may have aided the subject compared to the standardized photopic room testing   conditions (90 lux) of the Pelli-Robson examination. In addition, the enhanced contrast of CAPTCHA images may be above the threshold of how much intraretinal or subretinal pathologies affects one's visual function, which could explain why CAPTCHA performance was not significantly different in patients with active disease versus control subjects. e exact pathophysiology of decreased contrast sensitivity has not been fully elucidated. Poor performance on contrast sensitivity testing is seen in a variety of diseases including glaucoma, optic neuritis, and macular degeneration [21]. Arend et al. reported a series of 20 patients with diabetes mellitus without clinically significant macular edema with better than 20/25 vision. e authors found a significant correlation between foveal avascular zone area and contrast sensitivity, suggesting parafoveal capillary dropout may have contributed to a decrease in contrast sensitivity [25]. García et al. performed a retrospective study on 62 multiple sclerosis patients and found a significant correlation with optic nerve disease activity and contrast sensitivity performance. Our study adds to the existing literature demonstrating poor contrast delineation in patients with intraretinal fluid or subretinal fluid [26,27]. e strengths of this study are its prospective nature, which allowed for well-defined inclusion and exclusion criteria. In addition, OCT images obtained at the time of testing allowed for determination of disease activity in real time. Limitations include the lack of quantification of center vs. noncenter, involving fluid on OCT imaging and the absence of OCT images from pseudophakic controls.
Additionally, the technique of visual blurring using positive convex lenses has been described in the literature to assess its effect on the completion of visual tasks and the comprehension of visual cues [28,29]. Although visual blurring effectively matched the control population's visual acuity to the disease cohort, this technique is only meant to simulate retinal pathology and does not entirely mimic the pathology seen in DME and AMD.
In summary, visual function tests other than Snellen visual acuity are important in evaluating ocular diseases. In particular, contrast sensitivity should be considered in the clinical evaluation of active macular disease. Although visually impaired patients can struggle with CAPTCHA task completion, our study results show that neither intraretinal fluid nor subretinal fluid correlate with worsened CAPTCHA performance. We would expect, therefore, that using CAPTCHA to prevent hacking attempts would not exclude patients affected with active macular disease if their visual acuity allowed them to perform the test adequately. Further studies are needed to determine if CAPTCHA performance correlates to other structural markers of ocular pathology.

Conclusions
CAPTCHA performance is not a significant predictor of intraretinal fluid or subretinal fluid in patients with active macular disease. Ancillary tests of visual function, particularly contrast sensitivity, correlate to macular disease.
Further studies are needed to determine whether CAPTCHA performance is associated with other features of macular diseases.

Data Availability
e data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare that there are no conflicts of interest regarding the publication of this paper. Leona Ding is a biostatistician at the University of Washington.