A Pattern-Based OCT Metric for Glaucoma Detection

Abstract

Purpose: To develop and test a novel optical coherence tomography (OCT) metric for the detection of glaucoma based on a logistic regression model (LRM) and known patterns of glaucomatous damage.

Methods: The six variables of the LRM were based on characteristic patterns of damage seen on the OCT thickness maps of the ganglion cell layer plus inner plexiform layer (GCL+) and retinal nerve fiber layer (RNFL). Two cohorts were used to develop the LRM. The healthy cohort consisted of 400 individuals randomly selected from a real-world reference database (RW-RDB) of OCT widefield scans from 4932 eyes/individuals obtained from 10 optometry practices. The glaucoma cohort consisted of 207 individuals from the same 10 practices but with OCT reports with evidence of optic neuropathy consistent with glaucoma (ON-G). Specificity was assessed with 396 eyes/individuals from a commercial RDB. Sensitivity was assessed with individuals with ON-G from different optometry practices.

Results: For the new LRM metric, the partial area under the reciever operating characteristic curve (AUROC) for specificity >90% was 0.92, and the sensitivity at 95% specificity was 88.8%. These values were significantly greater than those of a previously reported LRM metric (0.82 and 78.1%, respectively) and two common OCT thickness metrics: global circumpapillary RNFL (0.77 and 57.5%, respectively), and global GCL+IPL (0.72 and 47.6%, respectively).

Conclusions: The new metric outperformed other OCT metrics for detecting glaucomatous damage.

Translational Relevance: The new metric has the potential to improve the accuracy of referrals from primary care to specialist care via risk scores and calculators, as well as glaucoma definitions for clinical trials. The individual variables of this model may also aid clinical diagnosis.

Introduction

Although it is generally agreed that optical coherence tomography (OCT) has become an essential structural test for glaucoma, there is disagreement on how best to interpret OCT results. Clinical trials, risk calculators, and clinicians often depend upon a strictly quantitative approach, typically based on conventional metrics readily available from the commercial OCT output. For example, in clinical trials, the most common OCT metric is global circumpapillary retinal nerve fiber layer (g-cpRNFL) thickness. 

We and others have provided evidence that g-cpRNFL, as well as other commonly used quantitative methods, miss serious glaucomatous damage that can be clearly seen on OCT RNFL and ganglion cell layer plus inner plexiform layer (GCL+) maps, as well as on visual fields.^1–12 Further, the damage missed by the quantitative approaches typically involves the central macula. On the other hand, this damage can be detected with a semiquantitative approach that includes a subjective evaluation of quantitative OCT information.^4,12 However, although a semiquantitative approach involving subjective decisions may be the best approach for clinical decisions, a strictly automated approach that depends on a single metric is required for other purposes such as clinical trials and for improving referrals from primary care to specialist care via risk scores and calculators. 

Recently, Fukai et al.¹³ used a multivariable logistic regression model (LRM) with backward stepwise selection, which reduced the number of OCT variables in a model down to six. Their approach has two important advantages over a single metric approach. First, it takes into account damage to the macular region, the region essential for everyday function, as well as cpRNFL measures. In fact, four of their six variables include measurements of the GCL+ of the macula, whereas only two of the variables include the cpRNFL. Second, the LRM produces a single continuous metric or score, from 0 to 100, making it viable for use in clinical trials and glaucoma risk scores and calculators. 

In the present study, we sought to improve upon the Fukai et al.¹³ approach in two ways. First, the six variables in their LRM were chosen from 312 candidates of various measures of cpRNFL and GCL+ thickness. We hypothesized that the performance of a LRM could be improved by choosing RNFL and GCL+ variables based on empirically validated models that describe patterns characteristic of glaucomatous loss.^{2,6,7,14–30} 

A second difference involves the choice of test sets for evaluating sensitivity. Fukai et al.¹³ found that their best performing LRM had 91% specificity and 85% sensitivity. In general, it is difficult to know what 85% sensitivity means, as sensitivity depends on the level of glaucomatous damage in the test sample. In this study, we tested the LRMs with two cohorts of data chosen to assure a wide range of glaucomatous damage. One cohort had clear damage and a diagnosis of glaucoma, whereas a second included eyes with a range of OCT patterns characteristic of glaucoma, including eyes with subtle damage. 

Thus, our primary goal was to develop an LRM based on empirically validated variables.^{2,6,7,14–30} Our secondary purpose was to compare the performance of our LRM against the Fukai et al. (F-N) LRM and to the two commonly used OCT global metrics for aiding in the detection of glaucoma in the clinic and in clinical trials. The implications for risk scores and calculators and endpoints for clinical trials, as well as for improving reports used in the clinic, are considered in the Discussion section. 

Methods

All participants had a 12 × 9-mm widefield OCT scan that included the optic nerve head and macula (Maestro2; Topcon Healthcare, Tokyo, Japan). Segmentation, disc, and fovea centering were utilized as provided by the commercially available software, without modifications. The commercial Hood Report was generated for all eyes using the Maestro2 software. The data included in the optic neuropathy consistent with glaucoma (ON-G) pattern cohort, 207 OCT-G cohort, and real-world reference database (RW-RDB) cohort, described below, came from optometry practices and were part of a retrospective study approved by the Advarra Institutional Review Board (IRB). Data in the clear ON-G cohort described below were derived from a retrospective chart review approved by the IRB of Illinois College of Optometry. Data receipt and analysis of all the data by the Columbia team were approved by the IRB of Columbia University. 

Participants

The study included five cohorts as described below. For the development of our LRM and the determination of percentile cutoffs, we used the following two cohorts: 

• • Real-world reference database—The RW-RDB database included 4932 OCT healthy eyes from 4932 individuals (Fig. 1, green rectangle).³¹ To obtain this RW-RDB with 4932 individuals, 6804 individuals were randomly sampled from over 23,000 individuals scanned in 10 practices that primarily focused on refractive care, as previously described.³¹ Using a reading center approach, the 4932 individuals in the RW-RDB were judged to have OCT Hood Reports without pathology and/or scan artifacts.³¹ Because of its size, the RW-RDB allows for more precise percentile cutoffs compared to existing commercial RDB. A comparison of the RW-RDB to the RDB in the commercial Maestro2 can be found in Hood et al.³¹ A random sample of 400 individuals (Fig. 1, dashed green rectangle) was used in the development of the LRM equation described below.
• • 207 OCT-G cohort—These 207 eyes (Fig. 1, bottom, dashed red rectangle) were from the same 10 optometry practices as the RW-RDB. They were part of a cohort of 1872 eyes/individuals excluded from the RW-RDB based on scan quality or pathology as judged by the reading center approach previously described.³¹ These 207 eyes had good-quality scans and a range of damage consistent with ON-G based strictly on the OCT. The means ± SD for the g-cpRNFL and g-GCL+IPL thicknesses were 78.91 ± 16.19 µm and 58.98 ± 8.61 µm, respectively. These 207 individuals (Fig. 1, dashed red rectangle) were used in the development of the LRM equation described below.

For the estimation of performance measures (i.e., sensitivity and specificity), we used the following three cohorts: 

• • Maestro2 396 RDB (healthy test set)—This cohort, which was used to test the specificity of the evaluated OCT metrics, consisted of 396 of the 398 eyes in the currently available commercial test set.^32,33 Two eyes from the 398 were excluded: one eye had optic neuropathy consistent with glaucoma, and the other had a pattern of damage characteristic of post-geniculate damage. The means ± SD for the 396 g-cpRNFL and g-GCL+IPL thicknesses were 104.7 ± 11.9 µm and 71.3 ± 7.0 µm, respectively.
• • Clear ON-G cohort—This is one of the two cohorts used for testing the sensitivity of the OCT metrics. It consisted of 52 eyes from 52 individuals tested at the Illinois College of Optometry, as described in La Bruna et al.³⁴ These eyes all had glaucoma according to a glaucoma expert (author MC). They were also judged to be Clearly ON-G based solely on the Hood Report by two OCT experts (authors DCH and MT). In addition, for inclusion, an eye had to have a 24-2 visual field with an abnormal pattern standard deviation and/or abnormal glaucoma hemifield test. The means ± SD for 24-2 mean deviation (MD) were −6.4 ± 6.3 dB, and the means ± SD for g-cpRNFL and g-GCL+IPL thicknesses were 72.2 ± 15.4 µm and 54.7 ± 8.4 µm, respectively.
• • ON-G pattern cohort—This is the second cohort used to assess the sensitivity of the alternative metrics. In this case, we sought to include eyes with a full range of glaucomatous damage, including eyes with subtle ON-G damage. This cohort (Fig. 1, solid red rectangle) consisted of 183 eyes from 183 individuals from three optometry practices that were not part of the 10 practices used for the RW-RDB and 207 OCT-G cohorts. Two OCT experts judged the ON-G pattern based only on the Hood Report. For inclusion, an individual had to be 60 years of age or older and have at least one eye that showed evidence of an arcuate-like pattern of thinning on the RNFL, GCL+IPL, and/or RNFL en face maps. If both eyes of an individual met the criteria, the eye with the most subtle (i.e., least extensive damage and/or deep patterns) was chosen for the ON-G pattern cohort. Although this cohort consisted of eyes with a range of damage consistent with ON-G, it also included any eye with an arcuate thinning apparent even if there were no indications on the RNFL and/or GCL+IPL probability maps. The means ± SD for the g-cpRNFL and g-GCL+IPL thicknesses were 81.0 ± 16.5 µm and 61.3 ± 7.2 µm.

Logistic Regression Models

The F-N LRM is model 3 described in Fukai et al.¹³ For this model, we used the coefficients provided in Fukai et al.¹³ For the Hood–Tsamis (H-T) LRM, the six variables were based on known patterns of damage to the OCT macular GCL+IPL and cpRNFL.^{2,6,7,14–30} The values of these variables were calculated from information available from the commercial Maestro2 instrument. Figure 2 provides a schematic representation, as well as the equation, for each variable. The first three variables, which involve the thick central region of the GCL+IPL around the fovea, take advantage of the empirical observation that damage to this region typically starts in the temporal sectors (TS and TI) and proceeds to the nasal sectors (NS and NI), as indicated by the red arrows in Figure 2.¹⁴ The first two variables are designed to detect early damage to the macular temporal inferior (TI+I) and temporal superior (TS+S) regions, whereas the third variable (TS+TI) detects more advanced glaucoma affecting both TI and TS regions. For the cpRNFL variables, we made use of the finding that the superior vulnerability zone (SVZ) and inferior vulnerability zone (IVZ) are the most affected regions.^{2,16,29,35–37} The clock hour cpRNFL thickness measures provided by the commercial software were used to calculate the cp-SVZ and cp-IVZ variables as described in the legend to Figure 2. The last variable, G+G, is the sum of the global (g-) thicknesses of the GCL+IPL and cpRNFL and is chosen to detect overall thinning of the cpRNFL and/or GCL+IPL layers. 

The weights of the six variables for the H-T LRM were determined using the 207 OCT-G eyes and 400 randomly selected OCT healthy eyes from the RW-RDB (Fig. 1, green dashed rectangle). All six variables described above and shown in Figure 2 were significant with P < 0.001. The final equation of the H-T LRM is as follows:  

\begin{eqnarray*} && Y = 354.87x\left( {\frac{{TI + I}}{{TS + S}}} \right) + 350.09x\left( {\frac{{TS + S}}{{TI + I}}} \right)\\ && - 4.11x\left( {\frac{{TS + TI}}{{NS + NI}}} \right) + 13.05x\left( {\frac{{IVZ}}{{SVZ}}} \right)\\ && + 14.38x\left( {\frac{{SVZ}}{{IVZ}}} \right) - \ 0.09x\left( {G + Gmac} \right) - 716.35\end{eqnarray*}

 

Figure 1.

 

A flow diagram illustrating the relationship of the four cohorts used in this study from 13 optometry practices. The RW-RDB cohort with 4932 individuals (solid green rectangle) came from 10 practices. A total of 6804 individuals were sampled from over 23,000 individuals to obtain an approximately flat age distribution of about 5000 individuals with OCTs free of significant artifacts and pathology as previously described.³¹ The individuals in the OCT-G cohort with 207 individuals (dashed red rectangle) came from the individuals rejected from the 6804 individuals. They had OCT reports with optic neuropathy consistent with glaucoma as judged using a reading center method. This cohort was used to develop the weights of the H-T index of the LRM, along with a random sample of 400 individuals (green dashed rectangle) from the RW-RDB cohort. The ON-G pattern cohort with 183 individuals (solid red rectangle) was obtained from three different optometry practices. They had good-quality scans and arcuate patterns of thinning on the thickness and/or en face maps. They were used to test the specificity of the metrics. Datasets used for development of the LRM have dashed outlines, and datasets used for performance testing have solid outlines.

Figure 2.

 

A schematic representation for each variable in the H-T LRM (A) Representation of the TI+I variable designed to detect early damage to the inferior region of the macula. The GCL+IPL thicknesses of the red highlighted sectors (TI+I) are added and compared to the sum of the thicknesses of the gray sectors (TS+S). This metric takes advantage of the empirical observation that damage to the inferior GCL+IPL region typically starts in the temporal sector (TI+I) and proceeds to the NI sector, as indicated by the red arrow.¹⁴ (B) Representation of the TS+S variable designed to detect early damage to the superior region of the macula. The GCL+IPL thicknesses of the red highlighted sectors (TS+S) are added and compared to the sum of the thicknesses of the gray sectors TI+I. This metric takes advantage of the empirical observation that damage to the superior GCL+IPL region typically starts in the temporal sectors (TS+S) and proceeds to the NS, as indicated by the red arrow.¹⁴ (C) Representation of the TS+TI variable designed to detect more advanced glaucoma affecting both the superior and inferior regions, which typically affects the temporal regions (TS and TI) more than the nasal regions (NS and NI). (D) The circumpapillary (cp) IVZ variable is designed to detect damage occurring first to the IVZ. This variable compares the cpRNFL thickness corresponding to the IVZ (clock hours 6 and 7, OD) and to the SVZ thickness (clock hours 11 and 12, OD). (E) The cp-SVZ variable is designed to detect damage occurring first to the SVZ. This variable compares the cpRNFL thickness corresponding to the two cpRNFL vulnerability zones. (F) Representation of the G+G variable designed to detect global or diffuse damage, calculated as the sum of the global GCL+IPL and global cpRNFL thicknesses.

A score, ranging from 0 to 100, was calculated based on the following equation:  

\begin{eqnarray*}\ Score = \ \frac{{{\rm{exp}}\left( Y \right)}}{{1 + {\rm{exp}}\left( Y \right)}} \times 100\end{eqnarray*}

Higher scores indicate a greater likelihood of glaucoma, not a probability of glaucoma. In other words, this score should not be taken as a probability of having or developing glaucoma (see Discussion below). 

Statistical Analysis

The performance of the two LRMs and two commonly used OCT metrics, g-cpRNFL thickness and g-GCL+IPL thickness, were evaluated with two approaches. First, based on receiver operating characteristic (ROC) analysis, the four areas under the ROC (AUROC) curve were compared against each other using DeLong's test for two correlated ROC curves. In addition, partial AUROC curves restricted to specificities >90% were estimated using the McClish correction.³⁸ Second, based on the values for the 95th percentile (pct) cutoff of the RW-RDB, the sensitivities of the LRMs and the two OCT metrics were compared using McNemar's test. 

The Six Variables of the H-T LRM

As a possible aid in the clinical interpretation of the H-T score, for each eye we also calculated the percentile of each of the six variables. These percentiles were based on the RW-RDB. For example, if a variable had a percentile of 2%, then that meant that ≤2% of the RW-RDB eyes had a lower value for that variable. For our purposes, we were only interested in the values of the six variables at the lower end of the RW-RDB distribution (see Discussion). 

Results

A Comparison of the Two Logistic Regression Models

The AUROC curve for the H-T LRM, 0.972, was significantly greater than the AUROC curve for the F-N LRM, 0.941 (P < 0.001, DeLong's test) (see Table 1). Given that these metrics are designed for glaucoma detection, sensitivities at low specificities are largely irrelevant. Therefore, we also examined partial AUROC (pAUROC) curves restricted to specificities above 90%. The H-T LRM again had the largest pAUROC curve, 0.92, as compared to 0.86 for the F-N LRM. Further, the sensitivity on the ROC at 95% specificity was found to be 88.8% for the H-T LRM, approximately 10% higher than for the F-N LRM (Table 1). 

Table 1.

 

AUROC, pAUROC, and Sensitivity at a 95% Specificity for the Two Models, H-T and F-T, and the Two Commonly Used Global Metrics, g-cpRNFL and g-GCL+IPL

 

To further test the hypothesis that the H-T LRM has better sensitivity than the F-N LRM, we set the cutoff values of each LRM to have a specificity of 95% based on the RW-RDB (see the footnote in Table 2 for these cutoff values). With the Maestro2 RDB as the test set of healthy eyes, the actual specificities of the two LRMs were close (H-T, 96.5%; F-N, 95.7%). Although the H-T LRM had fewer false positives (14 vs. 17) (Table 2), the difference was not significant (P = 0.65, two-tailed McNemar test). However, as predicted, the sensitivity of the H-T LRM was greater than the sensitivity of the F-N LRM, as detailed below, for the clear ON-G and the ON-G pattern cohorts. 

Table 2.

 

Comparison of the Specificity and Sensitivity of Four Metrics Based on the 95th Percentile Cutoff of the RW-RDB

 

Testing Sensitivity for Eyes With Clear ON-G

For the 52 clear ON-G eyes, the H-T LRM had better sensitivity (0 vs. 5 false negatives) when applying the 95th percentile cutoff (see Table 2). The difference was significant (P < 0.05, McNemar test). The five eyes missed by the F-N LRM showed clear loss of thickness characteristic of glaucoma. Figure 3 shows the Hood Report for all five false negatives. In two cases (Figs. 3A, 3E), there was extensive damage that included the entire donut region of the GCL+IPL, whereas in the other three there was arcuate damage with different degrees of damage to the central thick donut-shaped GCL+IPL region. Note that, whereas the example in Figure 3D had a F-N score at the 5.3 pct of the RW-RDB, close to the 95th percentile cutoffs, the scores for the other three panels fell at the 21.5 pct (Fig. 3A), 9.3 pct (Fig. 3B), 21.0 pct (Fig. 3C), and 9.0 pct (Fig. 3E) of the RW-RDB. 

Figure 3.

 

(A–E) The Hood Reports for five eyes that were detected by the H-T LRM but were missed by the F-N LRM in the 52 clear ON-G cohort. Note that in panels A and E there is extensive damage that includes the entire thick donut region of the GCL+IPL, whereas panels B and C show arcuate defects that include damage to the central donut. For the eye in panel C, the arcuate damage is largely outside the central donut.

Testing Sensitivity for Eyes With ON-G Patterns

For the 183 ON-G pattern eyes (Fig. 1, red rectangle), the sensitivity of the H-T LRM was 84.2% (29 false negatives) when applying a specificity of 95% (Table 2). This sensitivity is relatively high, especially considering that this cohort included eyes with subtle arcuate defects as indicated in Figures 4A and 4B, as well as eyes with relatively narrow arcuates as shown in Figure 4C. In any case, the sensitivity for the H-T LRM was significantly higher than for the F-N LRM (29 vs. 47 false negatives; P < 0.01, McNemar test) (Table 2). Note that the H-T scores for the three eyes in Figures 4A to 4C fell at the 7.4 pct (Fig. 4A), 10.3 pct (Fig. 4B), and 15.5 pct (Fig. 4C) of the RW-RDB. 

Figure 4.

 

(A–C) Hood Reports for eyes in the 183 ON-G pattern cohort that were missed by the H-T LRM. In some cases, such as A and B, these eyes had subtle damage in the form of narrow and shallow arcuates, although in some the arcuate was deep (C). (D) Eye detected by the H-T LRM but not the g-cpRNFL.

Comparison of the H-T LRM to Two Global Metrics

The H-T had better ROC measures and sensitivity than either of the two common single metrics. The AUROC and pAUROC of the H-T LRM were significantly greater than those measures for the g-cpRNFL and g-GCL+IPL (P < 0.001, DeLong's test) (Table 1). A direct comparison of the ROC curves for the four indices/metrics is provided in Figure 5. Both global metrics had more false negatives (i.e., they were less sensitive) than the H-T LRM for both ON-G cohorts (Table 2). The differences were significant (P < 0.01, McNemar's test). Of note, for the clear ON-G cohort, the g-cpRNFL, the most common OCT metric used in clinical trials, missed eight of the 52 eyes, but the H-T LRM missed none. At least seven of these eight eyes had clear damage to the central macular donut. Figure 4D provides an example with a g-cpRNFL value at the 10th percentile of the RW-RDB. Supplementary Material SA shows the Hood Reports for the remaining seven eyes. 

Figure 5.

 

ROC curves for the two LRM models and the two global thickness metrics. Sensitivity was based on the clear ON-G and ON-G pattern cohorts and specificity on the 396 eyes from the commercial RDB.

Discussion

The H-T LRM, which is based on known patterns of glaucomatous loss, had better performance (e.g., greater AUROC and pAUROC) than the F-N LRM (Table 1). Further, the H-T LRM had significantly better sensitivity at a specificity of about 95% than two common OCT g-metrics used in clinical trials (Table 2). These results have implications for improving OCT metrics for risk scores/calculators and clinical trials, as well as for improving reports to be used as a diagnostic aid in the clinic. 

H-T Score as a Standalone Metric

There are a few scores or indices that are based only on OCT thickness measurements and logistic regression modeling^{13,21,39–41}; the F-N risk score developed by Fukai et al.¹³ is the most similar to the one presented in this study. The main difference between the H-T score and their score/index is the method used for choosing the variables of the LRM. Whereas the six variables for the H-T score were based on characteristic patterns of glaucomatous damage seen on OCT, the six variables of the F-N LRM were chosen based on a statistical method to reduce 312 possible variables to six. However, like the F-N LRM, the six elements in the H-T LRM are based on GCL+IPL and cpRNFL thicknesses. Further, in both cases, four of the variables depend at least in part on macular measurements. Of note, Fukai et al.¹³ developed their risk (F-N) score for “population-based glaucoma mass screenings,” whereas other indices have been designed to detect early glaucoma in a routine clinical practice.^21,39 Below, we consider the use of the H-T score for each of these purposes. 

H-T Score as a Risk Score for Screening

As a single metric for screening, compared to the F-N LRM, the H-T LRM had better sensitivity for a comparable specificity (Table 2). Further, the AUROC of the H-T LRM was significantly greater (Table 1; Fig. 5); however, the comparison of sensitivity and specificity and/or AUROC tells only part of the story. Two methods can have similar specificity and sensitivity but differ in the number of eyes with more extensive damage, for example, of the central retinal area missed. Thus, the false negatives should be analyzed to assess the extent of damage present in the eyes missed by a risk score. This was the reason we included a clear ON-G cohort. The F-N LRM missed five of the 52 clear ON-G eyes (10%) at a 95% specificity level. These five included two eyes with extensive widespread loss of tissue in the macular region (Figs. 3A, 3E). Similarly, the two common global metrics also missed between eight and 12 of the 52 clear ON-G cohort at the same specificity (Table 2). 

Further, a specificity of 95% may be too low for some screening purposes. However, we have recently identified anatomical factors, such as fovea-to-disc distance and location of peak thickness of cpRNFL profile, that affect specificity and are exploring ways to use this information to de facto improve specificity for a given sensitivity.⁴² Finally, a word of warning about the use of risk scores in general and the H-T score in particular: These scores should not be taken as a measure of the probability of having or developing glaucoma unless specifically validated for this purpose. For the Hood Reports in Figures 6 and 7, we expressed the H-T score, as well as the values for the six variables, as a percentile that is based on the large RW-RDB; for example, a percentile of 5% would simply mean that 95% of the eyes in the RW-RDB would have a larger value. 

Figure 6.

 

(A) Modified version of the Hood Report that includes the percentile based on the 4932 RW-RDB of each of the six variables in the H-T LRM. (B) The six variables from Figure 2 are shown schematically, where the label (e.g., TI+I, TS+S) indicates the region of interest. The left column of panel A shows the percentile (significance level) based on the RW-RDB of the H-T score and the six variables of the H-T LRM. These are color coded according to the legend at the bottom of panel A, which is the color code used for the probability/deviations maps of the Hood Report.

Figure 7.

 

(A) Modified Hood Report for one of the ON-G pattern eyes with the H-T score at the fourth percentile. Here, none of the six LRM variables are at the first percentile, but TS+S is at the second percentile and is coded bright red. The associated schematic illustrates the region of the donut that is thinned based on the TS+S variable. Notice also that the SVZ variable is at the fourth percentile, suggesting that the SVZ thickness is thinner than the IVZ. Each variable and its percentile score are represented by a corresponding white circle on the report. For example, the white “a” circle that corresponds to the TI+I variable (a, on the left table) is represented by its schematic in the bottom middle of the report. (B) Same as A but with additional information that could be used for training. The arrows without a black border show the location of the TS+S sectors (red) that are thinner at the first percentile level and the cp-SVZ region that is thinner than the cp-IVZ (orange) region at the third percentile level. To appreciate the structure–structure comparison, the arrows with the black border are physically identical locations. Thus, for example, if the thinning of the IVZ region is real (i.e., not an artifact), then the orange arrows with the black borders show physically identical locations and should agree in showing a pattern of thinning.

H-T Score As an Aid for Detecting Glaucoma in a Routine Clinical Practice

Likewise, in a clinical practice, we do not want to miss most, if not all, of the eyes in the clear ON-G cohort. However, we have alternatives to a single OCT metric such as g-cpRNFL or the H-T score. The Hood Report⁴³ used in this study was developed to allow the clinician to perform better than single metrics. Those with experience reading this report are able to detect characteristic tissue loss missed by the g-cpRNFL and/or g-GCL+IPL thickness.^2,15,43–45 However, experience and training are needed for optimal performance. To aid in the learning process, we developed the Columbia University (CU) OCT-based method with a decision tree to aid the clinician.^4,12 Although this method improves performance, it too takes time to learn and ultimately depends upon learning to interpret the relation among the various elements of the entire report. 

For clinical use, the six variables of the H-T LRM provide six metrics that have the potential to quantitatively bridge the gap between quantitative metrics and a partially subjective method such as the CU OCT-based method for reading Hood Reports. A possible modified version of the Hood Report in Figure 6A illustrates the general concept. The six variables in Figure 2 are shown schematically in Figure 6B, where the labels (e.g., TI+I, TS+S) indicate the region of interest, which is shaded light red in the schematics in Figure 6B. We are assessing if this region is significantly thinner than the region shown in gray in Figure 6B. The results are shown in the left column of Figure 6A as the percentile (significance level) based on the RW-RDB of the H-T score and the six variables of the H-T LRM. These are color coded using the color code shown at the lower left corner of Figure 6A. 

The H-T score of the eye in Figure 6A was at the second percentile of the RW-RDB (i.e., at the 98% specificity level of the RW-RDB) (Fig. 6A, red in top square of the left column). Of note, two of the six H-T variables (TS+S and G+G) had values at the first percentile (dark red). In fact, all 52 clear ON-G eyes had one or more of the six variables ≤ first percentile. In Figure 6A, the schematics that represent the TS+S and G+G are coded red on the report. 

Figure 7 shows the report for one of ON-G pattern eyes with the six variables and their associated regions on the report labeled (a) through (f) for ease of comparison. The H-T score of this eye was at the fourth percentile. Here, none of the 6 LRM variables are at the first percentile, but TS+S (Fig. 7b) is at the second percentile and is coded red. The associated schematic (Fig. 7b) on the report illustrates the region of the donut that is thinned based on the TS+S variable. Notice also that the SVZ variable (Fig. 7e) is at the fourth percentile, indicating that the SVZ thickness is significantly thinner than the IVZ. 

Figure 7B is the same as Figure 7A but with information that could be made available for training purposes. The arrows without a black border show the location of the TS+S sectors (red) that are thinner at the first percentile level and the SVZ region (orange), which is thinner than the IVZ (orange) region at the third percentile level. To appreciate the structure–structure comparison, the arrows with the black border are positioned at physically identical locations. Thus, for example, if the thinning of the IVZ region is real (i.e., not an artifact), then there should be evidence of glaucomatous damage at locations indicated by the orange arrows with black borders in Figure 7B. The same argument applies to the red arrows. 

Translational Relevance

The H-T score is an improvement over existing single metrics for purposes where a single metric is needed or required; for example, it may aid primary care providers in deciding on which individuals to refer to glaucoma specialist or for use in clinical trials. In addition, the H-T scores, and the associated six variables, offer a possible method for improving decisions in the clinic. In particular, the model provides a link between single metrics such as the H-T score and the subjective evaluation of elements (e.g., RNFL and GCL+IPL thickness maps) of the report as illustrated in Figures 6 and 7. 

Limitations and Improvements

Several limitations are worth noting. First, although the results are encouraging, the H-T LRM should be tested against other single metrics using a clinical reference standard based on other information such as visual fields, fundus exam, and/or photographs. Second, the H-T LRM also requires confirming on OCT instruments from other manufacturers. Third, the H-T metric has yet to be evaluated as a possible metric for monitoring progression. Finally, to be effective for screening, specificity and sensitivity should be improved. We are exploring ways to improve specificity and sensitivity by using anatomical measures such as fovea-to-disc distance,⁴² as well as the significance levels of the six metrics or variables of the H-T LRM. 

Conclusions

An OCT metric, the H-T score, had better sensitivity for detecting loss of tissue characteristic of glaucoma than did the most common OCT metrics used in clinical trials. The H-T score was developed based on six individual variables of cpRNFL and GCL+IPL regional thickness which were combined into a single metric with logistic regression analysis. Further work is needed to validate this approach with a clinical trial protocol; however, the new metric has the potential to improve the accuracy of referrals from primary care to specialist care via risk scores and calculators, as well as glaucoma definitions for clinical trials. In addition, a report that combines the H-T score with information about the percentile of the individual parameters may improve the clinician's use of the OCT for diagnosing glaucoma. 

Acknowledgments

The authors thank the optometry practices that contributed data to this analysis, including Park Slope Eye (New York, NY), Boerne Vision Center (Boerne, TX), Brandon Eyes (Middleton and Madison, WI), Visionworks (Tucker and Savannah, GA; Millville, NJ), and Total Eyecare (Lake Hopatcong and Denville, NJ). 

Supported by grants from Alcon Research Institute (DCH), Topcon (DCH), and National Eye Institute, National Institutes of Health (K99EY032182 to ET). 

Disclosure: D.C. Hood, Topcon (F, C, R), Heidelberg Engineering (F, C, R), Novartis (C); S. La Bruna, None; M. Durbin, Topcon (E); C. Lee, Topcon (E); A. Guzman, Topcon (E); T. Gebhardt, None; Y. Wang, None; A.L. Stowman, None; C.G. De Moraes, Carl Zeiss Meditec (C, R), Topcon (C), Heidelberg Engineering (C, R), Novartis (C), Thea Phar (C), Perfuse Therapeutics (C), Reichert (C, R), Ora Clinical (R); M. Chaglasian, Topcon (C); E. Tsamis, Topcon (R), Envision (C)