Abstract
Purpose.:
Trend analysis techniques to detect glaucomatous progression typically assume a constant rate of change. This study uses data from the Ocular Hypertension Treatment Study to assess whether this assumption decreases sensitivity to changes in progression rate, by including earlier periods of stability.
Methods.:
Series of visual fields (mean 24 per eye) completed at 6-month intervals from participants randomized initially to observation were split into subseries before and after the initiation of treatment (the “split-point”). The mean deviation rate of change (MDR) was derived using these entire subseries, and using only the window length (W) tests nearest the split-point, for different window lengths of W tests. A generalized estimating equation model was used to detect changes in MDR occurring at the split-point.
Results.:
Using shortened subseries with W = 7 tests, the MDR slowed by 0.142 dB/y upon initiation of treatment (P < 0.001), and the proportion of eyes showing “rapid deterioration” (MDR <–0.5 dB/y with P < 5%) decreased from 11.8% to 6.5% (P < 0.001). Using the entire sequence, no significant change in MDR was detected (P = 0.796), and there was no change in the proportion of eyes progressing (P = 0.084). Window lengths 6 ≤ W ≤ 9 produced similar benefits.
Conclusions.:
Event analysis revealed a beneficial treatment effect in this dataset. This effect was not detected by linear trend analysis applied to entire series, but was detected when using shorter subseries of length between six and nine fields. Using linear trend analysis on the entire field sequence may not be optimal for detecting and monitoring progression. Nonlinear analyses may be needed for long series of fields. (ClinicalTrials.gov number, NCT00000125.)
Baseline data and design of the OHTS have been described previously.
1,2 All OHTS participants signed a statement of informed consent before study entry after having the risks and benefits of participation explained to them. The institutional review boards at each participating clinical site approved their respective informed consent statements and procedures. The study adhered to the tenets of the Declaration of Helsinki.
All subjects enrolled in the OHTS had to have at least two reliable (fixation losses, false-negative, and false-positive responses <33%), achromatic, automated VF test results (Humphrey Field Analyzer using the 30-2 testing pattern; Carl Zeiss Meditec, Inc., Dublin, CA) that were within normal limits during the qualifying period. The OHTS analysis dataset available for this study contained all VF tests and endpoint determinations in the OHTS database as of March 2009. From the full OHTS dataset (n = 1636), we first removed from further consideration any eye that reached an endpoint that was determined by the endpoint committee to be due to causes other than POAG (261 eyes of 202 subjects). We then selected only those follow-up VF tests that were considered reliable (false-positives, false-negatives, and fixation loss all <33% if the Full Threshold Algorithm was used, false-positives <15%, false-negatives and fixation loss <33% if the Swedish Interactive Threshold Algorithm [SITA] was used).
A “Delayed Treatment Cohort” was formed, consisting of those eyes that had a change from observation to treatment at some point during follow-up. This analysis included eyes randomized initially to observation, and that began treatment either as a result of a POAG endpoint determination or at the transition between the first and second phases of OHTS (at which time treatment was offered to all originally untreated subjects in the observation arm). The visit at which an eye first was noted to be on treatment became known as the “split-point” for that eye.
For a given window length of W tests, four sequences of VF tests were analyzed per subject: BeforeW, the W most recent VF tests before the split-point; AfterW, the first W VF tests beginning at least 9 months after the split-point; BeforeAll, all VF tests before the split-point; and AfterAll, all VF tests beginning at least 9 months after the split-point.
A 9-month gap was left between commencing treatment and the start of the “after” sequence to ensure that the participant's treatment had a chance to stabilize, allowing their physician to determine a drug and dosage that resulted in attaining the target IOP. Subjects with fewer than W VF tests before and W VF tests after the split-point were excluded. We did not require that treatment be continuous once commenced.
Secondly, a “Continuous Treatment Cohort” was formed consisting of those eyes having no change in treatment status during the study; this, therefore, consisted of subjects randomized initially to treatment. For this cohort, the series was split chronologically into two equal parts, with a “split-point” halfway through the series. In this case, no gap to allow treatment to stabilize is needed. Therefore, the sequence AfterAll consisted of all VF tests after the split-point.
Linear regression of mean deviation (MD) over time was performed separately for each of the four sequences for each of the eyes selected. MDs were considered to be equivalent between the two testing algorithms. The rate of change of MD (MDR [dB/y]) was recorded for each eye, together with the standard error of the slope estimate, and a determination of whether “rapid deterioration” occurred during that period (defined as an MDR worse than −0.5 dB/y that was significantly negative, with P < 0.05).
An effect of the initiation of treatment on MDR was sought using data from both eyes of each individual (where available). A paired comparison was performed to determine the change in MDR that occurred at the split-point. Specifically, the change in MDR given by MDR “after” minus MDR “before” was set as the outcome of a generalized estimating equation (GEE) regression
12 with no independent variables. The resulting intercept term and
P value were used as estimates of the average change in MDR and its level of significance, respectively. This is analogous to a paired
t-test comparing the two MDRs, but using a GEE regression to account for the fact that there may be correlated data from two eyes for the same participant. Additionally, the proportion of eyes for which “rapid deterioration” occurred was calculated. These proportions were compared between the “before” and “after” series using McNemar's test.
This analysis was repeated for different window lengths W, ranging from W = 4 (assumed to be the shortest window length over which linear regression can provide a reasonable estimate of the rate of change) to W = 12 (the longest window length for which there were sufficient eyes with enough VF tests before and after their split-point). Note that as the window length W increases, the number of eyes for which there are sufficient VF tests in both sequences is reduced; therefore, the average MDR in the BeforeAll and AfterAll sequences varies depending on the value of W.
Our results indicated that when linear trend analysis techniques are used, using the entire available sequence of VFs can decrease the sensitivity to detecting known changes in the rate of functional progression. Shorter sequences of only the more recent fields before the split-point make linear trend analysis more sensitive to such changes, without compromising specificity. Series of between six and nine VFs provided the best sensitivity in this analysis. While shorter sequences could be expected to result in more “false-positive” cases wherein progression is flagged in stable eyes, no evidence of a significant reduction in specificity was found. To our knowledge, this is the first report suggesting that using the entire VF sequence may not be optimal for detecting and monitoring progression when using linear trend analysis techniques.
An implication of these findings is that the linear model for progression is suboptimal for long series of visual fields in ocular hypertensive eyes. This is not surprising, since such a model assumes a constant rate of change. Consider a 60-year-old patient, currently with an MD of −5 dB, progressing at −1.0 dB/y. Using a linear model and extrapolating the trend would imply that at the age of 40 their MD had been +15 dB, which clearly is nonsensical. Similarly, if a patient's disease has progressed to the extent that their sensitivity has reached 0 dB, no further decline corresponding to a linear model is possible. A linear model for progression would overestimate the amount of change in sensitivity once this floor has been reached. Nonlinear models for progression that allow for periods of zero apparent change at the start and at the end of a series are needed. The OHTS dataset used for our study does not contain a full range of disease severities (since very few subjects experienced severe vision loss by the end of the study), and so no specific nonlinear model can be advocated at this stage without the necessary data for validation. In the meantime, if linear models for progression are to be used, it is recommended that the sequence be plotted against time and inspected visually to detect acceleration (for individual patients), and/or analysis restricted to the more recent visual fields in the series (especially in the case of clinical studies). In addition, the use of a linear model to fit data measured on a logarithmic scale (dB) should be interpreted with caution. Although it simplifies how clinicians measure and understand progressive visual field change, one should be reminded that a given amount of dB loss in the early stages of the disease (e.g., 1.0 dB) corresponds to a smaller absolute sensitivity change than in later stages for the same dB loss. Using shorter series of tests should minimize this confounding effect when measuring rates of change in dB/y in eyes with normal to early damage in the visual field. This hypothesis must be tested in eyes with advanced field loss.
Clinical assessment of progression in glaucoma has included seeking evidence of acceleration in the rate of progression. Even if perimetry were free of variability, so that the rate of change could be known precisely, it still could not be assumed that a patient's rate of change is constant, with no need for future testing. A previously stable patient can start to progress rapidly, and we cannot yet predict the onset of this change. Therefore, in some respects, our findings are more in agreement with current clinical practice than the principle of using a patient's entire series of fields. Indeed, the EyeSuite program developed to analyze longitudinal series of results from the Octopus perimeter (Haag Streit International, Berne, Switzerland) by default assesses the rate of change over the most recent six tests, rather than the entire series. While it may seem counterintuitive not to use all the data (especially in light of the fact that the biggest advantage of trend analysis methods is that they make better use of the available data than event analysis techniques), as seen in the example in
Figure 2, there is a sound logical underpinning behind assessing the relatively recent rate of change.
The optimum series length to maximize sensitivity when aiming to detect progression will depend on many factors. In our study, series of six to nine VFs provided good sensitivity when tests were done at six-month intervals, spanning a period of 2.5 to 4 years. If testing were performed only annually (as often is the case for ocular hypertensive patients), four years may be insufficient, since it would provide a series of only five fields. However, the optimum still may be to analyze subseries of fewer than nine fields to detect shorter-term changes in the progression rate. By contrast, if three tests were to be performed annually, as has been recommended for some purposes,
9 it may be possible to increase the series length (hence, reducing variability about the estimate of rate of change) without compromising sensitivity of detecting rapid progression. Another factor to be taken into consideration is that some patients produce more variable VFs than others, and so may require longer series for progression to become apparent.
The optimum analysis method and optimum series length could depend on disease severity. The main justification for not always using the first few fields in the series with linear analysis is that the patient may be stable for some time before progression begins, as in the example in
Figure 2. A patient who already has developed a glaucomatous defect would be considered less likely to have a prolonged period of stability before progression accelerates. In addition, variability is much higher in more advanced disease, potentially making estimates of the rate of change based on fewer fields unreliable. As variability increases, robustness becomes more important, favoring linear models over nonlinear models with higher numbers of free parameters, but using longer series of fields to obtain more accurate estimates of the rate of change. Since this dataset does not contain a large number of cases of moderate or severe glaucoma, this conjecture would need testing in a different dataset.
Using the entire sequence, a significant change in MDR was observed at the split-point in the Continuous Treatment Cohort, with these eyes progressing more rapidly in the second half than in the first half of the study. No change in treatment status occurred during their sequence, as all patients in this cohort were treated from the start of the study (although the treatment given may have changed). It is possible that this is a chance characteristic of the data. However, even though the sample size is reduced when
W is large, 986 eyes still would be considered more than adequate. It may be that this effect is caused by a significant number of those eyes beginning to progress towards the end of the sequence. It also could be indicative of nonlinearity of progression, with sensitivities accelerating downwards, as would be consistent with our previous findings in another dataset that the current MD is predictive of the rate of subsequent change.
13 Finally, it also would be consistent with the presence of a learning effect causing sensitivities to rise over the first few fields of the series
14 ; however, all OHTS participants were required to have had previous fields indicating at least some familiarity with automated perimetry. Although we refer to this group as the Continuous Treatment Cohort because their treatment status did not change, it cannot be assumed that they were not undergoing any glaucomatous progression. The fact that progression may be accelerating in many of the eyes in the Continuous Treatment Cohort that have been managed consistently, biasing the results towards a more rapid MDR later in the study, makes it all the more impressive that initiating treatment had such a clear beneficial effect in the Delayed Treatment Cohort.
A caveat with the findings is that commencement of treatment could have been due to the participant reaching an endpoint in the first stage of the study, or could have been due to the decision to offer treatment to everyone in the second stage of the study. Participants who reached an endpoint in the first phase will have shorter sequences BeforeW available, and so would not be eligible for inclusion when using longer window lengths W. This could explain partly the greater change in MDR when using, for example, W = 4 instead of W = 8. However, as W becomes quite large (7 or 8 fields), such participants will form only a small proportion of the sample size n, and so it is unlikely that this is driving the main conclusions of the analysis. Notably, when the same analysis was performed varying W but consistently using all series with at least 10 fields before and after the split-point, the change in MDR still was greater for smaller W.
Our study used MD to generate a measure of the rate of functional change, corresponding to disease progression. MD is useful as a global measure in clinical trials, such as the OHTS, but is insensitive to deterioration of small scotomas in individual patients. Clinically, change in MD would be just one of several measures used to determine whether changes in treatment are necessary. Point-wise changes are more variable. However, the same principle would apply, and nonlinear methods developed for MD are likely to have similar benefits when applied to point-wise data.
The main conclusion to be drawn from our study is that using the entire series of test results for linear trend analysis actually may be detrimental to early detection of rapid visual field change, especially when that progression is sporadic or preceded by a period of stability. At this early disease stage, rates of change were underestimated consistently when the entire sequence was used. Use of shorter sequences improved the ability to detect slowing of the rate of progression at the time treatment was initiated. By contrast, use of the shorter sequences did not cause a significant increase in the number of series for which a change in rate was detected in the absence of a change in treatment status. These results underscore the need for nonlinear models for progression, while also providing a method to reduce the problem until such models have been developed and validated. Such techniques could make trend analysis more sensitive to changes in the rate of progression, allowing earlier detection and implementation of appropriate treatments.
Arizona.
California.
Richard S. Baker, MD, Charles R. Drew University
Fermin P. Ballesteros Jr, University of California, San Francisco
James D. Brandt, MD, UC Davis Department of Ophthalmology
Jeffrey J. Casper, MD, UC Davis Department of Ophthalmology
Anne L. Coleman, MD, PhD, Jules Stein Eye Institute, UCLA
Michael V. Drake, MD, University of California, San Francisco
Margarita X. Gonzalez, Jules Stein Eye Institute, UCLA
Simon K. Law, MD, PharmD, Jules Stein Eye Institute, UCLA
Michele C. Lim, MD, UC Davis Department of Ophthalmology
Shan C. Lin, MD, Department of Ophthalmology, University of California, San Francisco,
Ivan R. Schwab, MD, UC Davis Department of Ophthalmology
J. Rigby Slight, MD, University of California, San Diego
Robert L. Stamper, MD, Department of Ophthalmology University of California, San Francisco
Patricia W. Tam, University of California, San Francisco
Robert N. Weinreb, MD, Department of Ophthalmology, University of California, San Diego
Eunice Williams-Steppe, MA, Department of Ophthalmology, University of California, San Diego
Washington, DC.
Frank S. Ashburn, MD, Eye Associates of Washington, DC
Clete C. Clark, Washington Eye Associates, Physicians & Surgeons
Douglas E. Gaasterland, MD, Eye Doctors of Washington
Linda A. Manier, Eye Doctors of Washington
Karen D. Schacht, Eye Associates of Washington, DC
Arthur L. Schwartz, MD, Washington Eye Associates, Physicians & Surgeons
Howard S. Weiss, MD, Washington Eye Associates, Physicians & Surgeons
Florida.
Donald L. Budenz, MD, University of Miami School of Medicine, Bascom Palmer Eye Institute
Francisco E. Fantes, MD, University of Miami School of Medicine, Bascom Palmer Eye Institute
Steven J. Gedde, MD, University of Miami School of Medicine, Bascom Palmer Eye Institute
Richard K. Parrish, MD, University of Miami School of Medicine, Bascom Palmer Eye Institute
Georgia.
Allen D. Beck, MD, Emory University Eye Center
Anastasios P. Costarides, MD, PhD, Emory University Eye Center
Donna Leef, Emory University Eye Center
Thomas S. Harbin, MD, Eye Consultants of Atlanta
Lea M. Danner, Eye Physicians & Surgeons
Teresa A. Long, Eye Consultants of Atlanta
Paul McManus, MD, Eye Physicians & Surgeons
Charles J. Patorgis, OD, Eye Consultants of Atlanta
Shelly R. Smith, Eye Physicians & Surgeons
Julie M. Wright, Eye Consultants of Atlanta
Kentucky.
Maryland.
Donald A. Abrams, MD, Krieger Eye Institute, Sinai Hospital of Baltimore
Sameer I. Ahmad, MD, Department of Ophthalmology, Johns Hopkins School of Medicine
David S. Friedman, MD, MPH, Department of Ophthalmology, Johns Hopkins School of Medicine
Ramzi Hemady, MD, Maryland Center for Eye Care Associates
Eve J. Higginbotham, MD, Maryland Center for Eye Care Associates
Henry D. Jampel, MD, Department of Ophthalmology, Johns Hopkins School of Medicine
Irvin P. Pollack, MD, Krieger Eye Institute, Sinai Hospital of Baltimore
Kevin L. Powdrill, Maryland Center for Eye Care Associates
Harry A. Quigley, MD, Department of Ophthalmology, Johns Hopkins School of Medicine
Pradeep Y. Ramulu, MD, Department of Ophthalmology, Johns Hopkins School of Medicine
Alan L. Robin, MD, 6115 Falls Road, Third Floor
Donald J. Zack, MD, PhD, Department of Ophthalmology, Johns Hopkins School of Medicine
Michigan.
Juan L. Allen, Kresge Eye Institute, Wayne State University
Monica Y. Allen-Alexander, MD, Kresge Eye Institute, Wayne State University
Terry J. Bergstrom, MD, WK Kellogg Eye Center, University of Michigan
David A. Crandall, MD, Henry Ford Medical Center – Troy, Department of Ophthalmology
Deborah Darnley-Fisch, MD, Department of Ophthalmology, Henry Ford Medical Center – Fairlane
Aldo Fantin, MD, , Department of Ophthalmology, Henry Ford Medical Center – Troy
Melanie Gutkowski, Department of Ophthalmology, Henry Ford Medical Center – Troy
Julianne Hall, Department of Ophthalmology, Henry Ford Medical Center
Bret A. Hughes, MD, Kresge Eye Institute, Wayne State University
Nauman R. Imami, MD, Department of Ophthalmology, Henry Ford Medical Center
Mark S. Juzych, MD, Kresge Eye Institute, Wayne State University
Mark L. McDermott, MD, Kresge Eye Institute, Wayne State University
Sy E. Moroi, MD, PhD, WK Kellogg Eye Center, University of Michigan
John M. O'Grady, MD, Kresge Eye Center, Great Lakes Ophthalmology
Carol J. Pollack-Rindle, WK Kellogg Eye Center, University of Michigan
John M. Ramocki, MD, Kresge Eye Institute, Wayne State University
Michelle A. Tehranisa, WK Kellogg Eye Center, University of Michigan
Linda A. Van Conett, Kresge Eye Center, Great Lakes Ophthalmology
Minnesota.
Missouri.
Edward Barnett, MD, Department of Ophthalmology & Visual Sciences, Washington University Medical Center
Benard Becker, MD, Department of Ophthalmology & Visual Sciences, Washington University Medical Center
Anjali M. Bhorade, MD, Department of Ophthalmology & Visual Sciences, Washington University Medical Center
Jamie D. Kambarian, Department of Ophthalmology & Visual Sciences, Washington University Medical Center
Michael A. Kass, MD, Department of Ophthalmology & Visual Sciences, Washington University Medical Center
Carla J. Siegfried, MD, Department of Ophthalmology & Visual Sciences, Washington University Medical Center
New York.
Yokasta E. Abreu, MD, New York Eye Ear Infirmary
Jeffrey M. Liebmann, MD, New York Eye Ear Infirmary
Robert Ritch, MD, New York Eye Ear Infirmary
Celso A. Tello, MD, New York Eye Ear Infirmary
Ohio.
N. Douglas Baker, MD, Ophthalmology Surgeons Consultants
Robert J. Derick, MD, The Ohio State University, Havener Eye Institute
Kathleen A. Lamping, MD, University Suburban Health Center
Kathryne McKinney, The Ohio State University, Havener Eye Institute
Cheryl Vitelli, University Suburban Health Center
Paul A. Weber, MD, The Ohio State University, Havener Eye Institute
Oregon.
Pennsylvania.
G. Richard Bennett, MS, OD, Pennsylvania College of Optometry
Sheri G. Drossner, Scheie Eye Institute, University of Pennsylvania
Joan C. DuPont, Scheie Eye Institute, University of Pennsylvania
Eydie G. Miller-Ellis, MD, Scheie Eye Institute, University of Pennsylvania
Jane F. Niemczyk, Glaucoma Care Clinic
Jody R. Piltz-Seymour, MD, Glaucoma Care Clinic
Prithvi S. Sankar, MD, Scheie Eye Institute, University of Pennsylvania
Texas.
Peter T. Chang, MD, Baylor Eye Clinic, Baylor College of Medicine
Pamela M. Frady, Baylor Eye Clinic, Baylor College of Medicine
Ronald L. Gross, MD, Baylor Eye Clinic, Baylor College of Medicine
Silvia Orengo-Nania, MD, Baylor Eye Clinic, Baylor College of Medicine