**Purpose.**:
The purpose of our study is to develop and validate a model to predict visual field (VF) outcomes in patients with treated glaucoma.

**Methods.**:
Data from 587 eyes with treated glaucoma evaluated in a cohort were used to develop two equations to predict VF outcomes, one estimating the risk of progression (%) and another estimating the global rate of VF sensitivity change (decibels [dB]/year). These equations, which included variables associated with VF progression in a multivariable model, then were tested in another cohort (*n* = 62 eyes) followed for at least 4 years. Agreement, discrimination, and calibration of the model in the validation sample were assessed as main outcome measures.

**Results.**:
The mean difference between observed and predicted global rates of sensitivity change was 0.13 dB/year (95% confidence interval [CI] = 0.06 to 0.18 dB/year) and the mean difference between observed and predicted final VF mean deviation (MD) values was 0.37 dB (95% CI = 0.00 to 0.75 dB). The predictive model had moderate discriminative ability to estimate VF progression in the independent sample (c-index of 0.78, 95% CI = 0.59 to 0.97).

**Conclusions.**:
To our knowledge, this is the first attempt to generate and validate a risk model for patients with treated glaucoma. The prediction model showed moderate accuracy in estimating future VF outcomes in an independent glaucoma population, and may be useful for the objective assessment of risk of progressive VF loss.

^{ 1–5 }Nevertheless, risk assessment for eyes with established disc and field damage in clinical practice remains subjective and qualitative. Patients usually are classified as being at “high,” “moderate,” or “low” risk of glaucoma onset or progression on the basis of clinical features and risk factors. Such empirical approaches vary substantially among clinicians, and limit consistent and standardized case management, and the development of practice guidelines to enhance patient care.

^{ 6 }This would allow clinicians to personalize individual treatment approaches, and to determine better the risk and benefit of treatment modifications. This challenge has been met partly for patients with ocular hypertension. The Ocular Hypertension Treatment Study (OHTS) Group,

^{ 1 }in collaboration with the European Glaucoma Prevention Study (EGPS)

^{ 7 }developed a 5-year risk calculator that permits objective estimation of the risk of conversion to glaucoma among patients with statistically elevated intraocular pressure (IOP) and normal visual field (VF) tests.

^{ 8 }This type of risk calculator improves clinician assessment regarding treatment versus observation.

^{ 9–11 }

^{ 8 }This corresponds to a small proportion of patients in clinical practice, and none of those with established glaucomatous optic neuropathy and/or VF loss. Finally, the OHTS risk calculator uses baseline information to predict the future risk of progression over a 5-year period, but does not provide estimates of the rate of progression, and does not take into consideration the intercurrent variables that interact to increase the risk of progression, such as disc hemorrhage (DH),

^{ 12 }and IOP parameters, such as its peak, mean, and fluctuation.

^{ 13–15 }

^{ 11 }Based on clinical information, such as disease severity and life expectancy, calculators for treated glaucoma also would likely increase agreement among clinicians with regard to future management of individual cases. One potential benefit of risk calculation in treated glaucoma is that it would allow clinicians to estimate objectively the risk of VF progression in the mid- and long-term before spending time and resources on numerous VF examinations or waiting until progression actually occurred.

^{ 15 }A number of risk factors were associated significantly with progression in this population. Additionally, we used trend analysis to understand better the role of different risk factors on the rate of VF progression among patients treated with current modalities of therapy.

^{ 15 }and a prospective clinical trial at Bascom Palmer Eye Institute.

^{ 16 }Both studies followed the tenets of the Declaration of Helsinki, and were approved by the respective Institutional Review Board and Committee of Ethics. There were two parts to this study: 1) development of a calculator able to predict the risk (%) and rate (decibels [dB]/year) of glaucomatous VF progression based on data from a retrospective cohort (for didactic purposes this sample was named “reference” population), and 2) test the calculator in an independent cohort, and compare the predicted versus the observed outcomes (this cohort was termed the “validation” population).

^{ 15 }All patients were experienced at perimetry, and had had at least 8 VF tests (24-2 SITA-Standard, Humphrey Field Analyzer II; Carl Zeiss Meditec, Inc., Dublin, CA) during a mean follow-up time of 6.4 ± 1.7 (range 2–10) years. Using trend analysis to evaluate VF progression, we reported a number of clinical variables significantly (

*P*< 0.10) associated with rapid VF progression: age (years), central corneal thickness (CCT, microns), mean IOP (mm Hg), peak IOP (mm Hg), detection of disc hemorrhage, presence of beta-zone parapapillary atrophy (βPPA), presence of exfoliation syndrome (XFS), and follow-up time (years).

^{ 16 }The methods used to collect all variables were the same as the ones reported in GAPS.

^{ 15 }Inclusion criteria consisted of refractive error between −7.00 and +3.00 diopters (D), best corrected visual acuity ≥20/40, age range between 40 and 80 years, reliable standard automated perimetry (SAP, <33% rate of fixation losses, false positives, and false negatives) and no prior history of intraocular surgery except for uncomplicated cataract extraction. Subjects with ocular diseases other than glaucoma or cataract, best-corrected visual acuity <20/40, or unreliable SAP tests were excluded. All patients underwent a baseline examination consisting of a complete ophthalmic examination, including slit-lamp biomicroscopy, gonioscopy, Goldmann applanation tonometry, ultrasound pachymetry, dilated stereoscopic examination, and photography of the optic disc, and SAP testing. Follow-up SAP examinations were performed at 6-month intervals. All patients underwent a dilated eye examination with optic disc stereophotography at annual visits.

*P*< 0.05 on at least two consecutive examinations.

^{ 17 }We used the default definition of progressing VF locations provided by the manufacturer, that is a test point was deemed progressing if the rate of sensitivity decline was >1.0 dB/year at

*P*< 0.01. For an eye to be considered progressing, at least two adjacent points in the same hemifield had to meet the above criteria. This definition likely improves specificity to define progression, as it takes into consideration the topographic orientation of the nerve fiber layer bundles that are damaged in glaucoma.

^{ 15 }First, a logistic regression was used to evaluate the role of each IOP parameter on VF progression. All analyses were adjusted for the length of follow-up, which consisted of the time interval between the first and last VF tests entered in the analysis. Each variable was tested first in a univariable model. Those with

*P*< 0.25 then were entered in the multivariable analysis. Since glaucoma filtering procedures lead to more substantial IOP lowering than medical therapy, which has been shown to slow the rates of VF progression,

^{ 18,19 }we included the occurrence of any type of incisional glaucoma surgery during follow-up in the multivariable model. Given that IOP variability during follow-up differs between filtered and non-filtered eyes,

^{ 18,19 }we also added the interaction term “glaucoma surgery*mean IOP” to the model. A backwards elimination procedure then was used to derive the final model (alpha-level = 0.05). The aforementioned variables were used to generate two equations: A) an equation in which the dependent variable is the risk of progression (%) in a given number of years based on the independent variables derived from our reference population, and B) an equation in which the dependent variable is the global rate of VF change, or rate of VF progression (dB/year).

*p*is the probability of the outcome of interest, or dependent variable (VF progression);

*Logit*(

*p*) is the natural log of the odds ratio for progression;

*b*

_{0},

*b*

_{1},

*b*

_{2}…

*b*are the regression coefficients of the multivariable regression equation; and

_{k}*X*

_{1},

*X*

_{2},

*X*

_{3}…

*X*are the independent variables (age, CCT, IOP, etc.).

_{k}*y*is a dependent, continuous variable (global rate of VF change [dB/year]); β

_{0}, β

_{1}, β

_{2}… β

_{k}are the regression coefficients of the multivariable regression equation; and

*X*

_{1},

*X*

_{2},

*X*

_{3}…

*X*are the independent variables (age, CCT, IOP, etc.).

_{k}*MD*is the final MD,

_{f}*MD*is the baseline MD,

_{b}*δ*is the global rate of change (dB/year) obtained from equation B, and ΔT is the future time interval one wishes to predict the VF outcomes.

- Hosmer-Lemeshow test (test of goodness of the fit for the predictive model),
- Bland-Altman plots comparing predicted versus observed rates of VF change as a continuous variable (dB/year),
- Bland-Altman plots comparing predicted versus observed final VF MD values (dB),
- Linear regression coefficients (
*R*^{2}) comparing predicted and observed final VF MD values, - Calibration of predicted and observed risks (%) of progression,
- Calibration of global rates using categorical separation of subgroups based on quartiles of predicted rates of progression: very fast, fast, moderate, and slow progressors,
- Performance of the model to discriminate risk (%) of progression and observed outcomes (c-statistic).

^{ 8,21 }Categorical variables were compared between the reference and validation samples using the chi-square test. Independent samples

*t*-tests were used for comparisons of continuous variables. Discrimination was defined as the ability of a predictive model to separate glaucomatous eyes with and without progression. In other words, discrimination is an estimate of the probability that the model assigns a higher risk for those with progression compared with those who remained stable based on our predefined progression criteria.

^{ 22 }Discrimination was assessed by calculating the c-statistic as proposed by Harrell et al.,

^{ 23 }which is analogous to the area under the receiver operating characteristic curve. A c-index value of 0.5 indicates random predictions, whereas a value of 1.0 indicates perfect prediction.

^{ 8,21 }In each of the quartiles, the average predicted risk was compared with the average outcome during the follow-up time. Since a greater number of visual fields improves the performance of trend analysis to measure rates of progression minimizing the effect of VF variability,

^{ 24 }we tested the hypothesis that our model would predict the VF outcomes more accurately in eyes with longer sequences of tests. For that purpose, we performed a secondary analysis only on eyes with 10 or more VF tests. Computerized statistical analyses were performed using MedCalc (MedCalc software v.3.3; MedCalc, Inc., Mariakerke, Belgium) and SPSS version 16.0 (SPSS Inc., Chicago, IL).

*P*= 0.67). For the linear regression model, which used global rates of progression (dB/year) as a dependent variable, the coefficient of determination was

*R*

^{2}= 0.14, the adjusted-

*R*

^{2}was 0.13, and the multiple correlation coefficient was 0.37 at

*P*< 0.001. The Table 1 compares clinical characteristics between the reference and validation populations. The reference population had worse mean baseline VF MD, higher peak and mean IOP, greater prevalence of exfoliation syndrome and βPPA, and were followed longer.

**Table. 1**

**Table. 1**

BPEI ( n = 62) | NY-GAPS ( n = 587) | P Value | |

Age at baseline (years) | 67.4 ± 8.3 | 64.9 ± 13.0 | 0.03 |

Baseline MD (dB) | −3.7 ± 4.4 | −7.1 ± 5.1 | < 0.01 |

Vertical cup-disc ratio | 0.64 ± 0.1 | 0.72 ± 0.1 | < 0.01 |

Mean CCT (μ) | 539.4 ± 38.4 | 540.9 ± 37.3 | 0.97 |

Peak IOP (mm Hg) | 17.3 ± 4.5 | 19.9 ± 4.5 | < 0.01 |

Mean IOP (mm Hg) | 13.4 ± 3.3 | 15.2 ± 3.1 | < 0.01 |

N of eyes with DH (%) | 8 (13) | 53 (9) | 0.49 |

N of eyes with βPPA (%) | 42 (67) | 370 (63) | 0.35 |

Exfoliation syndrome (%) | 3 (4.8) | 84 (14) | 0.04 |

N of eyes undergoing glaucoma surgery (%) | 6 (9.6) | 206 (35) | < 0.01 |

Follow-up time (years) | 4.0 ± 0.9 | 6.4 ± 1.7 | < 0.01 |

*R*

^{2}= 0.90,

*P*< 0.01).

**Figure 1.**

**Figure 1.**

**Figure 2.**

**Figure 2.**

**Figure 3.**

**Figure 3.**

**Figure 4.**

**Figure 4.**

^{ 15 }Two equations were generated: one that provides the risk of glaucoma progression (%) based on pre-defined PLR criteria in a given number of years, and another that estimates future rates of VF change based on clinical characteristics. Assuming linearity of glaucoma VF progression, the latter equation allowed estimating the VF MD value for an individual eye in a given number of years. Despite differences between the two populations and inherent challenges related to populations seen in clinical practice, our model revealed a moderate performance in prediction VF outcomes. Therefore, the calculator was robust despite the significant differences in the reference and validation populations illustrated in the Table. This observation supports the generalizability of our model to other populations, evidently taking into account a margin of error and the fact that the model tended to overestimate rates and risk of progression.

^{ 25–29 }Among the first and most important is the Framingham Risk Score, derived from the Framingham Heart Study to estimate the 10-year risk for coronary heart disease outcomes (myocardial infarction and coronary obstruction).

^{ 28 }This model has proven useful as a means of stratifying patients based on their risk profiles, which allows customized, individualized intervention. In ophthalmology, the OHTS and the DIGS developed risk models to assess the 5-year risk of conversion to primary open-angle glaucoma among patients with ocular hypertension,

^{ 8,21 }and have proven clinically useful by improving agreement rates among clinicians about the time to start treatment for ocular hypertensive patients.

^{ 10,11 }Boland et al. showed that using the OHTS risk calculator changed treatment recommendations by glaucoma specialists.

^{ 10 }They also noted that decisions became more consistent and more confident, and the average risk threshold for recommending treatment resembled those suggested by expert opinion and cost-benefit analysis.

^{ 8,21 }and Framingham

^{ 28,29 }models revealed moderate performance when tested in different populations. In the OHTS and DIGS-derived models, the ability of the model to predict conversion to glaucoma based on probabilities and c-statistics revealed a moderate result (0.75 in both studies), which is similar to our finding in the much more complex scenario of established, treated disease. It is important to emphasize that the existing models provide risk estimates based on the natural history of the disease, and have been established based on the observation of prospective studies. Our model, on the other hand, included a heterogeneous population undergoing different modalities of treatment.

^{ 15 }The fact that treatment was not guided by standard protocols adds substantial source of variability to our model, even though its performance was comparable to what has been described for the other currently available models.

*R*

^{2}value of linear regression equation was low (0.13), which suggests that only 13% of the variance in the outcome variable (rates of progression) could be explained by the linear model. The causes of a small

*R*

^{2}in a linear regression are various, such as measurement error, biological variation, variables not entered in the model or sometimes entered as indirect estimates, and due to a non-linear relationship between outcome and predictor variables.

^{ 30 }These all are inherent limitations of using a mathematical model to explain a biological phenomenon that still is poorly understood, such as glaucoma progression. We are unable to compare our results directly with the OHTS/EGPS or DIGS calculators, since their publications do not report measures of goodness of the fit of their model, which was based on Cox proportional hazards modeling using an event-based approach. However, this observation underscores the importance of interpreting the results of the calculator with caution, not using its results to tailor glaucoma therapy, but rather providing an objective, numerical variable to measure risk and that could be added to the clinician repertoire of factors used to monitor glaucoma.

*. 2002;120:714–720. [CrossRef] [PubMed]*

*Arch Ophthalmol**. 2007;114:1965–1972. [CrossRef] [PubMed]*

*Ophthalmology**. 2000;130:429–440. [CrossRef] [PubMed]*

*Am J Ophthalmol**. 2009;116:200–207. [CrossRef] [PubMed]*

*Ophthalmology**. 2003;136:805–813. [CrossRef] [PubMed]*

*Am J Ophthalmol**. 2008;145:191–192. [CrossRef] [PubMed]*

*Am J Ophthalmol**. 2007;144:266–275. [CrossRef] [PubMed]*

*Am J Ophthalmol**Ocular Hypertension Treatment Study Group, European Glaucoma Prevention Study Group*, Gordon MO Validated prediction model for the development of primary open-angle glaucoma in individuals with ocular hypertension.

*. 2007;114:10–19.*

*Ophthalmology**. 2006;15:426–431. [CrossRef] [PubMed]*

*J Glaucoma**. 2008;17:631–638. [CrossRef] [PubMed]*

*J Glaucoma**. 2008;53:S11–S16. [CrossRef] [PubMed]*

*Surv Ophthalmol**. 2001;131:699–708. [CrossRef] [PubMed]*

*Am J Ophthalmol**. 2004;45:4346–4351. [CrossRef] [PubMed]*

*Invest Ophthalmol Vis Sci**. 2000;9:134–142. [CrossRef] [PubMed]*

*J Glaucoma**. 2011;129:562–568. [CrossRef] [PubMed]*

*Arch Ophthalmol**. 2011;95:1122–1127. [CrossRef] [PubMed]*

*Br J Ophthalmol**. 1996;80:40–48. [CrossRef] [PubMed]*

*Br J Ophthalmol**. 2002;18:489–498. [CrossRef] [PubMed]*

*J Ocul Pharmacol Ther**. 2010;149:258–264. [CrossRef] [PubMed]*

*Am J Ophthalmol**. 2009;127:1610–1615. [CrossRef] [PubMed]*

*Arch Ophthalmol**. 2005;123:1351–1360. [CrossRef] [PubMed]*

*Arch Ophthalmol**. 2004;23:2109–2123. [CrossRef] [PubMed]*

*Stat Med**. 1996;15:361–387. [CrossRef] [PubMed]*

*Stat Med**. 2008;92:569–573. [CrossRef] [PubMed]*

*Br J Ophthalmol**. 2003;18:1039–1052. [CrossRef] [PubMed]*

*J Gen Intern Med**Hong Kong Diabetes Registry*, Yang X So WY Development and validation of an all-cause mortality risk score in type 2 diabetes.

*. 2008;168:451–457.*

*Arch Intern Med**. 2011;114:1336–1344. [CrossRef] [PubMed]*

*Anesthesiology**. 2001;286:180–187. [CrossRef] [PubMed]*

*JAMA**. 2007;5:91–96. [CrossRef] [PubMed]*

*Am Heart Hosp J**. New York, NY: Oxford. 2010;258–259.*

*Intuitive Biostatistics*. 2nd ed