Understanding and identifying visual field progression

ABSTRACT

Detecting deterioration of visual field sensitivity measurements is important for the diagnosis and management of glaucoma. This review surveys the current methods for assessing progression that are implemented in clinical devices, which have been used in clinical trials, alongside more recent advances proposed in the literature. Advice is also offered to clinicians on what they can do to improve the collection of perimetric data to help analytical progression methods more accurately predict change. This advice includes a discussion of how frequently visual field testing should be undertaken, with a view towards future developments, such as digital healthcare outside the standard clinical setting and more personalised approaches to perimetry.

KEYWORDS:

• Gaucoma
• perimetry
• visual field metrics
• visual field progression

Introduction

Assessing visual function using perimetry is essential for glaucoma diagnosis and management.^1,2 While retinal imaging is now ubiquitous and is of tremendous utility in the management of glaucoma, the core goal of glaucoma treatment is to preserve visual function. Consequently, it is critical that visual function is assessed carefully in people with glaucoma and those with risk factors of developing glaucoma, to enable accurate evaluation of the current status of the vision of a person and whether there has been any deterioration to their vision.

The purpose of this narrative review is to discuss methods used to detect visual field change (hereafter referred to as ‘progression’) and to place these methods into varying clinical and research contexts. There are a wide variety of methods proposed in the research literature, implemented in medical device software, and used in clinical practice. This review will discuss the main classes of approaches used to identify visual field progression and compare different use cases for these methods. Consideration of the merits or otherwise of visual field assessment for diagnosis of glaucoma falls outside the scope of this review.

What limits the ability to detect visual field progression?

When considering the ability to determine visual field progression, it is important to recognise limitations in the accuracy and precision of clinical visual field testing. Any given visual field test result is a single snapshot in time that only allows an estimate of the visual field of a person. For locations in the visual field that have normal sensitivity, test–retest variability of visual fields as measured by current devices is typically very good, but it is well-established that test-retest variability in areas of damage can be very high (approximately half the dynamic range of the instrument).³ There are many sources of noise that contribute to test–retest variability of perimetric estimates. These include patient factors,^4,5 physiological factors associated with the disease process,⁶ how tests are administered,^5,7 and factors associated with the specific test software (for example).^3,8

Patient factors include learning and fatigue, changes in response criteria from test to test, and fixational eye movements. It is important to recognise the profound effect that instructions can play on response criteria, hence critical to ensure that people are provided with the same instructions at each visit regardless of the perimetrist in attendance.⁷ People with glaucoma report that visual field testing can be stressful, that they typically prefer to be guided through the procedure, and that a calm, quiet test environment is preferred.^9–11 Creating the most conducive environment for concentration on the task is likely to result in more accurate visual field estimates.

Algorithmic bias also contributes to test–retest variability in perimetric estimates. Most modern Bayesian perimetric algorithms bias outcomes towards either normal or highly damaged thresholds (around 0 dB) to speed up test procedures (for example).^8,12,13 This results in reduced precision of measurement when sensitivity is poor (between about 5 dB to 20 dB). Indeed, it has been shown that measurement variability in this range is so high that removing locations with thresholds estimated in this range from progression calculations does not change outcomes (i.e. these locations do not contribute constructively to the statistical determination of visual field progression).^14,15

Modern test algorithms also typically include some form of neighbourhood logic that considers the sensitivity of nearby locations when seeding the test procedure for a given location,¹⁶ or while updating sensitivity estimates during the procedure,^17–19 or as part of a post-processing procedure at the end of testing.^19,20 These processes have advantages in terms of test efficiency and can reduce test–retest variability, but also have the potential to mask small, localised defects when these are surrounded by areas of normal vision.^21,22

Steep scotoma borders are unable to be measured with precision due to spatial undersampling of the visual field by sparse test patterns (for example, the 24-2 test pattern places stimuli 6 degrees apart) and are smeared by even small fixational eye movements^23,24 (note, the 10-2 test pattern is discussed further below in the subsection ‘Impact of Test Pattern’). Indeed, a recent publication demonstrated using a combination of fundus tracking (which attempts to place the test stimulus at the same retinal location by tracking eye movements) and a test algorithm that did not include spatial biases, that visual field defects are measured as deeper and more localised than in other machines without those features.²²

Different clinical and research needs for progression analysis

Most research into statistical methods for analysis of visual field progression has considered either the needs for outcome measures for clinical trials or has analysed visual field data that has been returned from clinical trials. The needs of clinical trials may, at times, be different to those of clinical practice. In the context of clinical trials, there is increasing interest in devising new methods for detecting small amounts of progression (while keeping specificity intact) to improve the efficiency of clinical trials. If analytical methods can be devised that have improved sensitivity for detecting progression, this may lead to clinical trials being able to enrol fewer eyes or be of shorter duration, both of which remove barriers to testing of new therapeutics in glaucoma.

In clinical practice, in addition to determining whether a given eye has reached some defined endpoint to be classified as having progressed, the rate of progression is important for clinical decisions regarding individual patient management. The next section provides an overview of current methods for determining progression that are used in both clinical trials and clinical practice.

Methods for detecting progression in the literature have been broken into two classes: event-based, where progression is determined by comparing the current VFs to reference data and/or previously acquired baseline VF measurements; and trend-based, where some form of modelling (typically linear regression) is fit to all previously acquired VF data, and a judgement made based on the rate of change apparent in the model

In the following section the principles underlying these approaches are briefly reviewed. The section focusses on an overview of the classes of progression models, with particular emphasis on limitations of these approaches and proposed research solutions to reducing the impact of these limitations. There is currently no clear winner, however, aspects of this research are significantly influencing clinical trial designs and development of software for progression in commercial devices.

Event-based approaches

Clinical judgement is arguably the most commonly applied form of event analysis, where the clinician inspects the visual field data and compares to previous results subjectively based on whichever manual schema they have been taught. The relatively poor agreement in the outcomes of this process between clinicians is well recognised and creates the situation where there is no unified consensus on a gold-standard definition for a visual field progression event. The median Kappa statistic (a measure of inter-rater reliability where 1 is perfect agreement and 0 is no agreement) between ophthalmologists for subjectively determining whether a progression event has occurred has been reported as between approximately 0.30 to 0.55^25–27 which is moderate at best.

Numerous quantitative event approaches have been used in clinical trials and in perimetry device progression software. An illustrative example is the Guided Progression Analysis in the HFA software, which was derived from the Early Manifest Glaucoma Trial protocol.²⁸ The GPA compares the pattern deviation value on a pointwise level to the average of two selected baseline measurements.²⁹ The location is flagged if the measured pattern deviation at the subsequent visits falls outside the 95% test–retest variability of a group of people with glaucoma that were retested in close succession (and therefore assumed to be stable). When three points (not necessarily adjacent) reach the criteria on two repeated visits, then the eye is classified as having ‘Possible Progression’. A further confirmation leads to the classification of ‘Likely Progression’.

A limitation of event criteria is that it does not provide any basis for predicting future progression beyond the current test; that is, it does not provide an estimate of the rate of progression, just a probability that progression has occurred since baseline. Consequently, measurement noise in either the baseline data or the specific test of interest can be highly influential on the outcomes. As the number of visits since baseline increases, the chance of spurious results triggering a ‘progression event’ increases, and so methods should control for this in their specificity-sensititity tradeoffs.

A further limitation is that the criterion for change is determined from a population of observers that might be either more or less reliable than the individual observer that is being tested. Methods to personalise the criteria used for classifying change based on individual observer reliability have been proposed,³⁰ but not implemented commercially. There are many studies in the literature that have compared a wide variety of event criteria, on an equally wide variety of empirical and simulated longitudinal visual field series, to evaluate the trade-offs in terms of sensitivity and specificity for determining progression (for example).^31–33 A detailed discussion of this literature falls outside the scope of the current review, as recent advances in approach are limited.

Trend-based approaches

The most basic of the trend-based approaches is the use of linear regression either on the whole field (global), averages of groups of points (clusters), or on each point (point-wise). Typically, linear regression is used to determine whether there is a statistically different change from no change in visual field status (a line with a gradient of zero), although in a longer time series, ageing changes should also be considered. One way of largely correcting for ageing changes is to use indices that are corrected relative to age-matched normative data (for example: global mean deviation, or pointwise total deviation).

In addition to determining whether the eye is statistically likely to be progressing, it is clinically important to consider the rate of that progression. Progression rate is discussed further below in the context of test frequency, which influences the length of time it takes to reliably estimate progression rate.

Global trends (typically on mean deviation or a similar global index such as the HFA Visual Field Index) are often used as a coarse descriptor of visual field deterioration. By averaging over all locations, the disadvantage of point-wise variability can be reduced; however, glaucoma can result in localised defects that are masked by global indices. This is particularly the case in early disease. Analysis on a pointwise level can allow small regions to be detected, however, there is a trade-off between sensitivity and specificity, which needs to be carefully considered to avoid an unacceptable false positive rate.

A compromise between global and point-wise determination is to investigate change on pre-defined clusters of points. For example, the Octopus Perimeter (Haag-Streit AG, Switzerland) (Haag-Streit) software divides the field into 10 clusters and performs regression analyses on averages over these.³⁴ When compared to global and pointwise analyses, cluster trend analysis may detect deterioration sooner.^35,36

Regardless of the trend method used, because of the variability in visual field test estimates, lengthy series of visual fields of 8 or more test results may be needed to ensure adequate specificity in observers with higher levels of variability.³⁷ The issue of test frequency (in order to achieve this level of data in a reasonable period of chronological time) is addressed later in this review, however it is worth noting here that long test series collected over many years may become less likely to meet the assumptions of linearity that are commonly assumed in most commercial trend analysis methods. While there have been many papers in the literature that have compared different event criteria to various trend criteria (for example),^{31,32,38–40} when standard approaches are applied such as the pointwise event-based (GPA) and global trend – based methods available in the HFA, the ability to detect glaucomatous visual field progression is similar if tightly matched for specificity.³⁸

Recent advances on classic statistical approaches

Modern computational statistics have improved on the basic linear regression approach by incorporating known sources of variability into trend-based models in an attempt to improve the model estimates. Table 1 provides a summary of such approaches, alongside the motivation for the specific approach.

Table 1. Attributes of visual field testing that might be exploited to improve progression detection beyond basic linear regression and event analysis (column 1) with relevant methods that appear in the literature in the second column.

Source of variability or information	Methods
Global trends in long series of visual fields (e.g. more than 10 years) can exhibit exponential patterns of decay, not linear. This could be partly due to the summation of linear point-wise decays commencing at different times over the period. Long series of visual fields may not be suitable for traditional linear regression, particularly if locations have reached the 0 dB measurement floor. Note: there is little evidence that individual locations progress exponentially^Citation41	Exponential regression^Citation42 Glaucoma Rate Index^Citation43
At each visit, a person’s criteria, alertness, instructions, diurnal rhythm, etc might fluctuate. This could be due to time of day, diet, instructions from the perimetrist, sleep levels, and so on. Thus points in a VF series can undergo a ‘Global Visit Effect’ where all points are higher or lower than expected relative to other visits.	Hierarchical Bayesian Modelling incorporating the GVE^Citation44
In normal vision, visual field measurements form a Hill of Vision, where visual sensitivities decline with distance from the fovea. In glaucoma, damage often occurs in neighbouring locations and so spatial correlations between measurements at different locations can be strong. These can be exploited using methods more sophisticated than cluster analysis.	Analysis with Non-Stationary Weibull Error Regression and Spatial Enhancement (ANSWERS)^Citation45 Bayesian modelling [eg. SPROG^Citation46
Glaucoma often follows distinct spatial patterns of progression in SAP VFs. These have been described previously^Citation47 and more recently also extracted from data using machine learning methods.^Citation48 Generally, regression is performed on the groupings of points identified by the patterns in these methods.	Archetype analysis^Citation49 Gaussian Mixture Models^Citation50 Hierarchical clustering^Citation51 Deep Learning^Citation52 Variational Bayes Linear Regression^Citation53
It is well known that SAP measurements exhibit greater variability as thresholds are decreased (heteroskedasticity)^Citation3,Citation54	Hierarchical Bayesian Modelling^Citation41 Analysis with Non-Stationary Weibull Error Regression and Spatial Enhancement (ANSWERS)^Citation45
Floor effects as 0 dB is not bright enough	Hierarchical Bayesian Modelling^Citation41 Exponential regression^Citation55
Slope of regressions are left skewed: that is, lots of slopes are close to 0, but some are very big.	Linear Mixed Models with altered random effects distributions^Citation56

Combined trend-event approaches

There are several methods that combine the trend-based method of pointwise linear regression (PLR) to form a probability of progression at the end of a given VF sequence (an event). Two examples of this approach are the Permutation Analysis of Pointwise Linear Regression (PoPLR)⁵⁷ and ANSWERS⁵⁸ methods that combine the p-values of PLR at each location in the field to estimate an overall probability of calling a sequence of fields stable. A recent paper has shown the two methods give about the same sensitivity at matched specificity for detecting progression.⁵⁹ Another variant of this approach uses a Binomial test on p-values from pointwise linear regression.⁶⁰ Bayesian Hierarachical models have also been used explored for combining trend and event ideas.^41,61

Machine and deep learning methods

There has been an explosion in the application of neural network models (usually called ‘AI’) to many different problems in medicine and science. Approaches based on deep learning have been particularly effective in image analysis, hence useful for detecting progression in retinal images. Visual field measurements, however, contain far fewer data points than a retinal image, and hence are not readily suited to gain advantages from deep learning approaches.

Papers using neural models have shown good results in detecting progression, but none convincingly better than the other methods discussed above⁶² and so additional discussion is not included here. A further problem with neural models is that it is difficult to reason about why they make the decisions that they do, thus it might be more appropriate to apply them in a modified way that allows some interrogation.⁵²

The impact of test pattern

The statistical methods described in the literature have been largely developed and tested on data collected using the 30-2 or 24-2 test pattern of the Humphrey Field Analyzer, HFA (Carl-Zeiss Meditec, USA) because this device has been used in many of the major multicentre clinical trials related to glaucoma. Many visual field machines in practice use other test patterns, including those that are polar in orientation (eg. G-pattern of the Octopus Perimeter (Haag-Streit AG, Switzerland)),³⁴ or the test patterns of the Medmont perimeter,⁶³ or that have additional test locations in the central visual field (eg. the 24-2C of the HFA,⁶⁴ the New Grid of the Compass Perimeter (CenterVue, SpA Italy, iCare OY, Finland)).⁶⁵ Practical limitations exist to the density of test patterns, either due to device limitations (for example, the Humphrey Field Analyzer, HFA (Carl-Zeiss Meditec, USA) limits the addition of custom points to a minimum 1 degree grid), or pragmatically due to stimulus size limiting spacing. The same principles apply for detecting progression in these other test patterns, but some adjustments need to be made to maintain specificity (particularly for event or trend analyses where classifications are based on change being identified for a fixed number of locations). Similarly, new procedures have been proposed that dynamically alter the test pattern according to the individual visual field defects (for example, to test along scotoma borders.^66–68 Progression can be determined using standard event and trend methods, or methods such as PoPLR for these approaches, but care needs to be taken in designing the progression algorithms to ensure specificity is maintained.

In recent years, there has been increasing awareness of the possibility of central visual field loss in relatively early stages of glaucoma,⁶⁹ which has increased the usage and interest in the 10-2 visual field test pattern. Traditionally, the 10-2 was reserved for relatively late stage glaucoma (once it was deemed that the 24-2 was no longer able to effectively track islands of vision remaining within the central 10 degrees); however, contemporary practice suggests test pattern selection should take into account both the location of defects detected on prior visual fields and structural information. In patients with evidence of central loss, using the 10-2 pattern allows a mild reduction in the time taken to detect visual field progression, compared to the 24-2 alone,⁷⁰ and allows better estimation of rate of MD change.⁷¹ The presence of a baseline 10-2 defect may also have some mild predictive power for the likelihood of future 24-2 loss.⁷²

It is important to note though that glaucomatous spatial archetypes vary, such that some people will experience central visual field loss early in the disease, whereas others will demonstrate progressing visual field loss outside the 10-2 region. At present, careful clinical judgement is required to decide when to convert from one test to the other, to avoid missing progression of disease in either central or peripheral visual field. Alternately, using perimetric test patterns that incorporate additional central test points or that use a polar rather than cartesian distribution of test points can provide a ready means to increase sampling of the visual field.

The impact of stimulus size

The majority of clinical studies that have reported longitudinal data have used Size III stimuli (0.43 degrees of visual angle). As noted by others, this choice of stimulus size for conventional perimetry largely arose by historical accident rather than as an ideal scientific choice.⁷³ Wall et al.,⁷⁴ and related papers have experimented with Size V stimuli (1.77 degrees of visual angle) for detecting glaucomatous progression, but it is not a clear winner relative to Size III.

Choosing a stimulus size depends on many factors, a thorough discussion of which is beyond the scope of this review. These include: desired sampling of retinal ganglion cells in the context of spatial summation in glaucoma,⁷⁵ the impacts of optical scatter on small versus large stimuli,^76,77 sensitivity for detection of defects,⁷⁸ and measurement variability.⁷⁹ Of specific clinical relevance is that all these factors influence the measured perimetric threshold, so it is not sensible to swap stimulus size between tests. If stimulus size is changed, then a new baseline needs to be used for future progression analysis.

The impact of test frequency

The rate of progression is critically important to management decisions. Ideally, individuals with a rapid rate of visual field deterioration should be identified in as short a time as possible to tailor clinical management accordingly. Similarly, identifying those that are highly unlikely to be rapidly progressing is also important when deciding how to allocate clinical resources. Using linear regression on MD as the method for estimating the rate of progression, Chauhan et al.,⁸⁰ estimated the period of time and number of examinations required to measure differing rates of visual field decline with a fixed statistical power of 80% (see Table 1 of Chauhan et al.,⁸⁰ To detect very rapid decline (a change in MD of 4 dB over the first 2 years), in an observer with average variability, three examinations per year are required.

A follow-up study by Anderson et al., nicely demonstrates that not all individuals flagged with fast progression in short-time sequences actually have true underlying fast progression, so confirmatory signs may be advisable depending on the clinical context.⁸¹ Estimates of the prevalence of fast (<−1.0 to −2.0 dB/year) and catastrophic (<−2.0 dB/year) progression vary from approximately 5 to 15% depending on the population studied.^82–84

There are several suggested methods in the literature for increasing test frequency. One proposed method uses a modern fast thresholding approach and tests each eye twice at every clinical visit, which requires a similar overall test time as testing each eye once using older thresholding strategies.^85,86 An alternate method for increasing test frequency in the future may be to supplement visual field testing in the office with visual field testing at home using a portable device. A recent simulation study predicts that those with catastrophic visual field loss can be detected in 11 months with weekly home monitoring, relative to 2.5 years with regular 6 monthly clinic assessments.⁸⁷

Clinical recommendations and a view to the future

Regardless of the statistical methods used to advise the practitioner of the likelihood of progression and/or an estimate of the rate of progression, it is obvious that the higher the quality of the data collected at each visual field test, the more likely that meaningful visual field progression will be detected if present. Within the context of routine clinical care, all efforts should be taken to minimise sources of variability such as:

• ensuring instructions are consistent between visits;

• provision of a calm and supportive environment for testing;

• testing at a similar time of day where possible to try to match possible fatigue and alertness between test visits;

• ensuring that the patient is appropriately set up at the perimeter and that this is maintained throughout the test;

• and ensuring that the response time window is appropriate (some perimeters allow the response window to be slowed which can be particularly useful in people with visual field damage as response times are typically slower for stimuli that have a luminance intensity close to threshold).⁸⁸ Having a response time that is too fast causes anxiety and incorrectly labels responses as false positives (when the response is actually a genuine response provided to the preceding stimulus).

As visual field testing algorithms change over the lifetime of monitoring a patient, it is important to carefully consider the impact of changing test procedures on progression analysis. For example, it has been demonstrated that converting to SITA-Faster from SITA-Standard leads to similar visual field outcomes in those with normal vision or mild glaucoma; but results in improved MD values in those with moderate to severe glaucoma.⁸⁹ This algorithmic feature may mask disease progression in more advanced disease.

In the future, personalised versions of perimetry may enhance the ability to determine progression. For example, visual field test patterns that are customised to individuals to allow testing of progression at scotoma borders,^66–68 or adjustments to test algorithms to incorporate known features of individual observer variability.³⁰ Determining progression using retinal imaging is outside the scope of this review, however, consideration of the disease holistically necessitates consideration of both structure and function.

Future analytical methods will likely combine information from perimetry with structural imaging and risk-factors create multifactorial progression risk scores. In doing so, it is important to recognise that visual field testing and retinal imaging measure very different features and do not necessarily predict each other except in general terms.⁹⁰ In later stages of disease, considering visual field progression is critical because retinal imaging measures can reach the measurement floor and become unable to detect further deterioration.⁹¹

Finally, evidence for or against a fast rate of progression, which is clinically meaningful information, can be determined in a shorter time if more visual field tests are performed. While not currently widely adopted, visual field test frequency may be increased in the future via home monitoring. There are an increasing number of commercial providers of portable visual field devices suitable for home monitoring, and open-source versions of software are also available.⁹² Currently, further evaluation is required for many of these devices to ensure that they are rigorously validated against standard perimetry devices, to ensure that they are suitable for quantitative assessment of likelihood of visual field progression.

Not all patients are interested in home monitoring as a solution,¹⁰ and not all patients will be able to afford or have the digital literacy for these types of solutions. Developments of financial models for successful wide-spread deployment alongside clinical review mechanisms for the wealth of data that may be obtained at home are still in their infancy.

Disclosure statement

Authors AMM and AT have received research support from the following ophthalmic device manufacturers: Haag-Streit AG, Heidelberg Engineering GmBH, CREWT Medical Systems, CenterVue SpA, iCare OY.