Article Text
PDF
Abstract
Aim Devise simplified formulae, using preoperative clinical data, to give risk estimates of (1) failure and (2) proliferative vitreoretinpathy (PVR) following primary retinal detachment repair by vitrectomy.
Methods 641 patients were analysed as part of an RCT investigating use of 5-fluorouracil and low-molecular-weight heparin. Treatment status had no effect on success rates and did not therefore form part of the analyses. Preoperative risk factors for surgical failure and for PVR within 6 months of retinal detachment surgery were identified, and a multiple variable logistic regression model developed. Further analyses were performed to devise a simple points system to produce risk estimates of failure.
Results Three risk factors were related to failure—previous lens extraction (p=0.046), grade C PVR (p=0.039) and extent of detachment (p<0.001). Three risk factors were also related to failure due to PVR—vitreous haemorrhage (p=0.088), grade C PVR (p=0.044) and extent of detachment (p<0.001). There was good agreement between risk estimates produced by the points system and those calculated directly using a multivariate regression model. The points-system model gave an area under the receiver operating characteristic curve of 0.658. The receiver operating characteristic curve for the PVR model gave an area under the curve of 0.8399 suggesting greater diagnostic value.
Conclusions A simple points system may be used as a clinical guide to identify patients at higher risk of failure following retinal detachment repair by vitrectomy. This may help clinicians select appropriate surgical approaches and stratify cases in research and surgical training.
- Retinal detachment
- vitrectomy
- failure
- proliferative vitreoretinopathy
- risk factors
- formula
- retina
- treatment surgery
Statistics from Altmetric.com
- Retinal detachment
- vitrectomy
- failure
- proliferative vitreoretinopathy
- risk factors
- formula
- retina
- treatment surgery
Introduction
Recent reports of the primary success rate of rhegmatogenous retinal detachment repair range from 64 to 91%.1–6 Retinal redetachment following initial surgery may occur secondary to the development of proliferative vitreoretinopathy (PVR), failure of treatment, for example missed or partially treated breaks, or the development of new retinal breaks. The visual consequences of failure as a result of PVR are well documented, with only 11–25% of patients achieving a visual acuity of 20/100.7–10 Advances in vitreoretinal surgical techniques have resulted in a broader range of treatment strategies to improve primary success rates and visual outcomes. This is, however, dependent on surgeons being able to identify those patients at an increased risk of failure and modify surgical approaches accordingly. The potential development of adjuvant therapies also requires methods of selecting appropriate cases.
Some high-risk features may be obvious, for example the presence of a giant retinal tear, but others may be more subtle and may be additive. Preoperative characteristics have been identified that may be associated with an increased risk of surgical failure—the duration of symptoms, the extent of retinal detachment, involvement of inferior retinal quadrants, high myopia and hypotony.2 11–13 Analysis of risk factors associated with the development of PVR (the most common cause of surgical failure with an incidence of 5.1–11.7%) has been undertaken to identify possible targets for adjuvant therapies.14–17 Risk factors identified for PVR include aphakia, uveitis, the presence of vitreous haemorrhage, area of retinal detachment, the presence of PVR preoperatively, raised intraocular pressure and increased age.14 15 17
Previous work, using multiple variable analysis, has constructed formulae to allow clinicians to estimate the risk of surgical failure due to PVR; however, these are complex and can be difficult to use in clinical situations.14 18 Furthermore, these formulae may rely on intraoperative data to provide an adequate level of accuracy.
The aim of this study was to devise simplified formulae for (1) failure and (2) development of PVR following retinal detachment repair by vitrectomy, which would be of use both in research and in clinical practice. Using the results of analyses of retinal detachment characteristics, we developed a simple points system, a method for simplifying complex statistical models that produces risk estimates which may be calculated easily and used in a clinical setting without the use of computers and calculators. The model was derived from preoperative data collected in a prospectively randomised controlled trial on the use of adjunctive low-molecular-weight heparin and 5-fluorouracil in patients undergoing primary vitrectomy and gas exchange for retinal detachment repair. The methodology employed was based on that used to develop models of risk of coronary heart disease based on the Framingham study.19
Methods
Ethics approval was granted by the Moorfields Local Research Ethics Committee (CHAD 1005). Data were handled in accordance with the Data Protection Act and were recorded on a Microsoft Access database.
Collection of patient data
Six hundred and forty-one patients were recruited from two specialised vitreoretinal units between February 2001 and January 2005.20 In brief, patients presenting with a primary rhegmatogenous retinal detachment undergoing repair by vitrectomy and injection of intraocular gas were recruited to the trial and randomly allocated to treatment or placebo groups. The leading indications for primary vitrectomy surgery included retinal detachment with a posterior vitreous detachment owing to superior posterior, multiple or tractional retinal breaks, pseudophakia and/or an inadequate view of the fundus—for example, secondary to vitreous haemorrhage. Eyes without a posterior vitreous separation and/or a limited peripheral retinal detachment would normally undergo scleral buckling surgery.
The use of antimetabolites restricted recruitment to males over 16 years of age and postmenopausal women. Complicated retinal detachments were excluded, that is giant retinal tears (defined as peripheral retinal tears greater than 3 clock hours in circumferential extent), intended silicone oil tamponade and retinal detachment secondary to posterior penetrating trauma. Diabetic retinopathy, glaucoma, corneal opacity sufficient to impair surgical view, no light perception vision, inability to give informed consent and unwillingness to accept randomisation were also exclusion criteria. The primary outcome measure for this trial was retinal reattachment following primary vitrectomy without any reoperations at 6 months and was classed as a ‘failure’ if retinal redetachment was observed within the 6-month follow-up period. The trial adhered to the tenets of the Declaration of Helsinki.
Of the 641 patients recruited, 615 completed the study and were included in the analysis of risk factors for surgical failure. Ninety-six patients were classed as ‘failures,’ and of these, 37 failed due to the development of grade C PVR. Treatment status (with or without adjuncts) was found to have no effect on success rate and was therefore not part of the risk factor analysis for this study.20 To further investigate for any treatment effect, a sensitivity analysis using only placebo data was performed, and the same variables were identified as being associated with all failure outcome. Data were split into a training data set (n=465) and a validation set (n=150) for identification of risk factors for the model and testing its predictive powers by cross-validation.
Risk factor association to failure
The risk factors analysed had previously been associated with retinal detachment failure or PVR development14 15 21 and were, in addition, readily available at preoperative examination and could be measured accurately. Factors analysed were age, duration of symptoms, pathological myopia (defined as a spherical equivalent of −6 dioptres or more), intraocular pressure, presence of vitreous haemorrhage, fundus-obscuring vitreous haemorrhage, previous lens extraction, preoperative uveitis, number of retinal quadrants detached, number and distribution of retinal breaks, and presence of grade C PVR (Retinal Society classification).22 Risk factors were cross-tabulated against failure and a univariate logistic regression analysis undertaken to assess evidence of any association between each putative risk factor and failure. For these analyses, data from the training set only were included.
Of the risk factors analysed, there was evidence of association between failure from any cause and previous lens extraction, number of retinal quadrants involved and presence of grade C PVR (table 1). Failure due to PVR was associated with vitreous haemorrhage, presence of grade C PVR and the extent of retinal detachment (table 1).
Univariate logistic regression analysis of possible risk factors for anatomical failure following primary surgery
Development of multivariable model to predict surgical failure due to any cause
A multiple variable logistic regression model was developed relating the three risk factors identified by univariate analysis to the risk of failure (table 2). As part of model checking, interactions between the risk factors in the multiple variable model were examined using likelihood ratio tests, and little evidence of codependence was found. A single model was then constructed using the combined sample in order to develop the simplified scoring system. Summary statistics on the risk factors considered are shown in table 3.
Parameters of the multivariable model relating the three risk factors identified by univariate analysis to the risk of surgical failure
Summary statistics on risk factors and outcome. Failure defined as retinal redetachment within 6 months of primary surgery
Referent risk-factor profile
Risk factors were organised into meaningful categories and a reference value for each category determined (table 4, columns a–c). A referent risk factor profile was selected by choosing a base category for each risk factor (WiREF). The base category was assigned 0 points in the scoring system—for example, one quadrant of detached retina. Less healthy risk factor states were assigned positive points incrementally (Wij)—for example, two quadrants of retinal detachment scored 1, three quadrants of retinal detachment scored 2 and so on (table 4, column c).
Parameters of the multivariable model used to calculate point scores for each risk factor for surgical failure (columns labelled for ease of reference)
Determination of distance of risk factor from the base category in regression units
This was computed for each risk factor using βi (Wij–WiREF) (table 4, column e), where βi is the regression coefficient calculated from the multiple variable logistic regression model (table 2, table 4 column d).
Constant for the points system defined (B)
The constant for the points system was the number of regression units which corresponded to one point. We based this on the increase in risk associated with lens extraction, that is B=0.1922.
Point score assigned to each risk-factor category
Points associated with each category of each risk factor were computed by: Pointsij=βi(Wij–WiREF)/B (table 4, column f). Points were rounded to the nearest integer.
Risk of failure associated with each point total
The theoretical range of the points total using the scoring system was 0 to 12. The risk estimate was attached to each point total using the multiple logistic regression equation.
The point totals, when multiplied by the constant (B=0.1922) approximates to:
(the component of the model which depends on the risk factors under consideration.)
The model requires:
Therefore, an estimate of the intercept (β0) was added, that is β0=−2.8175.
This was substituted into the logistic model, and an estimate of risk for each possible total point score was generated (table 5).
Estimate of failure risk associated with each total point score for surgical failure
Development of multivariable model to predict surgical failure due to PVR
The same methodology was also used to develop a model to predict the risk of surgical failure due to PVR using the three risk factors identified by univariate analysis (table 6). Interactions between the risk factors in the multiple variable model were examined using likelihood ratio tests, and little evidence of codependence was found. A summary of this analysis is given in table 7.
Parameters of the multivariable model relating the three risk factors identified by univariate analysis to the risk of surgical failure due to proliferative vitreoretinpathy
Parameters of the multivariable model used to calculate point scores for each risk factor for failure due to proliferative vitreoretinpathy
Results
Estimates of risk of failure calculated using the scoring system
The risk estimates associated with each score are presented in table 5. Risk estimates have been grouped together to provide a guide to the increased risk of failure for a given range of scores—for example, a score between 0 and 3 is associated with a less than 10% risk of failure following retinal detachment surgery.
Using a Pearson χ2 test, there was little evidence of a poor model fit (p=0.73). The receiver operating characteristic (ROC) curve (figure 1) was plotted, and the area under the ROC curve was calculated to be 0.658.
Receive operating characteristic (ROC) curve for surgical failure.
Demonstration of the model using clinical examples
The following examples illustrate the correspondence between risks estimated by the logistic regression model and those approximated by the points system.
Case 1: a patient who has had lens extraction, has two quadrants of retinal detachment and has no PVR (grade C)
The estimate of risk using the point scoring system is 0.0961 (table 8).
Calculation of point score and associated estimate of risk for case 1: a patient who has had lens extraction, has two quadrants of retinal detachment but does not have proliferative vitreoretinpathy grade C
The exact risk estimate from the logistic model is
The points system gives an estimated risk of failure of 9.6%, and employing the model directly also gives an estimate of 10.3%.
Case 2: a patient who has not had lens extraction but has PVR grade C and all four retinal quadrants detached
The estimate of risk using the scoring system is 0.3311 (table 9).
Calculation of point score and associated estimate of risk for case 2: a patient who has not had lens extraction but has PVR grade C and all four retinal quadrants detached
The exact risk estimate from the logistic model is
The points system gives a failure risk of 33%, while the model directly gives 35%.
Estimates of risk of failure due to PVR calculated using the scoring system
The risk estimates associated with each score are presented in table 10. A score between 0 and 4 is associated with a less than 10% risk of redetachment due to PVR following vitrectomy. The ROC curve (figure 2) was plotted, and the area under the ROC curve was calculated to be 0.8399.
Estimate of failure risk due to proliferative vitreoretinpathy associated with each total point score for surgical failure
Receive operating characteristic (ROC) curve for failure due to PVR.
Demonstration of the model using clinical examples
The following example illustrates the correspondence between risk estimated by the logistic regression model and that approximated by the points system.
Case 3: a patient who has a vitreous haemorrhage but does not have grade C PVR and has one quadrant of retinal detachment
The estimate of risk using the scoring system is 0.0427 (table 11).
Calculation of point score and associated estimate of failure due to proliferative vitreoretinpathy for case 3; a patient who has a vitreous haemorrhage but does not have grade C proliferative vitreoretinpathy and has one quadrant of retinal detachment
The risk estimate from the logistic model is
The points system gives an estimated risk of failure of 4.3%, and employing the model directly gives an estimate of 4.5%.
Discussion
Our analysis of the risk factors for failure of vitrectomy surgery for retinal detachment has produced two simplified points systems. One estimates the risk of surgical failure from any cause, while a second formula was developed to estimate the risk of failure due to the development of PVR; both could be of use in research and in clinical practice. Subsequently, we have demonstrated good agreement between risk estimates produced by these points systems and those calculated directly using a multivariate regression model.
It is worthy of mention that univariate analysis on the training data identified cataract surgery as a risk factor for surgical failure due to any cause. The accompanying paper (where all data were used) estimated a higher risk of failure in subjects who had had cataract surgery, albeit not statistically significant at the 0.05 level. The strategy adopted when developing regression models depends on the purpose of the analysis, and different approaches to choosing a regression model can result in different models being developed. We felt it important to keep cataract surgery in the prognostic model because it is simple to assess, and its inclusion assists in prediction.
The risk formula developed to calculate the risk of failure from any cause following surgery used significant risk factors identified by regression analysis—previous lens extraction, the area of retina detached and preoperative grade C PVR. This model gave an area under the ROC curve of 0.658, indicating that our scoring system does not have absolute diagnostic precision (few diagnostic tests do); however, it provides a guide to clinicians allowing identification of patients with an increased risk of failure of primary retinal detachment surgery. Moreover, it uses routinely observed preoperative clinical data without reference to more complex formulae and additional tests. The accuracy of risk estimates is likely to be reduced when they are not calculated directly from logistic models, particularly when continuous variables form part of the model. In the model we developed, however, all the categories which had a significant impact on outcome were discrete variables, minimising the loss of information and precision in this way.
Our second model calculated the risk of failure due to the development of PVR. Although other formulae predicting the risk of PVR following vitrectomy have been published previously, they have required complex calculations, included per-operative data and have failed to gain widespread acceptance.14 18 23 For example, in Kon and coworkers' series,14 vitreous protein levels formed part of the analysis; furthermore there was a notably high PVR rate (29.4%) suggesting more complex primary detachments. Some preoperative factors, for example uveitis, previous laser or cryoretinopexy, were present in small numbers in our study, and it is conceivable that a possible association was not detected in our analysis. It is notable that the simplified formula to predict PVR had a greater diagnostic accuracy than the formula for failure due to all causes—the area below the ROC curve being 0.8399 compared with 0.658.
All patients analysed in this study underwent primary pars plana vitrectomy, and thus the use of the formula is directly applicable to those retinal detachment patients undergoing primary vitrectomy. Of these, only 30 of 641 (4.7%) had an additional scleral buckle at the time of surgery.24 Nevertheless, it is notable that studies investigating factors associated with failure following scleral buckling surgery have, in general, identified similar risk factors to those from vitrectomy studies, and the points system may be of value more broadly to all retinal detachments including those undergoing buckling surgery.2 11–13 Further prospective analyses would establish its value. The points systems we have developed give an estimate of the risk of failure within the 6-month trial duration following primary surgery. It is possible that some patients may have failed after 6 months; however surgical failure beyond 6 months is very uncommon, and the numbers are likely to be very small.
We present two simple points systems that may be used as clinical guides to identify patients at higher risk of failure following vitrectomy and gas tamponade for retinal detachments. This may allow clinicians to modify their surgical approach (eg, selecting silicone oil tamponade) or potentially to use an adjunct to maximise the primary success rate. In addition, the points system may be a useful tool to stratify cases when conducting clinical research. It may also be of value in training retinal specialists to foster an awareness of identifiable clinical features which may raise the risk of failure of primary retinal detachment repair.
References
Footnotes
Linked articles 190306.
Funding This study was supported by a major Ophthalmology grant from the Royal College of Surgeons Edinburgh, the Royal Blind School Edinburgh/Scottish War Blinded, the W.H. Ross Foundation for the Prevention of Blindness and the Chief Scientist Office Scotland (CZB/4/705). The authors acknowledge (a proportion of their) financial support from the Department of Health through the award made by the National Institute for Health Research to Moorfields Eye Hospital NHS Foundation Trust and UCL Institute of Ophthalmology for a Specialist Biomedical Research Centre for Ophthalmology. The views expressed in this publication are those of the authors and not necessarily those of the Department of Health.
Competing interests None.
Ethics approval Ethics approval was provided by the Moorfields Ethics Committee.
Provenance and peer review Not commissioned; externally peer reviewed.