Abstract
This study repeats previous meta-analyses of the prevalence and progression of diabetic retinopathy in pregnant populations with preexisting diabetes.
We have recently published a systematic review and meta-analysis1 of the prevalence and progression of diabetic retinopathy (DR) in pregnant populations with preexisting diabetes. In a recent letter by Zhou and Shen,2 it was suggested that linear random-effects models have a limitation in analyzing pooled estimates from studies with small to moderate sample sizes, such as those in our recent meta-analysis.1 We chose to use the Freeman-Tukey double-arcsine transformation (FTT) method for our random-effects model for reasons previously discussed.3 However, generalized linear mixed modeling (GLMM) is a relatively newer method that can avoid the mentioned issues.4 Therefore, as an exercise, we have repeated our meta-analyses using GLMM methods and present the outcomes from both the (published) FTT and GLMM methods with the inclusion of prediction intervals, as suggested by Zhou and Shen.2
Methods
Detailed methodology around study selection and statistical methods has been described in our original article.1 In addition, here, the Stata IC (Stata Corp) version 15.1 additional module command metapreg5 was used to construct random intercept logistic regression models (a relevant GLMM, as described by Stijnen et al6) for the meta-analytic outcomes of interest published previously.1 The Stata command metapreg was also used to estimate prediction intervals.
We have kept data from type 1 diabetes and type 2 diabetes pooled together in this analysis to allow for comparison with previously published results, but acknowledge that ideally the diabetes types would be addressed separately. However, at present, there are too few published studies of type 2 diabetes during pregnancy to elucidate whether this condition may come with differing risks of progression compared with the better-studied type 1 diabetes.
Results
The Table presents results from both previously published FTT models and the new GLMM method with the prediction intervals of each outcome of interest. The study team found that the results from the GLMM model were similar to the results produced using the FTT model, except for the outcome of progression from nonproliferative DR (NPDR) to proliferative DR (PDR). It is unclear if the discrepancy between the pooled estimates from these 2 analytical models arises because 4 of 16 studies included in this outcome of interest have 0 cases (Figure 3 in Widyaputri et al).1 These 0 event studies would be expected to be problematic for a logit- or logistic-transformation 2-step method, but should not affect the GLMM estimates.4
Table. Pooled Estimates and 95% CIs and Predictive Intervals From Random-Effects Meta-analysis of Diabetic Retinopathy (DR) in Pregnancy.
| Outcome of interest | Included studiesa | Rate per 100 pregnancies (95% CI) | 95% Prediction interval (from metapreg command)b | |
|---|---|---|---|---|
| Stata command metaprop_one with FTT option (previously published results) | Stata command metapreg for a random intercept GLMM modelb | |||
| Prevalence | ||||
| Any DR in early pregnancy | High-quality studies; similar grading scheme; T1D, 9 studies (N ranges, 12-205); T2D, 1 study (N = 160) | 55.5 (38.9-71.6) (eFigure 4A in the Supplement)1 | 55.1 (40.7-68.7) | 15.1-89.5 |
| Any DR around delivery | High-quality studies; similar grading scheme; T1D, 5 studies (N ranges, 16-205); T2D, 1 study (N = 160) | 59.5 (36.1-80.8) (eFigure 4B in the Supplement)1 | 59.7 (39.5-77.1) | 7.4-96.5 |
| PDR in early pregnancy | High-quality studies; similar grading scheme; T1D, 11 studies (N ranges, 12-205); T2D, 1 study (N = 160) | 6.0 (2.2-11.2) (eFigure 5A in the Supplement)1 | 5.5 (2.4-11.9) | 0.2-59.9 |
| PDR around delivery | High-quality studies; similar grading scheme; T1DM, 7 studies (N ranges,16-205); T2D, 1 study (N = 160) | 6.4 (1.7-13.4) (eFigure 5B in the Supplement)1 | 6.2 (2.6-14.3) | 0.2-61.4 |
| Progression | ||||
| None to any DR | High-quality studies; T1D, 15 studies (N ranges, 6-86); T2D, 1 study (N = 145) | 15.0 (9.9-20.8) (Figure 1)1 | 15.8 (11.3-21.6) | 7.7-29.4 |
| Worsened NPDR | High-quality studies; T1D, 16 studies (N ranges, 4-101); T2D, 2 studies (N of 15 and 19) | 30.9 (23.3-39.2) (Figure 2)1 | 30.3 (22.1-40.0) | 10.2-62.3 |
| NPDR to PDR | High-quality studies; T1D, 14 studies (N ranges, 4-101); T2D, 2 studies (N of 15 and 19) | 6.3 (3.3-10.0) (Figure 3)1 | 8.2 (5.5-11.8) | 5.1-12.8 |
| Worsened PDR | High-quality studies; T1D, 12 studies (N ranges, 2-25) | 37.0 (21.2-54.0) (Figure 4)1 | 38.2 (24.4-54.3) | 10.7-76.2 |
Abbreviations: DR, diabetic retinopathy; FTT, Freeman-Tukey double arcsine transformation; GLMM, generalized linear mixed modeling; NPDR, nonproliferative diabetic retinopathy; PDR, proliferative diabetic retinopathy; T1D, type 1 diabetes; T2D, type 2 diabetes.
N = the study-specific total subjects with relevant data.
Random intercept logistic regression model using Stata commands5: metapreg event denom, studyid (study) by (diabetesTypeString) model (random) ci (wilson) prediction.
As expected by its calculation, prediction intervals were wider than 95% CIs around pooled estimates. It is interesting, however, that prediction and 95% CIs seemed to be better aligned for progression estimates. In the case of progression of NPDR to PDR—where FTT and GLMM estimates differed moderately—the prediction and GLMM 95% CIs overlap almost perfectly (95% prediction interval, 5.1-12.8 vs 95% CI, 5.5-11.8 per 100 pregnancies).
Discussion
If one can accept the concept of a universe of trials and accept the use of prediction intervals in the context of a meta-analysis of single proportions, we could expect, at best, 1 in every 12 (and at worst 1 in every 3) pregnant individuals with preexisting diabetes without DR will go on to developing DR during their pregnancy. In those with preexisting diabetes and NPDR, the best case might be that 1 in every 10 pregnant individuals could expect their DR to worsen during pregnancy (increasing to 2 in 3 in the worst-case scenario).
Overall estimates of prevalence and progression from FTT and GLMM methods are extremely similar. Pregnant individuals with preexisting diabetes do develop DR during pregnancy, the severity of DR does worsen in a significant number throughout its course, and more evidence is required to truly know its effect on those with type 2 diabetes.
References
- 1.Widyaputri F, Rogers SL, Kandasamy R, Shub A, Symons RCA, Lim LL. Global estimates of diabetic retinopathy prevalence and progression in pregnant women with preexisting diabetes: a systematic review and meta-analysis. JAMA Ophthalmol. 2022;140(5):486-494. doi: 10.1001/jamaophthalmol.2022.0050 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Zhou S, Shen C. Statistical considerations for meta-analysis of diabetic retinopathy prevalence. JAMA Ophthalmol. Published online September 8, 2022. doi: 10.1001/jamaophthalmol.2022.3470 [DOI] [PubMed] [Google Scholar]
- 3.Widyaputri F, Rogers S, Lim L. Statistical considerations for meta-analysis of diabetic retinopathy prevalence [Reply]. JAMA Ophthalmol. Published online September 8, 2022. doi: 10.1001/jamaophthalmol.2022.3473 [DOI] [PubMed] [Google Scholar]
- 4.Lin L, Xu C. Arcsine-based transformations for meta-analysis of proportions: pros, cons, and alternatives. Health Sci Rep. 2020;3(3):e178. doi: 10.1002/hsr2.178 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Nyaga V. METAPREG: Stata module to compute fixed and random effects meta-analysis and meta-regression of proportions. Accessed on August 19, 2022. https://ideas.repec.org/c/boc/bocode/s458693.html
- 6.Stijnen T, Hamza TH, Ozdemir P. Random effects meta-analysis of event outcome in the framework of the generalized linear mixed model with applications in sparse data. Stat Med. 2010;29(29):3046-3067. doi: 10.1002/sim.4040 [DOI] [PubMed] [Google Scholar]
