This is a single-center and retrospective study (registered at ClinicalTrials.gov: NCT05631054). It was approved by the Ethics Committee of Tianjin Medical University Eye Hospital (approval number: 2022KY-27) and was conducted in accordance with the Declaration of Helsinki. We included 186 PDR patients (250 eyes) who underwent primary vitrectomy in Tianjin Medical University Eye Hospital from January 2016 to October 2021. We defined 12 months postoperatively as a long-term period because previous studies have indicated that patients could gain stable visual acuity after this period of time (7, 8). The key inclusion criteria were: (1) >18 years of age; (2) diagnosed with type I or II diabetes mellitus; (3) voluntarily signed an informed consent form; (4) presence of un-clearing VH (1–6 months); and (5) TRD involving or threatening the macular confirmed by ophthalmic examinations. The key exclusion criteria were: (1) a follow-up period of<12 months; (2) prior intraocular surgeries in the study eye (except cataract surgery); (3) missing medical records; (4) combined with non-diabetic complications (e.g., retinal vein occlusion, macular hole, keratopathy); (5) surgeries of the study eye within 3 months prior to the last follow-up (e.g., cataract surgery, intravitreal injection, re-vitrectomy, silicone oil extraction). At a 4:1 ratio, patients were assigned to either the training cohort (200 eyes) or the validation cohort (50 eyes) randomly.

Standard 23- or 25-gauge PPV was performed under retrobulbar anesthesia. Intravitreal injection of anti-vascular endothelial growth factor (VEGF) drugs was performed 3–5 days prior to surgery in eyes with broad basement adhesion of the fibrovascular proliferative membrane to the retina. The anterior, posterior, and peripheral vitreous body were substantially removed using a high-speed vitrectomy surgical system, and the posterior vitreous cortex was removed with the aid of triamcinolone acetonide. Moreover, neovascular membranes were dissected and/or peeled off. Endo-laser photocoagulation was finished, and the number of laser shots depended on the preoperative laser status and detached area of the retina. Depending on the condition of the retina, tamponade with a balanced salt solution, air, C3F8 gas, or silicone oil was used. Intravitreal injection of anti-VEGF drugs or dexamethasone intravitreal implant was performed, if necessary. Phacoemulsification and intraocular lens implantation were combined for visually significant cataracts. Generally, silicone oil was removed 3–6 months postoperatively when the retina was stable. The surgical procedures were all performed by one surgeon.

We collected data by reviewing the patients’ medical records. As systemic factors, we collected the information on sex, age, diabetes type, diabetes duration, hypertension, dialysis, coronary heart disease, cerebral infarction, and blood test (the details are listed in Table 1 ). The CKD epidemiology collaboration equation was used to calculate the estimated glomerular filtration rate (eGFR) (10). Moreover, the following ophthalmological parameters were collected: symptom duration, best-corrected visual acuity (BCVA) measurement, indication for surgery, history of pan-retinal photocoagulation, and history of anti-VEGF intravitreal injection. The primary outcome was the BCVA measurement at the last follow-up, and the secondary outcomes included the anatomical outcome and occurrence of neovascular glaucoma (NVG) and re-vitrectomy.

Low vision was defined as a BCVA<0.3, according to the standard of the WHO 2015 (8). According to BCVA at the last follow-up, the training cohort was divided into low vision and non-low vision groups for comparative analysis. Regarding the change in BCVA, an increase of ≥0.3 logarithmic minimum angle of resolution (logMAR), a change of<0.3 logMAR, and a decrease of ≥0.3 logMAR were defined as “improvement,” “invariant”, and “worsening”, respectively. For statistical analysis, the decimal visual acuity was transformed to logMAR. Values of 1.85, 2.3, 2.6, and 2.9 logMAR were assigned to counting fingers, hand movement, light perception, and no light perception, respectively (11). We selected 60 mL/min/1.73 m² and 30 mL/min/1.73 m² as the cutoff values for eGFR (12). A value >60 mL/min/1.73 m² (normal eGFR), 30–60 mL/min/1.73 m² (medium eGFR), and<30 mL/min/1.73 m² (low eGFR) indicated borderline/normal renal function, early renal insufficiency, and poor renal function, respectively. For continuous variables, statistical results were displayed as mean ± standard deviation (SD) when they were normally distributed and as the median and interquartile range (IQR) when they were non-normally distributed. Percentage and frequency were used for categorical variables. The statistical analysis of continuous variables was conducted using t-test, Mann-Whitney U test, and analysis of variance, while categorical variables were analyzed using chi-squared test or Fisher’s exact test.

The multivariate binary logistic regression analysis considered variables with a P-value<0.1 from the univariate analysis, and stepwise forward regression was used to explore variables with a P-value<0.05 as potential independent predictors. The variance inflation factor was calculated to test for multicollinearity amongst the independent variables. In accordance with the outcomes of the multivariate logistic regression analysis, a nomogram for estimating the probability of long-term low vision after diabetic vitrectomy was built.

The discrimination capacities of the predictive indicators were described using receiver operating characteristic (ROC) curves. Generally, a value >0.7 of the area under the ROC curve (AUC) is considered to indicate good discrimination. The threshold of the equation was determined by the maximum Youden Index (sensitivity + specificity - 1). The calibration was tested and depicted using the Hosmer-Lemeshow test and calibration curves. The nomogram was internally validated using the bootstrapping approach, and it was externally validated in an external cohort. All statistics were analyzed using IBM SPSS Statistics ver. 25.0 (SPSS, Chicago, IL, USA) and R software ver. 4.0.1 (R Project for Statistical Computing, Vienna, Austria). A two-sided P-value of<0.05 was considered statistically significant.

A total of 491 patients underwent primary vitrectomy for PDR at Tianjin Medical University Eye Hospital from January 2016 to October 2021. Among them, 186 patients (250 eyes) who met the eligibility criteria were included in this study. The mean follow-up duration was 32 ± 21 months. For all eyes, the mean postoperative BCVA was 1.53 ± 0.66 logMAR, and it increased to 0.8 ± 0.85 logMAR at the last follow-up (P<0.001). The total incidence of long-term low vision was 43.6%. The rate of improvement, invariant, and worsening of BCVA was 68.8%, 15.2%, and 16.0%, respectively. The rate of the final BCVA of ≥0.7, 0.3–0.7, 0.1–0.3, and<0.1 was 22%, 34.4%, 18.4%, and 25.2%, respectively. Final anatomical success was achieved in 235 eyes (94%). During the follow-up period, 40 eyes (16%) underwent re-vitrectomy for recurrent VH in 31 eyes and retinal detachment in 9 eyes, and 18 eyes (7.2%) developed NVG.

In the training cohort of 200 eyes, low vision occurred in 89 eyes (44.5%). Table 1 provides a list of the characteristics evaluated in this study. The results indicated that age, diabetes type, hypertension, eGFR, symptom duration, preoperative BCVA measurement, and TRD were significantly related to long-term low vision after diabetic vitrectomy (P<0.05).

Considering the practical clinical application, we stratified three continuous variables: age (≤55 years old, >55 years old), symptom duration (<6 months, ≥6 months), and eGFR (<60 mL/min/1.73 m², ≥60 mL/min/1.73 m²). According to stepwise forward regression analysis, four independent predictors were included in the risk prediction model: TRD (β=1.443, OR=4.235, P<0.001), symptom duration ≥6 months (β=0.954, OR=2.595, P=0.004), preoperative BCVA measurement (β=0.540, OR=1.716, P=0.033), and hypertension (β=0.645, OR=1.905, P=0.044). The result of the logistic regression analysis is presented in Table 2 . The equation of the model is:

where logitP is the linear predictive value and P is the predictive probability.

The variance inflation factor of each predictor in this model was<10, indicating there was no multicollinearity among the independent variables. The cutoff score that maximized the Youden index was 0.495 (sensitivity, 62.9%; specificity, 78.4%).

Then a nomogram was depicted to present the logistic prediction model ( Figure 1 ). The total points were calculated according to the conditions of different patients. More total points indicated a higher probability of occurrence of long-term low vision after diabetic vitrectomy.

The ROC curves for the model and each predictor are depicted in Figure 2 . The discrimination was good, as indicated by the AUC of 0.764 (95% CI: 0.699–0.829). The nomogram also exhibited good calibration on its calibration curve ( Figure 3 ). Moreover, the Hosmer–Lemeshow test confirmed the calibration (c² = 8.799, P=0.360). The AUC of the internal validation was 0.747 using bootstrapping with 1,000 replicates. The external validation cohort included 50 eyes, and long-term low vision occurred in 20 eyes (40%). There were no significant differences in all baseline characteristics between the two cohorts ( Table 3 ). In the external validation cohort, the model exhibited good prediction accuracy according to the AUC of 0.755 (95% CI: 0.619–0.891).

In the training cohort, eGFR differed significantly only in the univariate analysis between the low and non-low vision groups. We further analyzed the features and prognosis of all the patients with borderline/normal renal function (n=184), early renal insufficiency (n=43), and poor renal function (n=23). For baseline characteristics, patients with worse renal function were shown to have significantly longer symptom duration, more hypertension, lower hemoglobin levels, lower hemoglobin A1c levels, and higher fibrinogen levels (P<0.05). There was no significant difference in preoperative median BCVA measurement (normal eGFR group, 1.7 (1.0–2.0) logMAR; median eGFR group, 1.5 (1.0–2.3) logMAR; low eGFR group, 2.3 (0.8–2.3) logMAR; P=0.178). The long-term BCVA significantly improved in the three groups (P<0.01); however, those with lower eGFR experienced considerably worse vision outcomes (normal eGFR group, 0.4 (0.15–1.0) logMAR; median eGFR group, 0.7 (0.4–1.3) logMAR; low eGFR group, 0.8 (0.4–2.3) logMAR; P=0.011). Furthermore, there was no significant difference in BCVA changes among the groups (P=0.068). After adjusting for symptom duration and hypertension, renal function was still an independent predictor of negative visual outcomes in the multilinear regression analysis (P=0.024). The incidence of NVG in the renal insufficiency group (n=66) was substantially higher than that in the group of patients with borderline/normal renal function (13.6% vs. 4.9%, P=0.018). Finally, incidence of re-vitrectomy in the three groups did not differ significantly from one another (P=0.933).