Association of early anti-VEGF Therapy anatomical response with renal function and OCT biomarkers in diabetic macular edema Authors: Huixiang Chen, Yu Shang, Peizhen Yang, Yanping Zhou, Yuanzhi Yuan Published: 2026 ABSTRACT PurposeTo investigate the predictors of anatomical response in patients with diabetic macular edema (DME) following anti-vascular endothelial growth factor (anti-VEGF) therapy and establish a nomogram model for predicting the probability of anatomical response.MethodsThis study enrolled 200 DME patients treated with anti-VEGF regimen. Based on the reduction rate of central macular thickness (CMT) following treatment, patients were classified into an anatomical weak responder group (CMT reduction < 20%) and an anatomical responder group (CMT reduction ≥ 20%). Baseline clinical data and OCT biomarkers were analyzed with multivariate logistic regression. A nomogram model was constructed by using R software. Bootstrapping was used for model validation, receiver operating characteristic (ROC) curve and calibration curve were used for evaluating the discrimination and calibration of prediction model, and decision analysis curve (DCA) was used for evaluating the practicality of model.ResultsPredictors for anatomical response in DME patients are serum creatinine (Scr), CMT, photoreceptor outer segment length (PROSL), and cystoid macular edema (CME) presence as independent variables. The nomogram prediction model based on the above four predictors had good representativeness (Bootstrap method: precision: 0.820), differentiation [the area under curve (AUC) value: 0.819], and the DCA analysis showed that the prediction model, whose threshold probability was in the range of 0 to 1, had clinical practical value.ConclusionThe anatomical response to anti-VEGF treatment for DME is independently associated with baseline Scr, CMT, PROSL, and the presence of CME.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12886-026-05047-6. Introduction Diabetic macular edema (DME) has become the most common cause of vision loss among individuals with type 2 diabetes [1]. Current clinical guidelines for diabetic retinopathy (DR) [2], recommend intravitreal anti-vascular endothelial growth factor (anti-VEGF) therapy as the first-line treatment for DME. However, a substantial proportion of patients exhibit limited or inadequate response to this intervention. According to findings from the Diabetic Retinopathy Clinical Research (DRCR) Retina Network Protocol T, at least 32% of patients with DME continue to experience persistent macular edema despite receiving no fewer than four regular injections within a 24-week period [3]. This persistence of edema under intensive anti-VEGF regimens in controlled clinical trials suggests that real-world treatment outcomes may be even less favorable. Ineffective or minimally effective anti-VEGF injections not only increase the economic burden and potential adverse events for patients but also place additional strain on healthcare systems. Therefore, the establishment of a prognostic assessment system for anti-VEGF therapy could facilitate optimized treatment planning and provide evidence-based support for precision medicine decision-making. Optical coherence tomography (OCT) is a non-invasive imaging modality widely employed in the diagnosis of DME and the longitudinal assessment of therapeutic response. Central macular thickness (CMT) remains the most commonly utilized OCT biomarker in clinical practice. However, with the advancement of spectral-domain OCT, a growing number of additional biomarkers have been identified, including disorganization of the retinal inner layers (DRIL) and disruption of the external limiting membrane and ellipsoid zone (ELM/EZ) [4]. These OCT biomarkers have increasingly become critical indicators for evaluating anti-VEGF treatment efficacy, monitoring disease progression, and guiding decisions regarding the timing of retreatment. Creatinine is a widely used biomarker for assessing renal filtration function. In general, elevated serum creatinine levels correlate with impaired kidney function. Creatinine is not only associated with renal diseases but also closely linked to diabetic retinopathy (DR). A real-world prospective cohort study [5] demonstrated that higher serum creatinine levels are independent risk factors for DR, vision-threatening DR, and DME. Furthermore, another study [6] reported statistically significant differences in baseline serum creatinine levels between diabetic patients with and without macular edema. These findings collectively indicate a significant association between serum creatinine levels and DME. Nomograms, as intuitive tools for risk prediction, have been widely applied and validated in the medical field, particularly for tumor-prognosis assessment [7]. By integrating diverse prognostic determinants, they provide quantitative estimates of an individual’s probability of a specific clinical event, thereby addressing the need for comprehensive biological and clinical prediction models [7]. In the context of anti-VEGF treatment for DME, early prediction of therapeutic response is critical for optimizing drug selection, adjusting treatment intervals, and modifying of management strategies in a timely manner. Based on prior research, the current study incorporates OCT findings and renal function status to investigate the association between baseline clinical indicators and anatomical response in DME patients undergoing anti-VEGF therapy during a six-month follow-up period. A nomogram model was subsequently developed to predict treatment response, providing a practical tool for clinical application. This model has the potential to identify patients at risk of suboptimal outcomes and to support the adoption of more personalized treatment strategies in the management of DME. Methods Study patients The study was conducted in accordance with the principles of the Declaration of Helsinki and obtained ethical approval from the Ethics Committee of Zhongshan Hospital, Fudan University. A total of 200 patients (200 eyes) diagnosed with DME were enrolled from the Department of Ophthalmology between June 2022 and June 2024. The inclusion criteria were as follows: (1) individuals aged 18 years or older with a confirmed diagnosis of type 2 diabetes mellitus (T2DM) and comorbid DR; (2) baseline best-corrected visual acuity (BCVA) between 20/40 and 20/400 Snellen equivalent, and CMT ≥ 300 μm as measured by spectral-domain optical coherence tomography (SD-OCT) [8, 9]; (3) completion of at least 6 months of follow-up with complete baseline and 6-month follow-up data available; (4) to avoid potential bias arising from inter-eye correlation, only one eye per patient was included in the analysis. If both eyes met the inclusion criteria, the left eye was selected by default in accordance with a standard protocol to ensure consistency and eliminate investigator selection bias. The exclusion criteria were as follows: (1) presence of conditions other than DR that may lead to macular edema; (2) uncontrolled glaucoma or other ocular or systemic diseases that could interfere with visual function; (3) prior intravitreal injection of corticosteroids or anti-VEGF agents within three months before enrollment; history of panretinal photocoagulation (PRP) completed less than three months prior, or the requirement of concurrent retinal laser therapy during the 6-month observation period; (4) cataract surgery, vitreoretinal surgery, or intravitreal corticosteroid administration during the observation period; and (5) suboptimal SD-OCT image quality that prevented accurate qualitative or quantitative evaluation of key outcome measures. Data collection All enrolled patients received a three-loading-dose followed by pro re nata (PRN) treatment regimen. Patient data were retrospectively extracted from the hospital information system, including demographic characteristics and clinical variables such as sex, age, duration of diabetes, history of hypertension, history of coronary heart disease, cumulative number of intravitreal injections, and concomitant medications administered within the preceding 6 months. Laboratory parameters encompassed fasting blood glucose (FBG), glycated hemoglobin (HbA1c), serum creatinine (Scr), and glomerular filtration rate (GFR). Ophthalmic examination data included BCVA at baseline and 6 months (expressed in logarithm of the minimum angle of resolution, logMAR), intraocular pressure, and SD-OCT imaging findings. Measurement of biomarkers on OCT The researchers assessed OCT 512 × 128 macular cube scan images and high-resolution images findings, with evaluation parameters encompassing central macular thickness (CMT), diffuse retinal thickening (DRT), cystoid macular edema (CME), subretinal fluid (SRF), disorganization of retinal inner layers (DRIL), integrity of the external limiting membrane (ELM) and ellipsoid zone (EZ), hyperreflective walls (HRW), vitreomacular interface abnormalities (VMIA), including vitreomacular traction (VMT) and epiretinal membrane (ERM), as well as photoreceptor outer segment length (PROSL). DRIL is defined as the inability to discern the boundaries between the outer plexiform layer, the inner nuclear layer, and the ganglion cell layer–inner plexiform layer complex within a 1 mm diameter area centered on the macular fovea in the central horizontal retinal OCT scan. The integrity of the ELM/EZ is evaluated based on the continuity of the external limiting membrane and ellipsoid zone within this 1 mm macular region. HRW is characterized by the presence of high-reflectance foci localized on the walls of low-reflectance cystic cavities located between retinal layers. PROSL is defined [10] as the distance measured from the inner boundary of the ellipsoid zone band to the inner boundary of the retinal pigment epithelium. Horizontal OCT scans passing through the macular fovea are obtained for each patient. PROSL measurements (in µm) are performed at three specific locations—the foveal center, 500 μm temporal to the fovea, and 500 μm nasal to the fovea—using the built-in caliper tool, and the average value is computed for analytical purposes. Grouping criteria The CMT reduction rate was calculated as (CMT at six months minus baseline CMT) divided by baseline CMT. Based on the established 20% threshold [11, 12], patients in the study were categorized into two groups: the anatomical responder group (CMT reduction rate ≥ 20%) and the anatomical weak responder group (CMT reduction rate < 20%). Statistical analysis All statistical analyses were conducted using SPSS (version 27.0, IBM). In the univariate analysis, categorical variables were described as frequencies and proportions, and differences between groups were evaluated using the chi-square (χ²) test or Fisher’s exact test. The normality of continuous variables was assessed using the Kolmogorov-Smirnov (K-S) test. Normally distributed continuous variables were expressed as mean ± standard deviation (mean ± SD), and group comparisons were performed using the independent samples t-test. Non-normally distributed continuous variables were presented as median (P25, P75), and group comparisons were performed using the Mann-Whitney U test. Considering the potential clinical heterogeneity of baseline renal function, all renal functional indicators (including GFR and Scr) were incorporated into the multivariate logistic regression model as continuous variables rather than categorical variables. This approach fully preserves the original numerical information and granularity of renal function data, avoids information loss and artificial grouping bias caused by categorical staging, and improves the stability and reliability of the predictive model. Variables with a p-value less than 0.1 in the univariate analysis were included as potential predictors in the multivariable logistic regression model to identify independent factors associated with treatment response in patients with DME. A nomogram was developed using the rms package in R software. Internal validation of the nomogram was performed via bootstrapping with 1,000 resamples. Model performance was assessed in terms of discrimination, calibration, and clinical usefulness. Discrimination was assessed by the receiver operating characteristic (ROC) curve and the corresponding area under the curve (AUC); calibration was examined via the calibration plot; and clinical utility was evaluated using decision curve analysis (DCA). A p-value less than 0.05 was considered statistically significant. Results Characteristics of the enrolled patients A total of 200 patients (200 eyes) with DME were enrolled in this study based on predefined inclusion and exclusion criteria, consisting of 110 males and 90 females. The mean age of the participants was 60.53 ± 11.92 years, with a mean duration of diabetes of 12.98 ± 7.92 years. Among the enrolled patients, 114 had a history of hypertension and 23 had coronary heart disease. The baseline CMT was 419.59 ± 111.48 μm, and the baseline BCVA, expressed in logMAR, was 0.55 ± 0.26. The mean baseline HbA1c level was 7.68 ± 1.41%, which was above the normal reference range. Renal function parameters showed a mean GFR of 78.01 ± 29.00 ml/min/1.73 m² and a mean Scr level of 86.41 ± 35.20 µmol/L, both within normal limits. During the 6-month observation period, patients received a mean of 4.63 intravitreal injections, with 131 patients treated with conbercept (Kanghong, China), 22 with ranibizumab (Novartis, Switzerland), and 47 with aflibercept (Bayer, Germany). No intraocular adverse events, including elevated intraocular pressure or endophthalmitis, were reported throughout the study. Comparison of baseline characteristics Patients were categorized into the anatomical weak responder group (n = 111) and the anatomical responder group (n = 89) based on predefined grouping criteria. Baseline demographic and clinical characteristics were compared between the two groups, revealing statistically significant differences in BCVA, GFR, and serum creatinine levels. No significant intergroup differences were observed with respect to gender, age, duration of diabetes, DR subtypes, comorbidities, HbA1c levels, number of intravitreal injections, or types of administered medications (Table 1). Table 1Comparison of baseline demographic and clinical characteristics between the anatomical weak responder group and the anatomical responder groupBaseline characteristicsAnatomical weak responder group (n = 111)Anatomical responder group (n = 89)Statistic p Gender (male)58 (52.3%)52 (58.4%)χ² = 0.7610.383Age (y)61.04 ± 10.6859.91 ± 13.34t = 0.6470.518DM duration (y)10.0 (8.0, 20.0)11.0 (6.0, 20.0)Z = 0.0580.954DR subtypesχ² = 0.2430.622 NPDR76 (68.5%)58 (65.2%) PDR35 (31.5%)31 (34.8%)Hypertension (YES)61 (55.0%)53 (59.6%)χ² = 0.4260.514CHD (YES)11 (9.9%)12 (13.5%)χ² = 0.6200.431Number of injections4.58 ± 0.804.70 ± 0.83t = 1.0340.303Medicationsχ² = 5.5170.066 Conbercept66 (59.5%)65 (73.0%) Ranibizumab12 (10.8%)10 (11.2%) Aflibercept33 (29.7%)14 (15.7%)BCVA (logMAR)0.50 ± 0.250.62 ± 0.26t = 3.286 0.001 Scr (umol/L)85.0 (68.0, 112.0)68.0 (57.0, 85.0)Z = 4.094 < 0.001 GFR (ml/min/1.73m2)75.0 (50.0, 93.0)93.0 (67.5, 107.5)Z = 3.683 < 0.001 HbA1c (%)7.73 ± 1.407.62 ± 1.42t = 0.5100.610Categorical variables are presented as numbers (percentages); normally distributed continuous variables are presented as mean ± standard deviation; non-normally distributed continuous variables are presented as median (P25, P75);Statistically, significant p-values are displayed in boldDM = Diabetes Mellitus; NPDR = non-proliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy; CHD = Coronary Heart Disease; BCVA = best-corrected visual acuity; Scr = serum creatinine; GFR = glomerular filtration rate; HbA1c = glycated hemoglobin Table 1Comparison of baseline demographic and clinical characteristics between the anatomical weak responder group and the anatomical responder groupBaseline characteristicsAnatomical weak responder group (n = 111)Anatomical responder group (n = 89)Statistic p Gender (male)58 (52.3%)52 (58.4%)χ² = 0.7610.383Age (y)61.04 ± 10.6859.91 ± 13.34t = 0.6470.518DM duration (y)10.0 (8.0, 20.0)11.0 (6.0, 20.0)Z = 0.0580.954DR subtypesχ² = 0.2430.622 NPDR76 (68.5%)58 (65.2%) PDR35 (31.5%)31 (34.8%)Hypertension (YES)61 (55.0%)53 (59.6%)χ² = 0.4260.514CHD (YES)11 (9.9%)12 (13.5%)χ² = 0.6200.431Number of injections4.58 ± 0.804.70 ± 0.83t = 1.0340.303Medicationsχ² = 5.5170.066 Conbercept66 (59.5%)65 (73.0%) Ranibizumab12 (10.8%)10 (11.2%) Aflibercept33 (29.7%)14 (15.7%)BCVA (logMAR)0.50 ± 0.250.62 ± 0.26t = 3.286 0.001 Scr (umol/L)85.0 (68.0, 112.0)68.0 (57.0, 85.0)Z = 4.094 < 0.001 GFR (ml/min/1.73m2)75.0 (50.0, 93.0)93.0 (67.5, 107.5)Z = 3.683 < 0.001 HbA1c (%)7.73 ± 1.407.62 ± 1.42t = 0.5100.610Categorical variables are presented as numbers (percentages); normally distributed continuous variables are presented as mean ± standard deviation; non-normally distributed continuous variables are presented as median (P25, P75);Statistically, significant p-values are displayed in boldDM = Diabetes Mellitus; NPDR = non-proliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy; CHD = Coronary Heart Disease; BCVA = best-corrected visual acuity; Scr = serum creatinine; GFR = glomerular filtration rate; HbA1c = glycated hemoglobin Comparison of baseline demographic and clinical characteristics between the anatomical weak responder group and the anatomical responder group Categorical variables are presented as numbers (percentages); normally distributed continuous variables are presented as mean ± standard deviation; non-normally distributed continuous variables are presented as median (P25, P75); Statistically, significant p-values are displayed in bold DM = Diabetes Mellitus; NPDR = non-proliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy; CHD = Coronary Heart Disease; BCVA = best-corrected visual acuity; Scr = serum creatinine; GFR = glomerular filtration rate; HbA1c = glycated hemoglobin Comparative analysis of baseline OCT biomarkers demonstrated that the mean CMT was 383.09 ± 108.07 μm in the weak responder group and 465.11 ± 98.63 μm in the responder group, with significantly lower values in the former (p < 0.001). The baseline PROSL was 37.85 ± 3.06 μm and 37.06 ± 3.22 μm, respectively, showing no statistically significant difference (p = 0.078). Furthermore, the presence of CME and SRF differed significantly between groups (p < 0.001 and p = 0.009, respectively), with higher proportions observed in the responder group. No significant differences were observed between the two groups regarding the baseline presence of DIRL, disruption of EZ/ELM integrity, DRT, HRW, or VMAI (p = 0.067, p = 0.732, p = 0.887, p = 0.495, and p = 0.115, respectively) (Table 2). Representative longitudinal changes in retinal morphology and fundus appearance for an anatomical responder and an anatomical weak responder are illustrated in Supplementary Figure S1 and Supplementary Figure S2, respectively. Table 2Comparison of OCT biomarkers between the anatomical weak responder group and the anatomical responder groupOCT biomarkersAnatomical weak responder group (n = 111)Anatomical responder group (n = 89)Statistic p CMT(µm)383.09 ± 108.07465.11 ± 98.63t = 5.544 < 0.001 PROSL(µm)37.85 ± 3.0637.06 ± 3.22t = 1.7740.078DRIL (YES)37 (33.3%)41 (46.1%)χ² = 3.3670.067EZ/ELM disruption (YES)30 (27.0%)26 (29.2%)χ² = 0.1170.732DRT (YES)107 (96.4%)87 (97.8%)χ² = 0.0200.887CME (YES)57 (51.4%)70 (78.7%)χ² = 15.884 < 0.001 SRF (YES)27 (24.3%)37 (41.6%)χ² = 6.754 0.009 HRW (YES)16 (14.4%)16 (18.0%)χ² = 0.4670.495VMIA (YES)20 (18.0%)9 (10.1%)χ² = 2.4900.115Categorical variables are presented as numbers (percentages); Continuous variables are presented as mean ± standard deviation; Statistically, significant p-values are displayed in boldCMT = central macular thickness; PROSL = photoreceptor outer segment length; DRIL = disorganization of retinal inner layers; EZ/ELM = external limiting membrane and ellipsoid zone; DRT = diffuse retinal thickening; CME = cystoid macular edema; SRF = subretinal fluid; HRW = hyperreflective walls; VMIA = vitreomacular interface abnormalities Table 2Comparison of OCT biomarkers between the anatomical weak responder group and the anatomical responder groupOCT biomarkersAnatomical weak responder group (n = 111)Anatomical responder group (n = 89)Statistic p CMT(µm)383.09 ± 108.07465.11 ± 98.63t = 5.544 < 0.001 PROSL(µm)37.85 ± 3.0637.06 ± 3.22t = 1.7740.078DRIL (YES)37 (33.3%)41 (46.1%)χ² = 3.3670.067EZ/ELM disruption (YES)30 (27.0%)26 (29.2%)χ² = 0.1170.732DRT (YES)107 (96.4%)87 (97.8%)χ² = 0.0200.887CME (YES)57 (51.4%)70 (78.7%)χ² = 15.884 < 0.001 SRF (YES)27 (24.3%)37 (41.6%)χ² = 6.754 0.009 HRW (YES)16 (14.4%)16 (18.0%)χ² = 0.4670.495VMIA (YES)20 (18.0%)9 (10.1%)χ² = 2.4900.115Categorical variables are presented as numbers (percentages); Continuous variables are presented as mean ± standard deviation; Statistically, significant p-values are displayed in boldCMT = central macular thickness; PROSL = photoreceptor outer segment length; DRIL = disorganization of retinal inner layers; EZ/ELM = external limiting membrane and ellipsoid zone; DRT = diffuse retinal thickening; CME = cystoid macular edema; SRF = subretinal fluid; HRW = hyperreflective walls; VMIA = vitreomacular interface abnormalities Comparison of OCT biomarkers between the anatomical weak responder group and the anatomical responder group Categorical variables are presented as numbers (percentages); Continuous variables are presented as mean ± standard deviation; Statistically, significant p-values are displayed in bold CMT = central macular thickness; PROSL = photoreceptor outer segment length; DRIL = disorganization of retinal inner layers; EZ/ELM = external limiting membrane and ellipsoid zone; DRT = diffuse retinal thickening; CME = cystoid macular edema; SRF = subretinal fluid; HRW = hyperreflective walls; VMIA = vitreomacular interface abnormalities Univariate and multivariable analysis Based on intergroup differences between the anatomical responder group and the weak anatomical responder group, variables with p < 0.1—namely scr, GFR, BCVA, CMT, PROSL, DRIL, CME, SRF, and types of administered medications—were included as covariates in multivariable logistic regression analysis. Model fitting was performed using the forward selection method, yielding a final model that included scr, CMT, PROSL, and CME as significant predictors. The coefficient of determination (R²), Akaike information criterion (AIC), and Bayesian information criterion (BIC) for the model selected as Model 1 were 0.371, 219.841, and 236.333, respectively (Table 3). The univariate analysis of this study demonstrated a statistically significant difference in the presence of SRF between the two groups (p = 0.009), consistent with the findings reported by Szeto et al. [13]. Consequently, SRF was selected over CME as a variable in the development of Model 2 (Table 4). The R2, AIC, and BIC values of the two models were calculated to assess model goodness of fit (Table 5). Model 1 was determined to be the final model for anatomical response, with scr, CMT, PROSL, and CME identified as the independent predictors included in the final model. To test model robustness, a sensitivity analysis was performed within the largest subgroup treated with conbercept (n = 131). The subgroup multivariable logistic regression confirmed that all four primary predictors—Scr, CMT, PROSL, and CME—retained their independent statistical significance (all p < 0.05) with highly consistent odds ratios and directional trends. This demonstrates that the predictive performance of our core model is stable and not substantially confounded by the variation in anti-VEGF agents (Supplementary Table 1). Table 3Multivariable logistic regression analysis of anatomical response (Prediction Model 1)Variablesb valuestandard errorWald p OROR 95%CIscr-0.0340.00721.705 < 0.001 0.9670.953–0.981CMT0.0070.00213.972 < 0.001 1.0071.003–1.011PROSL-0.1920.0658.844 0.003 0.8250.727–0.937CME (YES)0.9870.3816.703 0.01 2.6841.271–5.667Statistically, significant p-values are displayed in boldScr = serum creatinine; CMT = central macular thickness; PROSL = photoreceptor outer segment length; CME = cystoid macular edema; OR = odds ratio Table 3Multivariable logistic regression analysis of anatomical response (Prediction Model 1)Variablesb valuestandard errorWald p OROR 95%CIscr-0.0340.00721.705 < 0.001 0.9670.953–0.981CMT0.0070.00213.972 < 0.001 1.0071.003–1.011PROSL-0.1920.0658.844 0.003 0.8250.727–0.937CME (YES)0.9870.3816.703 0.01 2.6841.271–5.667Statistically, significant p-values are displayed in boldScr = serum creatinine; CMT = central macular thickness; PROSL = photoreceptor outer segment length; CME = cystoid macular edema; OR = odds ratio Multivariable logistic regression analysis of anatomical response (Prediction Model 1) Statistically, significant p-values are displayed in bold Scr = serum creatinine; CMT = central macular thickness; PROSL = photoreceptor outer segment length; CME = cystoid macular edema; OR = odds ratio Table 4Multivariable logistic regression analysis of anatomical response (Prediction Model 2)Variablesb valuestandard errorWald p OROR 95%CIscr-0.0320.00720.517 < 0.001 0.9690.955–0.982CMT0.0090.00220.317 < 0.001 1.0091.005–1.012PROSL-0.1730.0627.82 0.005 0.8410.745–0.949SRF (YES)-0.0030.370.00010.9930.9970.483–2.057Statistically, significant p-values are displayed in boldScr = serum creatinine; CMT = central macular thickness; PROSL = photoreceptor outer segment length; SRF = subretinal fluid; OR = odds ratio Table 4Multivariable logistic regression analysis of anatomical response (Prediction Model 2)Variablesb valuestandard errorWald p OROR 95%CIscr-0.0320.00720.517 < 0.001 0.9690.955–0.982CMT0.0090.00220.317 < 0.001 1.0091.005–1.012PROSL-0.1730.0627.82 0.005 0.8410.745–0.949SRF (YES)-0.0030.370.00010.9930.9970.483–2.057Statistically, significant p-values are displayed in boldScr = serum creatinine; CMT = central macular thickness; PROSL = photoreceptor outer segment length; SRF = subretinal fluid; OR = odds ratio Multivariable logistic regression analysis of anatomical response (Prediction Model 2) Statistically, significant p-values are displayed in bold Scr = serum creatinine; CMT = central macular thickness; PROSL = photoreceptor outer segment length; SRF = subretinal fluid; OR = odds ratio Table 5Likelihood ratio test results for each anatomical response modelAICBIC R 2 Model 1219.841236.3330.371Model 2226.805243.2960.372AIC = Akaike Information Criterion; BIC = Bayesian Information Criterion; R2 = Coefficient of determination; Table 5Likelihood ratio test results for each anatomical response modelAICBIC R 2 Model 1219.841236.3330.371Model 2226.805243.2960.372AIC = Akaike Information Criterion; BIC = Bayesian Information Criterion; R2 = Coefficient of determination; Likelihood ratio test results for each anatomical response model AIC = Akaike Information Criterion; BIC = Bayesian Information Criterion; R2 = Coefficient of determination; Nomogram development A nomogram was developed based on multivariable logistic regression analysis of the selected variables associated with anatomical response, with the predicted outcome representing the probability of achieving such a response. The model results were visualized in Fig. 1. In clinical application, each predictor variable is assigned a corresponding point value on the upper scale, reflecting its individual contribution to the overall score. The sum of these scores is then used to determine the corresponding probability of anatomical response for a given patient. A higher total score indicates a greater likelihood of achieving an anatomical response. Fig. 1Presents the nomogram of the anatomical response prediction model. The figure consists of seven rows: the first row indicates the scoring scale; the second through fifth rows represent individual variables; the sixth and seventh rows correspond to the total score and predicted probability, respectively. In clinical application, each predictor variable is assigned a corresponding point value on the upper scale, reflecting its individual contribution to the overall score. The sum of these scores is then used to determine the corresponding probability of anatomical response for a given patient Presents the nomogram of the anatomical response prediction model. The figure consists of seven rows: the first row indicates the scoring scale; the second through fifth rows represent individual variables; the sixth and seventh rows correspond to the total score and predicted probability, respectively. In clinical application, each predictor variable is assigned a corresponding point value on the upper scale, reflecting its individual contribution to the overall score. The sum of these scores is then used to determine the corresponding probability of anatomical response for a given patient Nomogram model test The ROC curve analysis demonstrated an area under the curve of 0.819 (95% CI: 0.761–0.876) (Fig. 2A). Internal validation was subsequently conducted using the Bootstrap method with 1,000 resampling iterations, resulting in a concordance index (C-index) of 0.820, indicating robust discriminative ability of the model. The calibration curve showed close alignment between predicted and observed probabilities, with negligible deviation from the ideal curve, thereby supporting the model’s accuracy in estimating the likelihood of anatomical response to anti-VEGF therapy in patients with DME (Fig. 2B). To further evaluate clinical applicability, decision curve analysis (DCA) was performed. The results revealed that, across the full range of clinically relevant threshold probabilities (0 to 1), the DCA curve for the proposed prediction model consistently surpassed both the “treat all” and “treat none” strategies, underscoring its potential utility in real-world clinical decision-making (Fig. 3). Fig. 2A The curve represents the ROC curve of the anatomical response prediction model. The diagonal line represents the reference line (AUC = 0.5, no diagnostic value). The area under the ROC curve (AUC) of the model was 0.819 (95% CI: 0.761–0.876); B displays the calibration curve of the same model. The Apparent Curve indicates the model’s initial predicted probability, the Bias-corrected Curve reflects the calibrated probability derived from bootstrap resampling, and the Ideal Curve represents perfect prediction agreement between observed and predicted probabilities. The calibration curve showed close alignment between predicted and observed probabilities, with negligible deviation from the ideal curve A The curve represents the ROC curve of the anatomical response prediction model. The diagonal line represents the reference line (AUC = 0.5, no diagnostic value). The area under the ROC curve (AUC) of the model was 0.819 (95% CI: 0.761–0.876); B displays the calibration curve of the same model. The Apparent Curve indicates the model’s initial predicted probability, the Bias-corrected Curve reflects the calibrated probability derived from bootstrap resampling, and the Ideal Curve represents perfect prediction agreement between observed and predicted probabilities. The calibration curve showed close alignment between predicted and observed probabilities, with negligible deviation from the ideal curve Fig. 3The red curve represents the decision curve analysis (DCA) of the prediction model. The black diagonal line indicates the reference line corresponding to the net benefit when all patients receive treatment, whereas the black horizontal line represents the reference line for the net benefit when no patient receives treatment. When the red curve lies above both reference lines, it suggests that using the prediction model for clinical decision-making within the associated threshold probability range yields a higher net benefit compared to either treating all or no patients The red curve represents the decision curve analysis (DCA) of the prediction model. The black diagonal line indicates the reference line corresponding to the net benefit when all patients receive treatment, whereas the black horizontal line represents the reference line for the net benefit when no patient receives treatment. When the red curve lies above both reference lines, it suggests that using the prediction model for clinical decision-making within the associated threshold probability range yields a higher net benefit compared to either treating all or no patients Discussion DME is one of the leading causes of vision impairment in individuals with diabetes mellitus [14]. Intravitreal anti-VEGF therapy has emerged as the standard of care for DME involving the central fovea. However, clinical evidence indicates considerable heterogeneity in treatment response across individuals. Approximately 25% to 40% of patients exhibit suboptimal responses or persistent macular edema despite adherence to standard treatment protocols [3, 15–17]. This variability may be influenced by multiple factors, including injection frequency, intraocular concentrations of inflammatory cytokines, baseline VEGF levels, morphological characteristics of edema, presence of mechanical vitreomacular traction, and structural integrity of the photoreceptor layer [3, 15–17]. Early identification of individual patient responses to anti-VEGF therapy may allow for timely adjustments in clinical management, such as transitioning to intravitreal corticosteroid injections or adopting a more intensive treatment regimen [12]. Therefore, the identification of predictive risk factors for therapeutic outcomes is critical to supporting evidence-based clinical decision-making. Previous studies have demonstrated a significant association between serum creatinine levels and the development and progression of DME [5, 6]. Multivariable logistic regression analysis indicates that elevated baseline creatinine levels are independently associated with a diminished anatomical response to anti-VEGF therapy (OR = 0.967, 95% CI: 0.953–0.981; p < 0.001). This result is consistent with the clinical findings of Warid et al. [18], who observed that patients with higher baseline serum creatinine levels experienced less improvement in visual acuity and retinal thickness following bevacizumab treatment for refractory DME. However, their study employed only univariate analyses, whereas the current investigation adjusts for multiple confounding variables through multivariable modeling, thereby strengthening the validity of the observed associations. From a pathophysiological perspective, renal dysfunction may compromise therapeutic outcomes through two potential pathways: first, a reduced glomerular filtration rate impairs the clearance of VEGF, resulting in increased systemic VEGF concentrations [19]; second, the accumulation of uremic toxins contributes to further breakdown of the blood-retinal barrier. Clinical evidence [20] indicates that in patients with renal impairment undergoing hemodialysis, central foveal thickness decreased markedly by 305 μm (from 473 μm to 168 μm), with a concurrent three-line improvement in BCVA. These observations suggest that improving renal function may act synergistically to enhance the efficacy of anti-VEGF treatment. Notably, our results revealed an inconsistency where baseline renal markers (Scr and GFR) differed significantly between the two groups, whereas chronological diabetes duration did not. While individual variations in long-term glycemic control, the phenomenon of “metabolic memory“ [21], or the well-documented discrepancy between true asymptomatic disease onset and formal clinical diagnosis in type 2 diabetes [22] may explain why functional renal impairment does not strictly parallel recorded diabetes duration, an exhaustive investigation into this chronological divergence falls outside the primary scope of this study. The current research focuses strictly on evaluating baseline functional parameters as immediate, direct predictors of early anti-VEGF responsiveness in DME. Currently, DME is characterized by three distinct morphological patterns: DRT, CME, and SRF. Multivariable logistic regression analysis revealed that the presence of CME at baseline was an independent predictor of anatomical response (OR = 2.684, 95% CI: 1.271–5.667, p = 0.010). This finding is consistent with the study by Roh et al. [23], which reported a mean improvement in BCVA of 5.75 ETDRS letters greater in CME patients compared to the non-CME group following bevacizumab treatment, as well as a significant reduction in CMT (223.71 ± 224.52 μm; p = 0.005). Comparable results were also documented by Wu et al. [24], further supporting the association between baseline CME and improved therapeutic outcomes. In the present study, SRF demonstrated statistical significance only in univariate analysis (p = 0.009). Previous studies [25] have shown that eyes with CME exhibit a higher density of microaneurysms in the perifoveal retinal capillary network than those with DRT or SRF, suggesting greater vascular permeability around the fovea and may enhance uptake of anti-VEGF agents, thereby contributing to the more favorable treatment response observed in CME patients. A cross-sectional study by Deák et al. [26] demonstrated that cysts in the outer nuclear layer exceeding 100 μm in diameter, as well as cystic lesions in the inner nuclear layer, are associated with decreased retinal sensitivity. Prior research [27] suggests that these cystic lesions are primarily attributable to extensive cytoplasmic swelling and necrosis of Müller cells, along with concomitant neuronal degeneration. Early intervention during the early stages of cyst formation may therefore promote resolution of Müller cell edema and facilitate timely recovery of retinal function. Although the concept of PROSL has been introduced for some time, we found limited research investigating its association with the prognosis of anti-VEGF treatment. Ozkaya et al. [28] demonstrated a progressive reduction in mean PROSL across patient groups, ranging from healthy individuals to those with diabetes mellitus (DM) without DR, followed by patients with DR, and further to individuals with DME. This trend decline suggests that PROSL may serve as an indicator of DR severity. The pathophysiological basis for PROSL shortening likely involves atrophic changes in the outer retina secondary to central ischemia and impaired perfusion [29]. In the study, univariate analysis did not reveal a statistically significant association between baseline PROSL and anatomical response (p = 0.078). However, given the established relationship between PROSL and DR progression and its pathogenesis, along with potential confounding factors such as age and gender, the true relationship may have been masked in the analysis. Consequently, PROSL was incorporated into a multivariable model, and the adjusted analysis confirmed a significant independent association, indicating PROSL as a predictive factor for treatment outcomes. Moreover, emerging evidence [30] suggests that PROSL may be a more robust prognostic marker than CMT, as macular edema may resolve with timely intervention, whereas recovery of photoreceptor structural integrity and length often continues beyond the resolution of fluid. In contrast, Sardana et al. [31] evaluated 12-week anti-VEGF therapy in DME patients and reported a negative correlation between baseline PROSL and 12-week CMT (p < 0.05). An analogous trend was observed in the overall cohort of this study; however, the absence of adjustment for key confounders like baseline CMT may compromise the validity of direct comparisons, highlighting the necessity of multivariable modeling to assess the independent contribution of PROSL to treatment prognosis. A substantial body of research [4, 9, 13] has explored the relationship between biomarkers and treatment prognosis; however, previous studies have largely remained at the level of establishing statistical correlations, resulting in limited translational impact in clinical practice. In addition, heterogeneity in study design—particularly with respect to inclusion and exclusion criteria, therapeutic regimens, and timing of outcome evaluation—has contributed to inconsistent findings across the literature. Moreover, there remains a notable absence of visualized prediction tools capable of clearly conveying the strength and implications of these associations. Campos et al. [32] utilized a multivariable logistic regression model to identify short-term prognostic indicators in patients with DME, examining 122 affected eyes and assessing 11 clinical parameters. In comparison, both the present study and Campos’s investigation included eight common variables: gender, duration of diabetes, HbA1c, history of hypertension, CMT, BCVA, EZ integrity, and presence of SRF. However, several important distinctions are evident. First, our study incorporates systemic factors by accounting for diabetic microvascular complications, specifically including serum creatinine levels as a covariate. Second, the current investigation explicitly accounts for CME and DRIL—both of which have been increasingly recognized as key morphological determinants of treatment outcomes. Finally, our study introduces a visual representation of the developed logistic regression model, thereby enhancing its utility for clinical decision support. The current study systematically analyzed patients’ baseline data to characterize demographic profiles, renal function parameters, and OCT biomarkers associated with anatomical response. The results established a robust foundation for the development of a predictive model. Multiple logistic regression analysis was employed to identify four independent predictors of anatomical response: serum creatinine level, central macular thickness, photoreceptor outer segment length, and the presence of cystoid macular edema. And a nomogram was developed to predict the anatomical response in DME patients receiving anti-VEGF therapy. Subsequently, model validation demonstrated strong performance, with a C-index of 0.820 indicating good discriminative ability. The calibration curve showed close agreement with the ideal curve, while ROC curve analysis and decision curve analysis demonstrated favorable predictive accuracy and clinical utility. This nomogram may serve as a practical tool for clinicians to identify patients likely to exhibit favorable or poor anatomical outcomes early in treatment, thereby facilitating timely clinical decisions and enabling more informed discussions regarding prognosis and therapeutic planning at the initiation of therapy. The primary strengths of the current study are as follows: First, it elucidates the association between serum creatinine levels and response to anti-VEGF therapy, and further confirms the independent impact of creatinine on treatment outcomes through a multivariable regression model that adjusts for baseline demographic and clinical parameters as well as OCT biomarkers. Second, we constructed and validated a novel, visual predictive model of anatomical response following anti-VEGF therapy in DME patients. We acknowledge the following limitations of the study. First, the model was developed using retrospective data from a single center, which may introduce selection bias due to hospital-specific patient admission practices. The inclusion criteria required baseline OCT and renal function test results, as well as high-quality OCT images, which may exclude patients with financial constraints, poor medical adherence, or ocular opacity (e.g. cataract, vitreous hemorrhage). Second, the relatively small sample size precluded an independent validation cohort; instead, the model was constructed using a single modeling group and internally validated via bootstrap resampling, limiting its generalizability to external populations. Third, the main outcome was response at six months post-treatment initiation, which provided insufficient insight into long-term outcomes. Fourth, baseline urinary albumin-to-creatinine ratio (ACR) data were unavailable due to the retrospective nature of this ophthalmic study. Consequently, we could not fully apply the complete KDIGO dual-classification framework that combines both GFR and ACR. Future prospective studies should systematically include baseline ACR metrics to enable a more rigorous and comprehensive assessment of the renal-retinal axis in DME. Fifth, our reliance on a fixed baseline CMT threshold of ≥ 300 μm—while serving as a simplified, pragmatic cutoff in line with retrospective literature—does not account for subtle variations in normal retinal thickness across different genders and ethnic groups. This might potentially limit the direct generalizability of our predictive model to other specific ethnic cohorts with varying physiological baselines until further validated in diverse, multi-ethnic populations. Conclusion The current study demonstrated that the anatomical response to anti-VEGF therapy in DME patients was independently associated with baseline serum creatinine levels, central macular thickness, outer segment length of photoreceptors, and the presence of cystoid macular edema. A nomogram-based predictive model was subsequently developed to estimate the likelihood of anatomical response using baseline clinical parameters, thereby supporting more informed decision-making in clinical practice. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1