Diabetic retinopathy (DR) remains a leading cause of irreversible vision loss globally, affecting 36.0% of the 537 million diabetes patients worldwide as of 2021, with projections indicating a 51% increase by 2045 (1). Severe non-proliferative diabetic retinopathy (NPDR) constitutes approximately 18.3% of DR cases and carries a 15-40% annual risk of progression to sight-threatening proliferative disease without timely intervention (2). The Early Treatment Diabetic Retinopathy Study (ETDRS) established the efficacy of laser photocoagulation for high-risk PDR over three decades ago (3), but contemporary management has evolved toward combination therapies integrating intravitreal anti-vascular endothelial growth factor (VEGF) agents with targeted laser, demonstrating superior anatomical and functional outcomes in pivotal trials such as CLARITY and Protocol W (4, 5).

Despite improved therapeutic options, substantial interindividual variability persists in treatment response. Recent analyses indicate 30-45% of severe NPDR patients exhibit suboptimal outcomes following anti-VEGF-laser combination therapy, manifesting as disease progression, persistent retinal pathology, or need for unplanned interventions (6, 7). This heterogeneity stems from multifactorial determinants including ischemic burden, glycemic dynamics, and systemic comorbidities (8), yet current clinical decision-making relies predominantly on qualitative assessments without validated predictive tools. Existing DR progression models focus either on natural history (e.g., WESDR risk calculator) or anti-VEGF monotherapy in diabetic macular edema (9, 10), leaving a critical gap for personalized prognostication in combination-treated severe NPDR.

The absence of predictive biomarkers specifically for anti-VEGF-laser synergy represents a significant clinical barrier. While optical coherence tomography angiography (OCTA) metrics show promise in research settings (11), their limited accessibility in routine practice necessitates development of pragmatic models integrating readily available parameters. Nomograms—visual calculators translating multivariate regression into point-based risk scores—have demonstrated utility in oncology and ophthalmology (12, 13) but remain unexplored for combination therapy response in NPDR. Furthermore, emerging evidence suggests systemic factors like glycemic variability and nutritional status modulate therapeutic efficacy (14), warranting incorporation into comprehensive prediction frameworks.

To address these gaps, this study aimed to develop and validate a clinically deployable nomogram that integrates multimodal predictors (demographic, metabolic, ocular, and treatment-related) to quantify individualized response probabilities for anti-VEGF-laser combination therapy in severe NPDR patients. Using rigorously curated retrospective data from a high-volume tertiary center, we specifically sought to: (1) identify key determinants of therapeutic success through machine learning-enhanced variable selection; (2) construct a multivariate prediction model with calibration and discrimination metrics meeting TRIPOD standards; and (3) translate the model into an intuitive nomogram for point-of-care risk stratification. This tool is designed to optimize patient selection, guide treatment intensification, and facilitate personalized management in this vision-threatening condition.

This study establishes the first clinically applicable nomogram for predicting treatment response to combined anti-VEGF and laser therapy in severe NPDR patients. Our model identifies four key predictors—HbA1c variability, FA non-perfusion area, DRSS severity level, and serum albumin—that collectively provide robust discrimination (AUC 0.821 derivation, 0.754 validation) and calibration. The nomogram addresses a critical unmet need in diabetic retinopathy management by enabling personalized prognosis estimation prior to treatment initiation. These findings advance our understanding of therapeutic response heterogeneity in this high-risk population and provide actionable insights for clinical decision-making.

HbA1c variability emerged as the strongest predictor of poor treatment response (adjusted OR 0.63 per 5% increase), surpassing traditional glycemic metrics like baseline HbA1c. This aligns with emerging evidence that glucose fluctuations induce greater oxidative stress and endothelial damage than chronic hyperglycemia alone (20). Our results extend recent findings by Mao et al., who demonstrated HbA1c variability predicts microvascular complications independent of mean HbA1c (21). Mechanistically, oscillating glucose levels exacerbate retinal inflammation through nuclear factor-kappa B (NF-κB) activation and VEGF upregulation (22), potentially diminishing anti-VEGF efficacy. NF-κB activation plays a pivotal role in the regulation of inflammation in retinal cells, where it induces the expression of various inflammatory cytokines, including IL-1β and TNF-α, that contribute to retinal damage and microvascular complications. Furthermore, the VEGF pathway is strongly influenced by NF-κB signaling, which upregulates VEGF expression, promoting angiogenesis and exacerbating retinal neovascularization, a hallmark of diabetic retinopathy. Recent research shows that inhibition of NF-κB can reduce VEGF levels and mitigate inflammatory responses, providing a therapeutic strategy for retinal diseases (23, 24). Additionally, NF-κB activation in retinal glial cells has been shown to enhance oxidative stress and cell migration, leading to further inflammation and vascular changes, which are critical in the progression of diabetic retinopathy (25). This biological plausibility supports our observation that each 5% increase in HbA1c CV reduced predicted response probability by 0.15 in the nomogram. These data underscore the importance of glycemic stability monitoring alongside conventional HbA1c targets in diabetic retinopathy management.

The significant association between FA non-perfusion area and treatment resistance (aOR 0.68 per 10% increase) highlights retinal ischemia as a key therapeutic determinant. Our findings corroborate optical coherence tomography angiography (OCTA) studies by Xu et al., who identified capillary nonperfusion as a predictor of anti-VEGF requirement in diabetic retinopathy (26). However, our use of widefield fluorescein angiography (FA) provides superior peripheral ischemia assessment, particularly relevant in severe non-proliferative diabetic retinopathy where peripheral pathology predominates (27). In our cohort, the mean FA non-perfusion area was 37.4 ± 12.8%, with substantial variability reflecting the heterogeneous nature of retinal ischemia in diabetic eye disease. The clinical implications of this finding are substantial for treatment planning, as widefield FA assessment can be integrated into pre-treatment stratification protocols to identify patients with extensive non-perfusion who may benefit from treatment intensification beyond standard combination therapy (28). Our regression analysis shows that each 10% increase in non-perfusion area decreased response likelihood by approximately 12%, suggesting that patients with more extensive ischemia might benefit from either more frequent anti-VEGF dosing or earlier supplemental panretinal photocoagulation targeting ischemic quadrants. While we acknowledge inherent limitations in FA-based ischemia quantification including variable contrast timing affecting perfusion appearance, projection artifacts in areas of hemorrhage, and inter-grader variability (29), we implemented standardized acquisition protocols and semi-automated quantification with manual verification to enhance reliability. The pathophysiological basis likely involves ischemia-driven upregulation of non-vascular endothelial growth factor angiogenic factors (platelet-derived growth factor [PlGF], angiopoietin-2 [Ang-2]) that bypass VEGF blockade (30). This explains why patients with greater non-perfusion area had lower response rates in our cohort, suggesting such cases may benefit from early supplemental therapies like panretinal photocoagulation or next-generation dual-pathway inhibitors.

Baseline DRSS severity independently predicted diminished treatment efficacy (aOR 0.72 per grade increase), with ETDRS level 53 patients showing 40% lower response probability than level 43 counterparts. This observation challenges the assumption that combination therapy uniformly benefits all severe NPDR subtypes. Pathologically, higher DRSS grades reflect advanced neurovascular unit degeneration and capillary dropout, creating a microenvironment resistant to vascular stabilization (31). Our data align with Protocol W subgroup analyses where baseline DRSS modified aflibercept efficacy in NPDR (32), but we uniquely quantify this relationship for combination therapy. Importantly, the nomogram assigns differential weights to DRSS levels, enabling precise risk stratification unavailable in current binary classifications.

Serum albumin represents a novel systemic predictor of treatment response (aOR 1.85 per 0.5 g/dL increase), a finding with significant clinical implications. Hypoalbuminemia reflects chronic inflammation and endothelial dysfunction, both implicated in diabetic retinopathy progression (33). Mechanistically, albumin maintains endothelial integrity by scavenging reactive oxygen species and inhibiting VEGF expression (34). Our study demonstrates its predictive value specifically in NPDR combination therapy. The 17% absolute probability gain per 0.5 g/dL increase suggests nutritional optimization and inflammation control may synergize with ocular treatments—an easily modifiable factor overlooked in current guidelines.

Our nomogram’s performance (AUC >0.75 in both cohorts) compares favorably with existing diabetic retinopathy prediction tools. The DRCR.net Protocol T model for anti-VEGF monotherapy in DME achieved AUC 0.68 (35), while machine learning algorithms using OCT biomarkers reached AUC 0.79-0.85 in research settings (36). The practical advantage of our model lies in its clinically accessible predictors, avoiding specialized imaging or genetic testing required by alternatives like other model (37). The robust validation performance (calibration slope 0.95, Hosmer-Lemeshow P=0.661) further supports real-world applicability. Importantly, decision curve analysis confirms clinical utility across therapeutically relevant thresholds, with net benefit peaking at 20% probability where clinical uncertainty is greatest.

The nomogram enables several practice-changing applications. First, it identifies high-probability responders (>75%) who may require less intensive monitoring. Second, it flags low-probability responders (<40%) who might benefit from treatment augmentation—such as earlier transition to faricimab (VEGF-angiopoietin-2 [Ang]2 inhibitor) in high-ischemia cases (38), or continuous glucose monitoring (CGM)-guided glycemic control for high-variability patients (39). Third, it provides objective criteria for clinical trial enrichment, potentially accelerating therapeutic development. The point-based scoring system facilitates rapid bedside calculation, addressing the implementation gap seen in complex machine learning models.

Several limitations warrant consideration. First, the retrospective design introduces potential selection bias, though temporal validation mitigates this concern. Second, FA analysis had 15% missing data despite standardized protocols. Future iterations could incorporate optical coherence tomography angiography (OCTA)-based nonperfusion metrics that correlate strongly with FA (26). Third, the model doesn’t account for specific laser parameters, though our analysis found no significant association. Fifth, our assessment of HbA1c variability was constrained by the retrospective nature of the data, with potential inconsistencies in the number and timing of available measurements. While we required a minimum of four measurements over the previous year for CV calculation, the irregular intervals between measurements may have affected our estimation of glycemic variability. Future prospective studies with standardized, equally-spaced HbA1c sampling protocols would strengthen the reliability of this predictor. Finally, we didn’t analyze molecular biomarkers like inflammatory cytokines that might enhance precision.

This study establishes a validated nomogram integrating glycemic variability, retinal ischemia, disease severity, and systemic nutrition to predict anti-VEGF-laser combination therapy response in severe NPDR. The model’s robust performance and clinical accessibility address a critical need for personalized treatment strategies in this sight-threatening condition. Future work should focus on prospective validation, integration with OCTA biomarkers, and assessment of nomogram-guided treatment algorithms.