Diabetic macular edema (DME) is the most common vision-threatening complication arising from diabetic retinopathy (DR).Globally, approximately 5.5% of individuals with diabetes are affected by DME (Im et al., 2022).Intravitreal injection of anti-vascular endothelial growth factor (anti-VEGF) agents has become the first-line treatment for DME owing to their potent anti-permeability and anti-angiogenic effects (Ehlers et al., 2022). However, clinical practice reveals significant inter-individual variability in response to anti-VEGF therapy. Approximately 30%–40% of patients continue to exhibit persistent edema or suboptimal visual improvement despite regular treatment (Zhang et al., 2022; Hussain and Ciulla, 2016). These patients often require additional therapeutic interventions such as laser photocoagulation, corticosteroid injections, or even surgical treatment (Hussain and Ciulla, 2016; Hatamnejad et al., 2024). Therefore, early identification of high-risk individuals who are likely to have poor treatment responses and the development of individualized intervention strategies have become critical challenges in current DME management.

In recent years, artificial intelligence (AI) has shown great promise in the management of chronic ocular conditions, including dry eye disease (Li et al., 2024; Wei et al., 2024), myopia (Xiao et al., 2024), and age-related macular degeneration (AMD) (Waldstein et al., 2020; Gong et al., 2024; Wan et al., 2025). In the field of DR, AI applications have primarily focused on early screening and diagnosis. AI systems have demonstrated high efficiency and accuracy in DR screening, enabling automated analysis of fundus images without human intervention to detect referable DR and DME. These systems have achieved sensitivity and specificity comparable to, or even exceeding, those of human graders in certain settings (Bellemo et al., 2019; Farahat et al., 2024; Yuan, 2019; Guo et al., 2023). However, compared to screening and diagnosis, the application of AI in predicting treatment response to anti-VEGF therapy in DME remains in its early stages of exploration.

Based on this, this study aims to develop a predictive model by collecting pre-treatment ocular parameters and blood-based biomarkers as predictor variables, with post-treatment resolution of macular edema as the outcome variable. This model is expected to assist clinicians in identifying potential non-responders at the initial stage of diagnosis and treatment planning, allowing for more targeted and personalized intervention strategies.

To our knowledge, this is the first study to integrate pre-treatment structural OCT features with a comprehensive panel of systemic blood biomarkers to predict response to anti-VEGF therapy in DME patients. This approach extends beyond previous models that relied solely on imaging data by capturing systemic factors that may influence treatment outcomes (Meng et al., 2024; Mao et al., 2022). In addition, by incorporating SHAP analysis and nomogram visualization, our model offers enhanced interpretability for the prediction of treatment response, providing insights into the contribution of each feature and distinguishing our approach from prior black-box machine learning methods.

Anti-VEGF therapy is currently the first-line treatment for DME. However, its efficacy varies widely among individuals. Predicting treatment response is critical for optimizing therapeutic strategies and achieving personalized care in DME management (Rasti et al., 2020). Previous studies have developed machine learning models using imaging biomarkers alone to predict anti-VEGF treatment outcomes in DME (Allingham et al., 2017; Leng et al., 2024). We expanded on these approaches by incorporating systemic metabolic and hematologic markers, which may account for additional variability in treatment response. As a result, our integrated model achieved robust predictive performance while considering both ocular and systemic factors. For context, prior OCT-only models for DME have reported AUC values around 0.76–0.83 (Gallardo et al., 2021; Shah et al., 2017), whereas our multi-modal model achieved a higher AUC of 0.90 in the present study. Notably, an OCT-based deep learning model recently attained an AUC 0.99 on a small dataset (Song et al., 2025); our approach achieves comparable accuracy while offering greater interpretability. Unlike prior models that functioned as black boxes, the interpretability of our model—achieved through SHAP values and a nomogram—provides clinicians with clear insights into individual risk factors, an advantage that facilitates clinical decision-making. Among various algorithms tested, the logistic regression model demonstrated the best predictive performance and interpretability. Our findings identified several key predictors of post-treatment central retinal thickness reduction, including preoperative retinal edema >400 μm, DRIL, subretinal fluid, EZ integrity, and baseline CRT.

In this study, we compared the performance of five commonly used machine learning algorithms for predicting anatomical response to anti-VEGF therapy in patients with diabetic macular edema. Although nonlinear models such as MLP and RF also demonstrated favorable performance, we ultimately selected the logistic regression model as the preferred approach. This decision was based on its superior overall performance in the internal validation cohort (AUC = 0.90), along with consistently high sensitivity, specificity, and F1 score. Moreover, logistic regression offers a simpler and more interpretable structure, allowing for direct extraction of feature coefficients and facilitating the development of user-friendly clinical tools such as nomograms. Given the relatively limited sample size, logistic regression is also less prone to overfitting and tends to exhibit better generalizability. Therefore, we considered logistic regression to be the most appropriate choice in balancing predictive accuracy with clinical applicability in this context.

The association between baseline CRT and the anatomical response to anti-VEGF therapy has been widely established in previous clinical trials. Studies such as RESTORE and RIDE/RISE have demonstrated that eyes with greater initial CRT experience more substantial reductions in macular thickness and greater visual improvements following ranibizumab treatment (Wells et al., 2016; Pieramici et al., 2016). Real-world data similarly support this finding, with patients having CRT >400 µm showing greater anatomical improvement than those with thinner retinas (Mushtaq et al., 2014). Notably, the UK National Institute for Health and Care Excellence (NICE) guidelines recommend CRT >400 µm as a threshold for initiating anti-VEGF therapy, given the superior efficacy of ranibizumab over laser treatment in this subgroup (Amoaku et al., 2020). Our study aligns with this evidence, suggesting that eyes with more severe edema at baseline (CRT >400 µm) are more likely to exhibit pronounced anatomical responses. This may be attributed to increased VEGF levels and severe blood-retinal barrier disruption in these patients, resulting in greater responsiveness to anti-VEGF agents that effectively neutralize VEGF and reduce vascular leakage (Funatsu et al., 2002; Gurung et al., 2023).Nevertheless, baseline CRT alone cannot fully predict treatment outcomes, which also depend on factors such as baseline visual acuity and OCT morphologic features.

DRIL is an OCT biomarker reflecting the loss of distinct retinal inner layer boundaries within the central 1 mm zone. It is considered a negative prognostic indicator of visual function in DME. Sun et al. first reported that greater baseline DRIL extent was associated with poorer visual acuity, and that early changes in DRIL correlated with subsequent visual outcomes (Das et al., 2018; Sun et al., 2014). Further studies have suggested that DRIL reflects chronic edema- or ischemia-related damage to Müller and bipolar cell axons, potentially hindering rapid resolution of edema after therapy (Munk et al., 2022; Midena et al., 2023; Lai et al., 2023). Our model also indicated that severe baseline DRIL was associated with a lower likelihood of central retinal thickness reduction. We hypothesize that DRIL may reflect underlying neurodegeneration and capillary nonperfusion, limiting the capacity for structural recovery even after VEGF suppression.

Subretinal fluid is a distinct OCT feature in DME, and several studies have shown its presence to be associated with better response to anti-VEGF therapy (Sophie et al., 2015; Hwang et al., 2019; Kwon et al., 2021). For instance, subgroup analyses from the RIDE/RISE trials revealed that patients with baseline subretinal fluid had a higher probability of achieving complete macular dryness (CFT ≤250 µm) after ranibizumab treatment (Sophie et al., 2015). Another prospective study similarly reported that eyes with subretinal fluid exhibited greater CRT reductions and better visual outcomes after anti-VEGF therapy (Huang et al., 2022). Our findings are consistent with these reports and suggest that subretinal fluid is a positive predictor of anatomical improvement. The underlying mechanism may relate to the highly VEGF-driven nature of subretinal fluid-DME, where severe outer blood-retinal barrier disruption leads to fluid accumulation in the subretinal space. In such cases, VEGF blockade can promptly reduce leakage and resolve subretinal fluid (Kwon et al., 2021; Sonoda et al., 2013).

Interestingly, our study found that EZ integrity was a negative predictor of central retinal thickness reduction after anti-VEGF treatment. Although many previous studies have identified intact EZ as a favorable prognostic factor for visual acuity recovery (Hsieh et al., 2023), its role in anatomical improvement is less clear and remains underexplored. EZ continuity generally reflects relatively mild, early-stage disease, with less structural disorganization and lower VEGF levels. Therefore, while these eyes have better preserved photoreceptor function and potential for visual improvement, their baseline macular thickness is often closer to normal, leaving limited room for further reduction. In contrast, disrupted EZ often implies chronic, VEGF-driven damage, greater edema, and worse baseline anatomy, which may respond more dramatically to anti-VEGF therapy in terms of CRT reduction. This inverse relationship has been supported by several real-world studies showing that eyes with greater baseline CRT and structural disruption—often including EZ loss—achieve more pronounced CRT reductions, whereas structurally intact eyes exhibit minimal changes post-treatment (Chaturvedi et al., 2025; Saxena et al., 2022). This suggests that intact EZ may reflect a lower disease burden and lower VEGF expression, making edema less responsive to anti-VEGF therapy in terms of structural reversal. In these cases, although CRT reduction is limited, visual function may still be preserved or improved. Therefore, EZ integrity appears to serve as a dual indicator—predicting better functional outcomes but less dramatic anatomical responses.

To enhance clinical applicability, we transformed the final logistic regression equation into a user-friendly nomogram that allows rapid bedside risk stratification. Clinicians can plot the five baseline predictors—pre-treatment CRT >400 μm, pre-operative retinal edema, presence of subretinal fluid, disorganization of the inner retinal layers, and intactness of the ellipsoid zone—assign the corresponding points, and sum them to obtain a Total Points score. This score can then be vertically projected onto the bottom scale to estimate the probability of achieving a ≥20% reduction in central retinal thickness after anti-VEGF therapy. Internal validation demonstrated excellent calibration and discrimination of the nomogram (AUC = 0.90). In practice, patients with a high predicted probability can proceed with standard anti-VEGF regimens; those with an intermediate probability should be monitored more closely and considered for early adjunctive therapy; and those with a low probability may benefit from prompt evaluation of alternative or combined treatments. The transparency and interpretability of this tool provide an intuitive quantitative basis for individualized management of DME. Moreover, explainable AI methods such as SHAP improve model transparency by revealing how each variable contributes to the prediction, thereby enhancing clinical interpretability and decision-making. Pair-wise SHAP-dependence results indicate that the effects of baseline CRT, retinal oedema, sub-retinal fluid, DRIL and EZ integrity act largely additively in our cohort, implying that the logistic model’s behaviour is dominated by main effects. This additivity not only enhances interpretability but also supports the construction of a straightforward main-effect nomogram for clinical use.

This study has several limitations. First, it was conducted at a single center with a limited sample size, which may introduce selection bias. External validation in larger, multicenter cohorts is needed to improve generalizability. Second, the study focused on anatomical response (≥20% CRT reduction) and did not include visual acuity outcomes or neuro electrophysiological monitoring, which limits the comprehensiveness of treatment assessment (Scott et al., 2024; Kremers and Huchzermeyer, 2024). Third, the follow-up period was relatively short, and the model’s ability to predict long-term outcomes remains untested. Finally, while systemic metabolic variables were included, important biological markers such as VEGF and inflammatory cytokines were not measured. Future studies should incorporate these factors to enhance model performance.

In this study, we developed and validated a machine learning model with good predictive value for assessing anatomical response to anti-VEGF therapy in patients with DME. Preoperative retinal edema >400 μm, DRIL, presence of subretinal fluid, and EZ integrity were identified as key factors influencing the likelihood of central retinal thickness reduction following treatment. Future research involving larger, multicenter cohorts and incorporation of more comprehensive biological and clinical markers is warranted to further refine and optimize the model.