Construction of a machine learning-guided prediction model for the efficacy of anti-VEGF treatment in diabetic macular edema

Citation  

Cui Haoqiang, Xiao Kunhong, Lu Wenrui,  et al. Construction of a machine learning-guided prediction model for the efficacy of anti-VEGF treatment in diabetic macular edema[J]. Chin J Exp Ophthalmol, 2025, 43(11):1024-1034. DOI: 10.3760/cma.j.cn115989-20250608-00186.

Objective  To establish machine learning models to predict visual improvement and anatomical response after anti-vascular endothelial growth factor (VEGF) treatment in patients with diabetic macular edema (DME).

Methods  A multi-algorithm machine learning predictive modeling study based on retrospective clinical data was conducted.A total of 225 patients with DME who received their first intravitreal anti-VEGF injection at Fuzhou University Affiliated Provincial Hospital were enrolled between January 2023 and April 2025.According to data completeness, 204 cases were included in the visual recovery prediction model and 201 cases were included in the anatomical response prediction model.Baseline data included optical coherence tomography (OCT) features and blood biomarkers.The primary outcomes were defined as an improvement of ≥1 line in visual acuity and a reduction of ≥20% in central retinal thickness (CRT) after anti-VEGF treatment.Feature selection was performed using univariate logistic regression and Lasso regression.Four machine learning algorithms, logistic regression (LR), decision tree, multilayer perceptron, and random forest, were trained and validated.Model performance was evaluated using accuracy, sensitivity, specificity, area under the receiver operating characteristic curve (AUC), and decision curve analysis.The best-performing model was further interpreted using SHAP analysis, and a nomogram was constructed for clinical application.This study adhered to the Declaration of Helsinki.The study protocol was approved by the Ethics Committee of Fuzhou University Affiliated Provincial Hospital (No.K2025-03-064).Written informed consent was obtained from each subject.

Results  Among the 37 baseline variables, five key predictors were identified for the outcome of ≥20% CRT reduction: baseline CRT, baseline CRT ≥400 μm, presence of subretinal fluid (SRF), disorganization of the retinal inner layers (DRIL), and integrity of the ellipsoid zone (EZ).Among the four models, the LR model had the best performance, with an accuracy of 0.88, sensitivity of 0.94, specificity of 0.70, and an AUC of 0.94 (95% confidence interval [ CI]: 0.87-1.00).SHAP analysis showed that baseline CRT ≥400 μm, DRIL, SRF, and baseline CRT contributed positively to the outcome, while EZ integrity was a negative predictor for CRT reduction.For the outcome of ≥1-line visual improvement, two key predictors were identified: baseline best corrected visual acuity (BCVA) and EZ integrity.Both baseline BCVA and EZ integrity were negative predictors for ≥1-line visual improvement.The LR model also had the best performance in the internal validation cohort, with an accuracy of 0.71, sensitivity of 0.67, specificity of 0.75, and an AUC of 0.76 (95% confidence interval [ CI]: 0.61-0.91).A visual nomogram was developed based on the selected predictors and the best-performing model.By converting patient-specific clinical characteristics into scores, clinicians can calculate a total score and estimate the probability of achieving a reduction of ≥20% in CRT and a ≥1-line improvement in visual acuity after anti-VEGF therapy.

Conclusions  Machine learning-based model building can effectively predict visual and anatomical response following anti-VEGF treatment in DME patients.Logistic regression shows robust predictive performance for both outcomes.Identification of key predictors, especially OCT features such as EZ integrity, SRF, and DRIL, may aid in guiding treatment expectation assessment and personalized intervention strategies.Nomogram constructed in this study shows good clinical applicability and may serve as a decision-support tool to improve the precision of DME management.

KEYWORDS：

Diabetic macular edema; Vascular endothelial growth factor; Predictive model; Nomogram; Diabetic retinopathy