AI Prediction of Treatment Response in DME

Overview

Diabetic macular edema (DME) is a leading cause of vision loss among individuals with diabetes mellitus. Although intravitreal anti-vascular endothelial growth factor (anti-VEGF) therapy is the standard treatment for center-involved DME, treatment response varies considerably between patients. Recent advances in artificial intelligence (AI), including machine learning (ML) and deep learning (DL), have enabled automated analysis of retinal imaging biomarkers to predict anatomical and functional treatment outcomes. This study aims to systematically evaluate published evidence regarding AI-based prediction models and imaging biomarkers used to predict treatment response in patients with DME. The review will assess the predictive performance of AI models, identify the most important imaging biomarkers, compare different AI approaches and imaging modalities, and summarize methodological strengths, limitations, and research gaps to support future development of precision ophthalmology.

Diabetic macular edema is one of the most common causes of visual impairment in patients with diabetic retinopathy. Despite the widespread use of intravitreal anti-VEGF agents, corticosteroids, laser photocoagulation, and combination therapies, individual responses remain highly variable. Early identification of patients who are likely to respond to specific treatments may improve visual outcomes while reducing unnecessary treatment burden. Advances in retinal imaging, particularly optical coherence tomography (OCT) and optical coherence tomography angiography (OCTA), have enabled detailed characterization of retinal structural and vascular biomarkers associated with treatment outcomes. Commonly investigated biomarkers include central retinal thickness, intraretinal cysts, subretinal fluid, hyperreflective retinal foci, disorganization of the retinal inner layers (DRIL), ellipsoid zone integrity, external limiting membrane integrity, choroidal thickness, choroidal vascularity index, retinal fluid volume, vascular density, and foveal avascular zone parameters. Artificial intelligence techniques, including conventional machine learning and deep learning algorithms, increasingly integrate retinal imaging features with demographic and clinical variables to predict functional and anatomical responses to therapy. Reported prediction models have demonstrated promising diagnostic performance, although considerable variability exists regarding imaging modalities, model architecture, validation strategies, outcome definitions, and reporting standards. This systematic review will comprehensively evaluate published studies investigating AI-based prediction of treatment response in patients with diabetic macular edema. The review will identify imaging biomarkers contributing to predictive performance, compare machine learning and deep learning approaches, evaluate different retinal imaging modalities, summarize reported model performance metrics including area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1-score, and assess methodological quality using validated risk-of-bias assessment tools. Where sufficient homogeneous data are available, a random-effects meta-analysis will be conducted to quantitatively synthesize model performance. The findings are expected to identify robust imaging biomarkers, highlight current limitations of AI prediction models, and provide recommendations for future research and clinical implementation of AI-assisted precision medicine in diabetic macular edema.