# Integration of peripheral blood-based systemic inflammatory indices and retinal imaging using interpretable machine learning for predicting anti-VEGF treatment response in macular edema secondary to retinal vein occlusion

## TL;DR

This study uses machine learning to predict how well patients with retinal vein occlusion-related macular edema will respond to anti-VEGF treatment by combining blood inflammation markers and retinal imaging data.

## Contribution

The novel contribution is an interpretable machine learning framework integrating systemic inflammation and retinal data to predict treatment response and identify systemic inflammation as a modifiable target.

## Key findings

- Key predictors of treatment response include central macular thickness and systemic inflammation indices like NLR and SII.
- Exacerbated systemic inflammation reduces retinal treatment benefits and increases non-response risk.
- A web-based risk calculator was developed to identify non-responders early and guide personalized treatment.

## Abstract

Macular edema secondary to retinal vein occlusion (RVO-ME) demonstrates considerable inter-individual variability in response to anti-VEGF therapy. While current research has predominantly focused on ocular imaging features and intraocular cytokine profiles, the role of systemic inflammation remains underexplored. This study proposes an interpretable machine learning (ML) framework that integrates peripheral blood-based systemic inflammatory indices with retinal imaging data to predict anatomical outcomes following anti-VEGF treatment in RVO-ME, and to elucidate underlying systemic inflammation–retinal structure interactions.

This single-center retrospective study included 202 RVO-ME patients receiving a standardized three-injection anti-VEGF regimen. Clinical data, retinal imaging parameters, peripheral blood cell counts, and derived systemic inflammatory indices were collected. Feature selection used least absolute shrinkage and selection operator (LASSO) and Boruta algorithms. Nine ML models were developed and optimized through Bayesian hyperparameter tuning with five-fold cross-validation for model selection, followed by independent test set validation. SHapley Additive exPlanations (SHAP) and Generalized Additive Models (GAMs) provided interpretation and mechanism exploration. A web-based risk calculator was deployed for clinical translation.

Central macular thickness before the third injection (CMT-2), minimum neutrophil-to-lymphocyte ratio (NLR-min), and minimum systemic immune-inflammation index (SII-min) emerged as key predictors. The Random Forest model performed optimally. SHAP and GAMs revealed that exacerbated systemic inflammation (SII-min and NLR-min) attenuated retinal structural treatment benefit (CMT-2), while concurrent elevations in inflammatory and structural burden markedly increased non-response risk. Counterfactual simulation suggested therapeutic gains from targeting systemic inflammation. The calculator based on the optimal model offers visual decision support for early non-responder identification.

This study identifies systemic inflammation–retinal structure synergy as a key mechanism underlying anti-VEGF treatment heterogeneity in RVO-ME, highlights systemic inflammation as a modifiable therapeutic target, and supports personalized treatment strategies to improve clinical outcomes.