# A risk prediction model for neovascular glaucoma secondary to proliferative diabetic retinopathy based on Boruta feature selection and random forest

## TL;DR

This study creates a model to predict the risk of neovascular glaucoma in patients with proliferative diabetic retinopathy using machine learning techniques.

## Contribution

A novel risk prediction model for neovascular glaucoma using Boruta feature selection and random forest algorithms.

## Key findings

- The model achieved 90.74% accuracy and an AUC of 0.87 in predicting neovascular glaucoma.
- Calibration curves showed reliable predictions within a 0.4–0.8 probability range.
- Decision curve analysis confirmed the model's clinical net benefit across various thresholds.

## Abstract

Neovascular glaucoma (NVG) is one of the most severe complications of proliferative diabetic retinopathy (PDR), carrying a high risk of blindness. Establishing an effective risk prediction model can assist clinicians in early identification of high-risk patients and implementing personalized interventions to reduce the incidence of vision impairment. This study aimed to develop and evaluate a risk prediction model for NVG in PDR patients based on the Boruta feature selection method and random forest algorithm to improve clinical predictive performance.

This retrospective study included 365 PDR patients treated at Shenzhen Eye Hospital between January 2019 and December 2024, comprising 269 controls (non-NVG) and 96 cases (NVG). The Boruta feature selection method was employed to identify key features associated with NVG development in PDR. A risk prediction model was then constructed using the random forest algorithm. Model performance was evaluated based on accuracy, sensitivity, specificity, and area under the curve (AUC). Additionally, calibration curves and decision curve analysis (DCA) were used to assess clinical utility. All data analyses and modeling were performed in R (version 4.2.3).

The Boruta algorithm selected 12 significant predictive features. The random forest-based model achieved an accuracy of 90.74%, sensitivity of 82.14%, specificity of 93.75%, and an AUC of 0.87, demonstrating strong predictive performance. Calibration curves indicated reliable prediction probabilities within the 0.4–0.8 range. Decision curve analysis revealed substantial clinical net benefit across threshold probabilities of 0.2–0.8.

The Boruta-guided random forest model developed in this study exhibits excellent predictive performance and clinical applicability for assessing NVG risk in PDR patients.