Deep learning-based identification of eyes at risk for glaucoma surgery

Area of Science:

• Ophthalmology
• Medical Artificial Intelligence
• Glaucoma Research

Background:

• Glaucoma management requires timely surgical intervention for uncontrolled cases.
• Predicting the need for glaucoma surgery is crucial for effective patient management.
• Current methods may not fully leverage multimodal data for precise surgical risk assessment.

Purpose of the Study:

• To develop and evaluate a deep learning model (DLM) for predicting the risk of surgical intervention in eyes with uncontrolled glaucoma.
• To assess the DLM's performance across various time horizons using multimodal ophthalmology data.
• To identify key predictive features for glaucoma surgery using explainability techniques.

Main Methods:

• A retrospective study of 4898 eyes from adult glaucoma patients was conducted.
• A deep learning model was constructed using visual field (VF), optical coherence tomography (OCT), and clinical data.
• Model performance was evaluated using area under the receiver operating characteristic curve (AUC) and precision-recall curve (PRC), with SHAP for feature importance.

Main Results:

• The DLM achieved an AUC of 0.92 for predicting surgery within 3 months.
• Clinically useful AUC values (0.8) were obtained for predictions up to 3 years.
• Intraocular pressure (IOP), Mean Deviation (MD), and Retinal Nerve Fiber Layer (RNFL) thickness were identified as critical predictors.

Conclusions:

• Deep learning models can effectively identify eyes at high risk for glaucoma surgery within defined timeframes.
• Predictive accuracy decreases with longer forecasting horizons.
• Clinical implementation of these DLMs can aid in referring patients for timely surgical evaluation by glaucoma specialists.