Deep learning in glaucoma referral: Performance assessment using a real-world setting

Area of Science:

• Ophthalmology
• Artificial Intelligence in Healthcare
• Medical Imaging Analysis

Background:

• Glaucoma diagnosis relies on specialist evaluation, leading to high referral rates.
• Current referral pathways can be inefficient, with a low prevalence of glaucoma in referred patients.
• Developing automated screening tools is crucial for optimizing healthcare resource allocation.

Purpose of the Study:

• To assess the efficacy of a deep learning (DL) model as a pre-referral filter for detecting referable glaucoma.
• To evaluate the model's performance using colour fundus photographs (CFPs) compared to expert clinical judgment.

Main Methods:

• A retrospective cohort study included 96 patients referred for glaucoma evaluation.
• A DL model analyzed CFPs, with a simulated referral triggered if the output exceeded a threshold of 0.73.
• Model predictions were compared against glaucoma diagnoses made by expert clinicians.

Main Results:

• The DL model achieved 100% sensitivity in identifying glaucoma cases.
• A model-led referral system could have reduced overall referrals by 37.5% without missing any glaucoma diagnoses.
• The model demonstrated a specificity of 0.65 and a positive predictive value of 0.41.

Conclusions:

• The deep learning model accurately identified all glaucoma patients, demonstrating its potential to streamline referrals.
• Implementing this AI-driven system could significantly reduce unnecessary specialist consultations.
• The model's performance supports its use as an effective filter in the pre-referral stage for glaucoma care.