Glaucoma is major cause of irreversible blindness and is characterized by optic nerve damage and visual field loss. Screening for glaucoma is challenging due to lack of a simple, accurate, cost-efficient and standardized process. Artificial intelligence, (AI) especially deep learning (DL) algorithms have potential to automate glaucoma detection, but have to be evaluated in real world settings, before public deployment. This study aims to evaluate the screening accuracy of a DL algorithm for glaucoma detection using colour fundus photographs (CFP) in a pragmatic randomised control trial (RCT). The algorithm will be tested in 1040 eligible patients with diabetes, recruited from the Diabetes \& Metabolism Centre's clinics under the Singapore Integrated Diabetic Retinopathy Program (SiDRP) and randomized to 2 arms: AI-assisted model vs current standard of care (grader assessment). The performance of both arms will be compared to performance of study ophthalmologist in diagnosing glaucoma. We hypothesize that the DL model has better screening performance in detecting glaucoma in the community, compared to the current practice method.
Background: Glaucoma is the leading cause of irreversible blindness worldwide, characterized by optic nerve damage and visual field loss. Screening for glaucoma remains challenging due to lack of a simple, standardized, and cost-effective test. Artificial intelligence (AI), especially deep learning (DL), offers potential to improve and standardize glaucoma detection. However, its performance must be prospectively validated in real-world settings before public deployment. Aim: To evaluate the accuracy and cost-effectiveness of a DL algorithm using colour fundus photographs (CFP) as a clinical decision support tool for glaucoma detection in a real-world setting. Methods: A two-centre, single-blind, pragmatic randomized controlled trial (RCT) will be conducted among 1,040 adults with diabetes recruited from the Diabetes \& Metabolism Centre (DMC) and SingHealth Polyclinics-Bukit Merah under the Singapore Integrated Diabetic Retinopathy Programme (SiDRP). After fundus imaging, participants will be randomized 1:1 to AI-assisted grading or current manual grading by graders at the SiDRP reading center (520 subjects per arm). Diagnostic performance will be compared against the gold-standard glaucoma diagnosis, determined via comprehensive ocular examination including intraocular pressure measurement, visual field testing, optical coherence tomography, and dilated fundus assessment. Cost-effectiveness will be evaluated using a cohort-based Markov model to estimate lifetime costs and incremental cost-effectiveness ratios (ICERs) of the two glaucoma screening strategies. Clinical Significance: Integrating AI into glaucoma screening can address resource constraints and streamline detection. This study will provide real-world evidence on the accuracy and cost-effectiveness of AI-based screening. If validated, it could be integrated into national screening programs to enhance early detection, reduce unnecessary referrals, and prevent avoidable blindness through a cost-efficient, scalable approach.
Study Type
INTERVENTIONAL
Allocation
RANDOMIZED
Purpose
DIAGNOSTIC
Masking
SINGLE
Enrollment
1,040
A Vision Transformer model to detect glaucoma from fundus photos
Control group with current practice model by human graders
Singapore National Eye Centre
Singapore, Singapore, Singapore
RECRUITINGEvaluation of model performance
To compare the model performance in accuracy, sensitivity, specificity, positive predictive value and negative predictive value between the new AI-assisted clinical model and the current practice model in detecting glaucoma, with reference to the expert panel's standards.
Time frame: At study completion (after all fundus images have been graded and data collection is finalized; approximately within 12 months of study initiation)
Evaluation of time efficiency
To compare time efficiency between the AI-assisted clinical model and the current practice model, defined as the total time (in seconds) taken per participant for the entire screening process, from image access to final grading decision, recorded in real time during the grading session.
Time frame: At study completion (after all fundus images have been graded and data collection is finalized; approximately within 12 months of study initiation)
Evaluation of Grader's Acceptance
To assess graders' acceptance and satisfaction with the AI-assisted clinical model compared to the current practice model in detecting glaucoma. Assessment will be conducted through brief in-task prompts during the grading process and through a structured post-study questionnaire.
Time frame: At study completion (after all fundus images have been graded and data collection is finalized; approximately within 12 months of study initiation)
This platform is for informational purposes only and does not constitute medical advice. Always consult a qualified healthcare professional.