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

PURPOSE: Evaluate a deep learning model's performance as a pre-referral filter for referable glaucoma using colour fundus photographs. METHODS: Retrospective cohort study. Patients referred for glaucoma evaluation in 2021 were included. Referred glaucoma suspects underwent testing (visual acuity, colour fundus photography, OCT and visual fields) followed by specialist review. Clinician evaluation was classified as accepted for face-to-face visit (further classified as glaucoma diagnosis and clinically suspicious) or discharged. Hypertensive (IOP ≥ 24 mmHg), on IOP-lowering treatment, or patients who had undergone previous glaucoma surgery were excluded as they require glaucoma follow-up regardless. A deep learning model analysed the colour fundus photographs and a simulated referral was considered if the model's output exceeded the manufacturer's recommended threshold (0.73). The model's referral decision was compared against glaucoma experts' glaucoma diagnosis and treatment; performance was measured. RESULTS: A total of 96 patients were included in the study; the mean age was 66.5 (SD: ±15.5) years; around a third of the included participants were male (n = 33; 34.4%). Of these referred patients, 37 were accepted by the glaucoma department (38.5%); 19 were found to have glaucoma (19.8%). A model-led system would have prevented over a third of referrals (36; 37.5%) while not missing any glaucoma cases. The model presented high sensitivity (1.00; 95% CI 0.82-1.00), and reasonable specificity (0.65; 95% CI 0.53-0.75) and positive predictive value (0.41; 95% CI 0.27-0.57). CONCLUSIONS: Glaucoma rate in ophthalmologist-referred patients was low. A deep learning-led system would have accurately referred all patients with glaucoma and reduced unnecessary observations. Current model performance can act as a filter between referring and receiving specialised healthcare.