AI-(EM)POWERED SCREENING: GLAUCOMA POST STUDY

Abstract

Purpose : Describe the setup of an AI-powered glaucoma screening pilot in a primary care setting.

Methods : Population-based study started in March 2023. Residents between 55 and 65 years from Lisbon North Area registered at their GP-center (ACES Lisboa Norte) were randomly invited for screening. Volunteers would undergo disc-centred fundus photography and intraocular pressure (IOP) measurement. An AI algorithm (MONA GLC, MONA.health, Belgium) evaluated the need for referral to ophthalmologist for glaucoma evaluation using the fundus image. Subjects were referred to the glaucoma clinic if either IOP 24mmHg or AI output 0.73. Clinical data and demographics were prospectively collected. An overall 72 item form was created per patient, which included reasons for refusal, family history, literacy, and socioeconomics. Internal circuit audit of the first 6 months was performed. Clinicaltrial.gov:NCT05875090.

Results : In this preliminary cohort, 402 out of 409 subjects (98.2%) accepted the invitation. Reasons for refusal included: followed by his/her ophthalmologist (n=4), did not value risk (n=3). Recruited patients were 52,2% male and 95.2% Caucasian. More than half had elementary-level education and ten (2.5%) had a family history of glaucoma. Six (1.5%) had a known diagnosis of glaucoma. Mean IOP and AI output scores were 14.1mmHg (±3.8); and 0.56 (±0.12), respectively. Thirty-seven patients (9.2%) were referred, of which 32 (7.9%) due to high AI score. All but one referred subjected agreed to hospital visit. Perimetry analysis showed 4.5% (18/402) of patients showing glaucomatous-related visual field damage, with mean MD -4.5dB [IQR: -6.0/-1.5], and 14 eyes showing at least moderate (MD<-6dB) defects. AI-positive eyes had a mean ppRNFL of 87mm (±17.3) and BMO of 238mm (±43.5). Twenty-eight subjects (77.7%) of the fully evaluated subjects were scheduled for follow-up, 7 (19.4%) were returned to the GP-centered pathway due to low risk, 2 excluded from screening due to wrong referral (n=1) and significant ocular comorbidity (AION, n=1) and 1 sent to Medical Retina due to diabetic retinopathy.

Conclusions : AI-powered glaucoma screening demonstrated great acceptability. It has detected a slightly higher than expected prevalence of glaucoma, with a significant number of moderate-to-advance cases. The current setup seems to show little attrition rate, thus suggesting the possibility for a sustained circuit between primary care setting and Hospital-based clinics.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.