Artificial intelligence-based glaucoma screening in primary care: a cross-sectional study and economic viability analysis of an ongoing trial

Abstract

<h2>Summary</h2><h3>Background</h3> Glaucoma is a leading cause of blindness and is often undiagnosed. Owing to novel artificial intelligence (AI) technology, population-wide screening could now be possible. Our aim was to evaluate whether AI-based glaucoma screening is clinically and economically feasible within publicly funded primary care. <h3>Methods</h3> This cross-sectional study and economic analysis was part of the ongoing Glaucoma Portugal Screening Trial (Glaucoma POST; NCT05875090). In this substudy, we used individual participant baseline data from a single primary care screening facility in Lisbon, Portugal. Participants aged 55–65 years, with and without diabetes, were randomly selected from primary care screening registries and invited for glaucoma screening (ie, fundus photography and intraocular pressure) through existing diabetic retinopathy infrastructure. We excluded people who had comorbidities precluding attendance (eg, bilateral blindness or severe mobility limitations). Fundus photographs were analysed with an AI algorithm (MONA-GLC) to generate a glaucoma risk score. Referral to hospital for specialist evaluation was triggered by either a positive AI result or an intraocular pressure of 24 mm Hg or higher. Images were also independently graded by six glaucoma experts. Referred participants underwent visual field testing; glaucoma was diagnosed with adapted Thessaloniki Eye Study criteria. We assessed the feasibility, diagnostic performance, and cost-effectiveness over a 10-year horizon of integrating AI-based glaucoma screening into an existing diabetic retinopathy screening programme within a real-world primary care setting. We compared AI-based screening with standard care (ie, patients referred from primary care to hospital-based general ophthalmologists, with onward referral to glaucoma specialists as appropriate). <h3>Findings</h3> Screening and recruitment was done between March 1 and Dec 31, 2023. Of 1038 invited individuals, 671 (65%) attended screening. After excluding 42 participants who missed visual field testing, 629 were included in the performance and prevalence analyses. Glaucoma was diagnosed in 40 of these 629 participants (6% [95% CI 5–9%]). From the 671 people screened, the AI algorithm referred 66 participants (10%), compared with 118 referrals (18%) through the adjudicated expert assessment. AI-algorithm sensitivity was 78% (95% CI 62–89%) and specificity 95% (93–97%). In a setting in which minimal staff training was required, AI-enhanced glaucoma screening achieved an incremental cost-effectiveness ratio of €1725 per quality-adjusted life-year (76·3% probability of cost-effectiveness at a threshold of €20 000 per quality-adjusted life-year) at 1% prevalence, becoming cost-saving at prevalences of 2% or more. <h3>Interpretation</h3> Our results suggest that AI-enabled screening in primary care can be cost-effective and can reduce unnecessary expert referrals, even though a one-off screening round might not capture every case. AI could enhance efficiency and detection, enable earlier treatment, and prevent avoidable blindness. <h3>Funding</h3> None.

Ähnliche Arbeiten