The Cost-Effectiveness of an AI-Based Population Wide Screening Programme of Primary Open-Angle Glaucoma in the Netherlands

OBJECTIVES: To assess the cost-effectiveness of AI-based screening of fundus photos for primary open-angle glaucoma (POAG) in the Dutch population aged 50 to 75 years.

METHODS: Traditional population-wide glaucoma screening is generally not cost-effective in developed countries; however it is hypothesized that AI-based screening could significantly reduce screening costs. AI-based screening starts with a screening invitation from the Dutch National Institute of Public Health to the target population to visit a primary care lab for a fundus photo, that is subsequently analyzed by AI. Positive cases are referred to ophthalmologists for further diagnosis and treatment in case of POAG. The current standard of care relies on opportunistic case finding. A health economic model, compromising a decision tree and Markov model, was developed to simulate screening, glaucoma progression, and treatment. Key inputs were AI screening sensitivity (85%) and specificity (95%), screening compliance (50%), and transition probabilities between glaucoma severity stages. Data were sourced from a longitudinal cohort study from the Dutch city of Rotterdam, literature, and expert estimates.

RESULTS: AI-based screening could detect glaucoma earlier than the current standard, reducing visual impairment by 0.76 months per screened individual in the general population. The base-case ICER from a societal perspective was €17.147 (€21.395 from a healthcare perspective), suggesting cost-effectiveness at a €20.000 threshold. However, the low prevalence leads to a high number needed to test to find one case (639), resulting in small QALY gains and thereby high sensitivity to certain parameters. Scenarios with different transition probabilities resulted in ICERs ranging from €9.379 to €71.074.

CONCLUSIONS: AI-based glaucoma screening in the Netherlands could be a cost-effective strategy to improve early detection, thereby reducing the burden of visual impairment. However, the results are sensitive to various parameters, especially the transition probabilities. A small-scale implementation of this screening model could help provide further insight into its (cost)-effectiveness.