Presentation (CTI)

OBJECTIVES: This study aims to evaluate human-AI collaboration strategies in diabetic retinopathy (DR) screening and identify the most cost-effective strategy in clinical practice.
METHODS: From a healthcare provider’s perspective, we developed a hybrid decision tree/Markov model to simulate DR screening pathways and disease progression with a hypothetical Chinese cohort of 100,000 individuals aged 18-79 years followed up over a lifespan (up to 80 years). Compared with manual screening (status quo), we assessed the costs and effectiveness of eight possible human-AI collaborative strategies. The nine screening strategies were assessed in five different age groups, with six different DR screening intervals, resulting in 270 different screening scenarios. Main outcome measures included incremental cost-effectiveness ratio (ICER) and cost per blindness year adverted. We used 3-time GDP/capita (US$ 38,052 in 2023) as the willingness-to-pay thresholds in China.
RESULTS: Annual ‘AI·M+M2’ screening in the 20-79 age group, where AI and humans performed independent grading with disagreement reviewed by a secondary grader for a final decision, was the most cost-effective among all the 270 scenarios. Compared with manual screening, this strategy could add 426 blindness-free years and 146 QALYs, with an additional investment of US$ 0.9m (ICER=US$ 6194/QALY; cost/blindness year adverted=$2116/year) in the simulated population. The health benefits of this strategy are equivalent to US$ 4.64m per 100,000 individuals, and this strategy remained optimal in both rural and urban Chinese settings. The ‘AI+M2[Se]’ screening strategy, which involved human graders reviewing certain negative cases flagged as high risk by AI, served as a cost-saving alternative. In this case, if human experts achieve a screening sensitivity of 97.9%, ‘AI+M2[Se]’ may become the most cost-effective screening strategy.
CONCLUSIONS: The ‘AI·M+M2’ screening is the most cost-effective human-AI collaborative strategy for DR screening in China. Human involvement remains essential and cost-effective in a setting where AI is highly mature and efficient.