JMIR Publications

ABSTRACT

Background:

Previous research has shown that artificial intelligence (AI) can lower unit cost and increase productivity of community-based diabetic retinopathy (DR) screening. However, evidence from low-middle income countries is lacking ,which may bias the decisions on the adoption of screening technology.

Objective:

To test whether AI-assisted DR screening is more worthwhile than manual grading models in the context of lower labor costs.

Methods:

We conducted the cost-effectiveness and cost-utility analysis with decision-analytic Markov models from the societal perspective to compare the actual cos, effectiveness and utility of two scenarios in telemedicine screening for DR: manual grading and AI-based assessment. Sensitivity analyses were done to gauge robustness of the results. Real world data was obtained from Shanghai Digital Eye Disease Screening program. The main outcomes were the incremental cost-effectiveness (ICER) ratio and incremental cost-utility ratio (ICUR). The thresholds of ICUR was set as one and three times the gross domestic product (GDP) per capita.

Results:

Under the status quo, the total expected costs for a 65-year-old resident were $3182.5 and $3265.4,while the total expected years of blindness avoided were 9.80 years and 9.83 years and the utilities were 6.86 QALYs and 6.87 QALYs in AI-assisted model and in manual grading model, respectively. The ICER was $2553.4 per year of blindness avoided and ICUR was $15217.0 per QALY, which indicated the AI-assisted model was not cost-effectiveness. Within the range of parameters included in the sensitivity analysis, with the increase of the compliance with referral increased after the adoption of AI, the increase of on-site screening costs in manual grading, and the decrease of on-site screening costs in AI-assisted model, the AI-assisted model could be the dominant strategy.

Conclusions:

This study fills an evidence gap in the field of AI-assisted diagnostic technology. It indicates that in the context of low labor costs in low and middle income countries (LMICs), AI-assisted DR screening system may not be more worthwhile than manual grading, since would decrease the long-term screening effectiveness and residents’ health utility but could not save enough costs. It suggests that the health economic value of AI based disease screening in the developing countries needs to be critically examined, since the labor cost settings are essentially different from those in developed countries.