Table 2. Summary of currently available literature on health economic analyses of artificial intelligence implementation for diabetic retinopathy in low- and middle-income countries.

AI: artificial intelligence, ARIAS: automated retinal image analysis system, DR: diabetic retinopathy, CEA: cost-effectiveness analysis, CUA: cost-utility analysis, CBA: cost-benefit analysis, CMA: cost-minimization analysis, ICER: incremental cost-effectiveness ratio, ICUR: incremental cost-utility ratio, QALY: quality-adjusted life year, DL: deep learning, HG: human grader, GDP: gross domestic product, LY: life years, THB: Thai baht, USD: United States dollar, R$: Brazilian real, WTP: willingness-to-pay

Reference	HEA/model	Models and Comparators	Primary Outcome Measure	Results	Strengths	Limitations
Huang et al. [23]	CEA/Markov model-based hybrid decision tree	AI screening was compared with ophthalmologist screening, in which fundus images were evaluated by ophthalmologists	ICER	Health system perspective- Relative to no screening, AI screening was more expensive with a cost of $180.19 but more effective with an incremental QALYs of 0.16. The ICER of the AI screening group compared with the no screening group was $1,107.63/QALY gained, less than the threshold of $30,765.09. Societal perspective- AI screening costs less than ophthalmologist screening ($1,683.23 versus $1,775.48). Relative to no screening, the ICER of AI screening was $10,347.12, below the cost-effective threshold $30,765.09.	Applied a more comprehensive system of prognosis after people were diagnosed with diabetes. In the study, health states included DR states, blindness, death, and the stable state after laser treatment, which reflect the natural progression of DR.	The utility values were partly derived from the results in other countries, which might not be exactly consistent with those in China. The ICER was compared with the per capita GDP of the whole country instead of rural China.
Lin et al. [24]	CEA and CUA/decision-analytic Markov model	AI screening was compared to manual-grading	ICER and ICUR	In the manual grading–based telemedicine screening, the total cost was US $3,265.40 with 9.83 years without blindness and 6.753 QALYs. In the AI-based telemedicine screening, the total cost was US $3,182.50, with 9.80 years without blindness and 6.748 QALYs. ICUR was US $15,216.96 per QALY and ICER was US $2,553.39.	Conducted a sensitivity analysis within wide ranges and identified the most influential variables affecting the decision to use AI and manual grading in telemedicine screening.	Mainly based on empirical data from Shanghai; therefore, it cannot be representative of all of China because of the huge regional and medical care differences between urban and rural areas.
Gomez Rossi et al. [25]	CEA/Markov model	AI screening was compared to standard screening of DR undertaken by ophthalmologists	Association of AI with QALYs	The mean cost was R $1,321 for AI and R $1,260.28 for diagnosis without AI. Both strategies yielded a very similar mean utility of 8.4 QALYs. The ICER was US R $-91,760. The acceptability curve showed that standard of care was more likely to be more cost-effective although higher WTP increased the uncertainty about the optimal strategy.	The main strength was its design, which modeled different AI technologies for detecting three different diseases such as melanoma, dental caries, and DR, and compared them against established medical practices.	Limited information available on the research, operation and overhead costs, and payment mechanisms involved in incorporating AI did not allow for generating detailed comparisons.
Fuller et al. [26]	CUA/Markov model	ARIAS-based DR screening was compared to standard, in-office dilated eye examinations	ICUR	A 23.3% reduction in cost (USD) in the ARIAS group compared with the standard practice (P < .001). Comparing the current practice to ARIAS screening, an ICUR of $258,721.81 was calculated, which was well beyond the assessed willingness-to-pay threshold of $100,000.	Results were analyzed over a five-year period.	Model only included direct costs of screening and treatment from the payor’s perspective and did not account for indirect costs to patients.
Srisubat et al. [27]	CUA/decision tree-Markov hybrid model	DL screening was compared to human graders	Total cost incurred by the health care system and the total QALYs gained per patient	From societal and provider perspectives, there was equal effectiveness in LY for HG and DL at 18.53, whereas QALYs were 12.857 and 12.862, respectively. From a societal perspective, DL cost was 163,478.16 THB vs. HG cost was 163,565.04 THB, representing an 87 THB cost difference in favor of the DL strategy. From a provider perspective, DL was found to have a higher incremental cost at 2,195 THB and the ICER was 512,955 THB.	Evaluated cost-savings over a lifetime horizon of patients.	Model did not consider the possibility that DL may flag more ungradable patients than HG in real-world scenarios, and therefore, false negatives were not accounted for.