Table 1.
Cost-effectiveness of AI-based DR screening with different model performances
| Scenarios | Sensitivity | Specificity | Cost per person (US$) | Incr. Cost in 251,535 population (million US$) | Effect per person (QALYs) | Incr. Eff (QALYs in 251,535 population) | ICER (US$/QALY) | NMB (million US$) |
|---|---|---|---|---|---|---|---|---|
| 1 (status quo) a | 0.933 | 0.877 | 6214 | - | 9.1689 | - | - | - |
| 2 | 0.882 | 0.903 | 6192 | −5.544 | 9.1630 | −1490 | 3719 | −40 |
| 3 | 0.897 | 0.897 | 6197 | −4.310 | 9.1648 | −1034 | 4169 | −28 |
| 4 | 0.909 | 0.892 | 6201 | −3.257 | 9.1663 | −664 | 4908 | −17 |
| 5 | 0.919 | 0.887 | 6205 | −2.245 | 9.1673 | −396 | 5664 | −10 |
| 6 | 0.925 | 0.884 | 6208 | −1.434 | 9.1680 | −226 | 6341 | −6 |
| 7 | 0.929 | 0.880 | 6211 | −0.720 | 9.1685 | −98 | 7377 | −2 |
| 8 | 0.936 | 0.873 | 6217 | 0.839 | 9.1693 | 99 | 8467 | 2 |
| 9 | 0.944 | 0.863 | 6225 | 2.867 | 9.1702 | 315 | 9101 | 7 |
| 10 | 0.947 | 0.856 | 6231 | 4.228 | 9.1705 | 409 | 10,346 | 8 |
| 11 | 0.951 | 0.847 | 6238 | 6.088 | 9.1709 | 513 | 11,879 | 10 |
| 12 | 0.954 | 0.837 | 6246 | 8.145 | 9.1713 | 606 | 13,435 | 11 |
| 13 | 0.958 | 0.824 | 6257 | 10.713 | 9.1717 | 699 | 15,329 | 11 |
| 14 | 0.963 | 0.804 | 6273 | 14.834 | 9.1722 | 839 | 17,681 | 11 |
AI artificial intelligence, DR diabetic retinopathy, Incr. incremental, QALY quality−adjusted life-year, ICER incremental cost-effectiveness ratio, NMB net monetary benefit, GDP gross domestic product.
aStatus quo was defined as the scenario with theoretically optimal model performance, identified by the cut-off point on the receiver operative curve. ICER was calculated by comparing each intervention scenario with the status quo. Scenario 2-7 were cost-saving, while scenarios 8–14 were cost-effective compared to the status quo. Among all, the minimum sensitivity was found at 88.2% in the most cost-saving scenario, while the minimum specificity was found at 80.4% in the scenarios with the greatest effect. The optimal cost-effective performance was determined with the highest effect at sensitivity of 96.3% and specificity of 80.4%. In the population of Lifeline Express, the prevalence of referable DR was 7.44%, and the willingness-to-pay level was determined as three times per-capita GDP (US$ 30,828). All scenarios were also compared with reference scenarios (the lower-cost non-dominated scenario and no screening) and the ICERs were less than the predefined willingness-to-pay level.