Abstract
Purpose :
Studies have shown that artificial intelligence (AI) algorithms can help screen for treatment-requiring retinopathy of prematurity (ROP), though it is unknown how cost-effective this is versus standard methods. This study evaluated the cost-effectiveness of autonomous and assistive AI-based ROP screening compared to telemedicine and ophthalmoscopic screening over a range of probabilities, costs, and outcomes.
Methods :
Decision trees created and analyzed in TreeAge Pro modeled outcomes and costs of four possible ROP screening strategies: ophthalmoscopy, telemedicine, assistive AI with telemedicine review, and autonomous AI with only positive screens reviewed. We assumed similar sensitivity for detection of severe ROP with a wide sensitivity analysis, but a higher specificity for ophthalmoscopy. Screening and treatment costs were based on Current Procedural Terminology codes, and opportunity costs to physicians were modeled. AI cost was assumed to be $30. Outcomes were based on the Early Treatment of ROP study, defined as timely treatment, late treatment, or correctly untreated. Incremental cost-effectiveness ratios were calculated at a willingness-to-pay threshold of $100,000. One- and two-way sensitivity analyses were performed comparing AI strategies to telemedicine and ophthalmoscopy, as was a probabilistic sensitivity analysis.
Results :
Autonomous AI was as effective and less costly than each other screening modality (Table 1). Cost of AI evaluation was the most important factor in the sensitivity analysis. AI-based ROP screening was cost-effective up to $17 for assistive and $43 for autonomous screening compared to telemedicine, and $51 and $73 compared to ophthalmoscopy. In the probabilistic sensitivity analysis, both AI screening modalities were cost-effective in over half of trials in all but one comparison (Figure 1).
Conclusions :
We demonstrate that AI-based screening strategies may be more cost-effective than traditional screening modalities across a range of parameters, and cost-effectiveness depends significantly on what cost is assigned to AI.
This is a 2021 ARVO Annual Meeting abstract.