<![CDATA[Background: Screening for retinopathy of prematurity (ROP) in low- and middle-income countries (LMICs) is limited by the shortage of trained examiners. Objective: We assessed the diagnostic accuracy of telemedicine and artificial intelligence (AI) for ROP screening in LMICs and summarised economic evaluations of screening. Methods: PubMed, Europe PMC, OpenAlex and Scopus were searched for accuracy studies from LMICs comparing remote grading of retinal images (telemedicine) or AI with ophthalmoscopy or expert grading, and for economic evaluations. Sensitivity and specificity were pooled with bivariate random-effects models; risk of bias was assessed with QUADAS-2 and certainty with GRADE. Results: Twenty-four accuracy studies from nine countries and 13 economic evaluations were included. For clinically significant ROP, telemedicine (6 datasets; 927 units) had a sensitivity of 92.8% (95% CI 82.6–97.3) and a specificity of 95.8 (87.6–98.7)%; AI (6 datasets; 7,346 units) had a sensitivity of 93.0 (81.5–97.6)% and a specificity of 77.3 (55.6–90.3)%. Specificity was higher in upper-middle-income settings. For any-stage ROP and plus disease, pooled sensitivities were 92.2% and 81.4%, and specificities 97.6% and 99.1%. Fifteen studies were at high risk of bias, and certainty was low to very low. Telemedicine was cost-effective in all five cost-utility analyses; no economic evaluation of AI from LMICs was found. Conclusion: Telemedicine detected clinically significant ROP with high sensitivity and specificity and was cost-effective in LMICs, although the certainty of evidence was low. AI achieved similar sensitivity with more false-positive referrals; its use as a triage step within telemedicine programmes requires prospective evaluation.]]>
