Retrospective validation of G-ROP, CO-ROP, Alex-ROP, and ROPscore predictive algorithms in two Chinese medical centers

PurposeTo evaluate the sensitivity and specificity of four predictive algorithms (G-ROP, CO-ROP, Alex-ROP, and ROPscore) for retinopathy of prematurity and compare their performances in the Chinese population.MethodsA retrospective study was conducted at two medical centers in China of infants born at Women's Hospital School of Medicine Zhejiang University and Yiwu Maternal and Child Health Hospital. A total of 1,634 infants who met the criteria and who were GA < 32 weeks or BW < 2,000 g according to Chinese guidelines for ROP screening were included. The ROP group was further grouped into severe ROP and mild ROP. The sensitivity and specificity of G-ROP, two simplified G-ROPs, CO-ROP, Alex-ROP, and ROPscore were analyzed.ResultsSevere ROP and any ROP were identified in 25 and 399 of 1,634 infants, respectively. According to the criteria of different models, 844, 1,122, 1,122, and 587 infants were eligible in the G-ROP, CO-ROP, Alex-ROP, and ROPscore, respectively. G-ROP had 96.0% sensitivity and 35.0% specificity for severe ROP. For two simplified G-ROPs (180 g and 200 g models), similar sensitivity was showed with original G-ROP and they had specificity of 21.8% and 14.0%, respectively. The sensitivity and specificity of Co-ROP were 96% and 64.3% for severe ROP, while Alex-ROP only had sensitivity of 56.0% and specificity of 61.4% for severe ROP. ROPscore had a sensitivity of 91.3% and a specificity of 62.4% for severe ROP. In 546 infants who met all 4 models' inclusion criteria and included 23 infants with severe ROP, the validation outcomes showed the sensitivity of G-ROP, ROPscore, CO-ROP, and Alex-ROP for severe ROP was 95.6%, 91.3%, 100%, and 56.0%, and their specificity was 38.0%, 60.8%, 39.9%, and 52.9%, respectively.ConclusionG-ROP, ROPscore, and CO-ROP had high sensitivity for severe ROP in the Chinese population, but both the sensitivity and specificity of Alex-ROP were low. CO-ROP (not high-grade CO-ROP) provided the best performance for severe ROP in a fair comparison. For further application, ROP screening models need to be adjusted by local populations.