Table 2.

Current studies on the validation of artificial intelligence technology.

Author	AI Device/Software	Condition	Dataset Used for Deep Learning	Test Sample	Outcome	Area under Curve (ROC)
Zheng et al. [87]	semi-supervised generative adversarial networks (GANs)	retinal disorders	877 OCT images	107,912 OCT images	semi-supervised GANs performed better than the reference supervised DL model	0.99
Adithya et al. [88]	offline deep learning algorithm (DLA)	vitreoretinal abnormalities (VRA)	4319 ocular ultrasound images	421 ocular ultrasound images	DLA showed high sensitivity detecting retinal detachment (97.4%) and choroidal detachment (100%)	0.939
Liu et al. [89]	deep learning system (DLS) for diabetic macular edema	diabetic retinopathy	4295 OCT images	matched images from the same dataset	DLS had 80% specificity and 81% sensitivity, while experienced graders had 59% specificity and 70% sensitivity	DLS scored 0.88, compared with 0.80 for the reference software
Bai et al. [90]	retinopathy of Prematurity AI (ROP.AI)	plus disease in ROP Plus disease in ROP	unspecified as ROP.AI is a proprietary software	8052 retinal images	84% sensitivity, 43% specificity, and 96% negative predictive value	0.75
Wagner et al. [91]	code-free deep learning-based classifiers (CDFL)	plus disease in ROP Plus disease in ROP	retinal images from 6141 neonates	338 retinal images	CFDL models conferred similar performance to senior pediatric ophthalmologists	0.989
Kemp et al. [92]	Medios AI software (FOP NM-10)	referable diabetic retinopathy (RDR)	unspecified as Medios AI is a proprietary software	2327 retinal images	Medios AI compared favorably with an experienced field grader	0.9648