Zhou Ashley, Li Zhuolin, Paul William, Burlina Philippe, Mocharla Rohita, Joshi Neil, Gu Sophie, Nanegrungsunk Onnisa, Bressler Susan, Cai Cindy X, Alvin Liu T Y, Moini Hadi, Sepehrband Farshid, Bressler Neil M, Kong Jun
Department of Ophthalmology, Johns Hopkins University School of Medicine, Baltimore, Maryland.
Department of Ophthalmology, Mayo Clinic, Rochester, Minnesota.
JAMA Netw Open. 2025 Jan 2;8(1):e2453770. doi: 10.1001/jamanetworkopen.2024.53770.
Determining spectacle-corrected visual acuity (VA) is essential when managing many ophthalmic diseases. If artificial intelligence (AI) evaluations of macular images estimated this VA from a fundus image, AI might provide spectacle-corrected VA without technician costs, reduce visit time, or facilitate home monitoring of VA from fundus images obtained outside of the clinic.
To estimate spectacle-corrected VA measured on a standard eye chart among patients with diabetic macular edema (DME) in clinical practice settings using previously validated AI algorithms evaluating best-corrected VA from fundus photographs in eyes with DME.
DESIGN, SETTING, AND PARTICIPANTS: Retrospective cross-sectional evaluation of deidentified fundus photographs matched to spectacle-corrected VA determined by technicians on eye charts among patients with a history of DME based on optical coherence tomography and at least 2 visits within 1 to 6 months of each other at a university-based clinic between January 2014 and December 2022. Data were analyzed from January 2023 to October 2024.
Previously validated AI algorithm evaluation of fundus photographs.
AI-determined VA mean absolute error (MAE) compared with actual spectacle-corrected VA.
Among 141 patients, the mean (SD) age was 63 (13) years, 71 (50%) were male, 2 (1%) were Asian, 42 (30%) were Black or African American, and 88 (63%) were White. Among 282 eyes at visit 1, 66 had nonproliferative diabetic retinopathy (NPDR) and DME, 38 had proliferative diabetic retinopathy (PDR) and DME, 101 had NPDR and no DME, and 77 had PDR and no DME. Among 564 images (282 eyes) at both initial and follow-up visits, MAE (SD) among eyes with NPDR, with or without center-involved DME (CI-DME), was 1.16 (1.00) lines on the eye chart for VA between 20/10 and 20/20 (67 images), and 1.44 (1.15) lines for between VA 20/25 and 20/80 (231 images). MAE (SD) among eyes with PDR, with or without CI-DME, was 1.92 (1.08) lines for VA between 20/10 and 20/20 (50 images), and 1.42 (0.97) lines for spectacle-corrected VA between 20/25 and 20/80 (150 images). Only 65 images had VA 20/100 or worse, precluding meaningful analyses.
In this cross-sectional study, AI evaluation of fundus photographs among patients with DME and VA 20/80 or better estimated spectacle-corrected VA within approximately 1 to 1.5 lines of actual spectacle-corrected VA. These results support use of AI evaluation of fundus photographs to determine spectacle-corrected VA among patients with DME globally, beyond ophthalmology offices.
在管理多种眼科疾病时，确定经眼镜矫正的视力（VA）至关重要。如果人工智能（AI）对黄斑图像的评估能够从眼底图像中估算出这种视力，那么AI或许可以在不产生技术人员成本的情况下提供经眼镜矫正的视力，减少就诊时间，或便于根据在诊所外获取的眼底图像对视力进行家庭监测。
在临床实践环境中，使用先前经过验证的AI算法，通过评估糖尿病性黄斑水肿（DME）患者眼底照片来估算其最佳矫正视力，从而估计这些患者在标准视力表上测得的经眼镜矫正的视力。
设计、设置和参与者：对去识别化的眼底照片进行回顾性横断面评估，这些照片与技术人员在视力表上确定的经眼镜矫正的视力相匹配，研究对象为有DME病史的患者，基于光学相干断层扫描，且在2014年1月至2022年12月期间于一家大学诊所彼此间隔1至6个月进行了至少2次就诊。数据于2023年1月至2024年10月进行分析。
对眼底照片进行先前经过验证的AI算法评估。
将AI确定的视力平均绝对误差（MAE）与实际经眼镜矫正的视力进行比较。
141例患者中，平均（标准差）年龄为63（13）岁，71例（50%）为男性，2例（1%）为亚洲人，42例（30%）为黑人或非裔美国人，88例（63%）为白人。在第1次就诊的282只眼中，66只患有非增殖性糖尿病视网膜病变（NPDR）和DME，38只患有增殖性糖尿病视网膜病变（PDR）和DME，101只患有NPDR且无DME，77只患有PDR且无DME。在初次和随访就诊的564张图像（282只眼）中，对于视力在20/10至20/20之间的NPDR眼，无论有无中心累及性DME（CI - DME），在视力表上的MAE（标准差）为1.16（1.00）行（67张图像），对于视力在20/25至20/80之间的为1.44（1.15）行（231张图像）。对于PDR眼，无论有无CI - DME，视力在20/10至20/20之间的MAE（标准差）为1.92（1.08）行（50张图像），对于经眼镜矫正视力在20/25至20/80之间的为1.42（0.97）行（150张图像）。只有65张图像的视力为20/100或更差，无法进行有意义的分析。
在这项横断面研究中，对于DME且视力为20/80或更好的患者，AI对眼底照片的评估所估算的经眼镜矫正视力与实际经眼镜矫正视力相差约1至1.5行。这些结果支持在全球范围内，在眼科诊所之外，使用AI对眼底照片的评估来确定DME患者的经眼镜矫正视力。
