JMIR Publications

ABSTRACT

Background:

Background:

Diabetic retinopathy is the leading cause of vision loss among working - age adults worldwide. Artificial intelligence assisted automated image reading has effectively alleviated the human resource challenges in large-scale remote screenings, yet there is limited analysis on the impact of complex ocular factors on the diagnostic efficacy of artificial intelligence.

Objective:

Objective:

To identify ocular factors interfering with artificial intelligence based diagnosis in diabetic retinopathy screening and optimize diagnostic performance.

Methods:

Methods:

This cross-sectional study screened residents with type 2 diabetes at primary health centers in underdeveloped regions. Following standard ophthalmic examinations, diabetic retinopathy grading was performed using EVisionAI and compared with expert grading. The influence of various ocular factors on the diagnostic performance of EVisionAI was subsequently evaluated.

Results:

Results:

A total of 1847 participants with type 2 diabetes, aged 32 to 91 years old, were enrolled between October 21, 2024, and November 12, 2024, of whom 1748 participants (3496 eyes) completed the screening process. Although EVisionAI demonstrates diagnostic capability comparable to ophthalmologists (Sensitivity: 90.61% [95% CI, 87.89% to 92.79%]; Specificity: 98.99% [95% CI, 98.52% to 99.31%]), ocular factors - including pupil size, refractive media opacity, and tessellated fundus - can variably impact EVisionAI's accurate identification of diabetic retinopathy lesions. When reach certain levels, these ocular factors lead to significant declines in EVisionAI's sensitivity (Refractive media opacity: 80.95% [95% CI, 68.71% to 89.36%]; Tessellated fundus: 82.86% [95% CI, 65.70% to 92.83%]). Compared to vision-threatening diabetic retinopathy lesions - which EVisionAI rarely misses - the correct identification of early-stage diabetic retinopathy is more susceptible to interference from ocular factors, particularly severe tessellated fundus changes. Most notably, severe refractive media opacity caused by vitreous degeneration almost invariably results in misdiagnosis of vision-threatening diabetic retinopathy.

Conclusions:

Conclusions:

Although EVisionAI achieves high sensitivity and specificity comparable to human graders in diabetic retinopathy screening, severe ocular factors can still diminish its diagnostic performance, particularly for early-stage diabetic retinopathy. Optimizing for these factors during remote screening - particularly for early-stage cases presenting with multiple ocular factors - can significantly enhance its recognition efficiency, and contribute to further improving the accuracy and referral efficiency for diabetic retinopathy in primary health centers with limited specialists.