The diagnostic accuracy of AI-assisted diabetic retinopathy screening in primary care: a prospective validation study

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Sensitivity 73.7%, Specificity 90.2% (all images); Sensitivity 80.0% (excluding poor quality)

• Primary Outcome: Diagnostic accuracy of AI-assisted diabetic retinopathy screening

• Nnt Or Sensitivity: Sensitivity 73.7% (all images), 80.0% (good quality images); NPV 98.3%

• Confidence Interval: Not reported

Clinical Application

Moderate - requires good quality fundus photography and may need quality control measures Relevant to Australian primary care and could complement existing diabetic retinopathy screening programs, potentially reducing ophthalmology referral burden Adults with type 2 diabetes in primary care settings requiring diabetic retinopathy screening

Abstract

OBJECTIVES: This study investigated the diagnostic accuracy of AI-assisted diabetic retinopathy screening in primary care, using ophthalmologist-led screening as the reference standard. METHODS: Patients with type 2 diabetes attending routine appointments at 10 primary care clinics underwent AI-assisted screening, followed by re-screening at an ophthalmology clinic. The quality of fundus images captured in primary care was independently assessed, and diagnostic accuracy was evaluated by comparing AI-assisted results with ophthalmologist results, including sensitivity, specificity, PPV, NPV, and AUC. Two analyses were conducted: one including all images and one excluding those of poor quality. RESULTS: Among 183 patients (336 images), 18.6% of images were classified as poor quality. When all images were included, the AI-assisted screening achieved a sensitivity of 73.7%, specificity of 90.2%, PPV of 31.1%, NPV of 98.3%, and AUC of 0.82. Excluding poor-quality images improved sensitivity to 80.0%, NPV to 98.7%, and AUC to 0.84. Additional ocular findings unrelated to diabetic retinopathy were observed in 96 patients, including confirmed or non-specific signs of glaucoma, cataract, age-macular degeneration, benign nevus and reduced visual acuity. CONCLUSION: AI-assisted screening in primary care shows potential for clinical application, but further validation in larger populations and improvements in image quality are needed before clinical implementation.

References

1. 1.Krogh, M., Ørskov, M., Nørgaard, T. L., Nielsen, M. G., Jensen, M. B., Jensen, M. S. A., Vorum, H., & Kristensen, J. K. (2026). The diagnostic accuracy of AI-assisted diabetic retinopathy screening in primary care: a prospective validation study. Scandinavian Journal of Primary Health Care. https://doi.org/10.1080/02813432.2026.2683982