Artificial intelligence for diabetic retinopathy screening in primary care: Diagnostic accuracy and real-world implementation

Background

Diabetic retinopathy is a leading cause of preventable vision loss, yet adherence to recommended retinal screening remains suboptimal. Artificial intelligence-based systems can analyze retinal photographs in primary care settings and may improve access to early detection. This review evaluates the diagnostic performance and practical implementation of artificial intelligence-assisted diabetic retinopathy screening outside traditional ophthalmology clinics.

Methods

A focused literature review was conducted using PubMed. English-language studies published from January 2024 through September 2026 were reviewed. Systematic reviews, meta-analyses, and prospective or real-world studies evaluating artificial intelligence-assisted diabetic retinopathy screening in primary care, endocrinology, or population-based settings were included. Outcomes of interest were sensitivity, specificity, image gradability, workflow integration, patient and staff acceptance, and referral outcomes.

Results / Findings

Ten publications were included. Recent meta-analyses reported patient-level sensitivities ranging from 87.7% to 94.0% and specificities from 89.0% to 90.6%. Real-world primary-care studies showed wider ranges, with sensitivities from 73.7% to 98.2% and specificities from 90.2% to 99.8%. Differences in image quality, screening thresholds, devices, and clinical workflow contributed to variability in performance. Poor-quality retinal images were associated with lower sensitivity in prospective studies. Some systems provided definitive nonmydriatic screening results in more than 95% of patients, and recent implementation studies reported favorable acceptance among patients and healthcare staff.

Discussion & Conclusion

Artificial intelligence-assisted retinal screening shows strong diagnostic performance and may support earlier detection of diabetic retinopathy in primary care and other non-ophthalmology settings. However, effectiveness in routine practice depends on image quality, device selection, workflow integration, and appropriate ophthalmic referral. These systems can expand access to diabetic eye screening but do not replace a complete ophthalmologic evaluation when disease is detected or images are not interpretable. Further studies should evaluate cost-effectiveness, long-term outcomes, and performance across diverse and underserved populations, including Hispanic/Latino communities.