Comparison of Retinopathy Screening Performed Using Conventional Methods and an Artificial Intelligence-Assisted System in Prediabetic and Diabetic Patients: A Single-Center Prospective Analysis
Background and Objectives: AI-assisted retinal imaging is increasingly proposed to expand diabetic retinopathy (DR) screening capacity, but most validation studies use reading-center standards, exclude prediabetes, and assess macular edema against photographic surrogates rather than optical coherence tomography (OCT). We compared an AI-assisted retinal screening system with routine ophthalmological examination in an unselected internal medicine screening population. Materials and Methods: In this prospective, cross-sectional diagnostic accuracy study, consecutive adults with prediabetes or diabetes underwent AI-assisted retinal imaging (index test) followed by masked ophthalmological examination (reference standard), with OCT as the reference for macular edema. Analyses were performed at the eye level (n = 714) and patient level (n = 357) using Wilson score 95% confidence intervals and Cohen kappa (κ) for agreement. Results: Among 357 patients (714 eyes; 83.8% diabetes, 16.2% prediabetes), DR prevalence by ophthalmological examination was 15.0% (eyes) and 17.9% (patients), versus 24.4% and 33.1% by AI (McNemar p < 0.001). Eye-level and patient-level sensitivity and specificity were 70.1% and 83.7% and 75.0% and 76.1%, respectively. Negative predictive value was high (93–94%) but positive predictive value was low (41–43%); agreement was moderate at the eye level and fair at the patient level (κ = 0.428 and 0.384, respectively). False-positive AI calls were associated with higher HbA1c and albuminuria (adjusted odds ratio for log-transformed ACR 1.23, p = 0.040); false negatives had milder disease. In prediabetes, DR prevalence was 1.7% and agreement was absent (κ = −0.03). AI sensitivity for OCT-defined macular edema was 45.5% (κ = 0.197), with no significant central macular thickness difference by AI status. Conclusions: The AI system showed high negative predictive value but systematically over-called DR, with only fair-to-moderate agreement and performance below accepted screening benchmarks, particularly in prediabetes and for macular edema. These findings support its use as a rule-out triage adjunct with ophthalmological confirmation of positive results, rather than an autonomous substitute, and argue for local recalibration before deployment.
Authors
- Kübra Gul Olke (ORCID: https://orcid.org/0000-0003-4486-3775)
- Hilmi Erdem Sümbül (ORCID: https://orcid.org/0000-0002-7192-0280)
- Ahmet Gazi Mustan (ORCID: https://orcid.org/0000-0002-9784-5063)
- Begüm Şeyda Avcı (ORCID: https://orcid.org/0000-0001-6149-9341)
- Ahmet Emin Demirtaş (ORCID: https://orcid.org/0009-0008-7771-3558)
- Erdinç Gülümsek (ORCID: https://orcid.org/0000-0003-0026-2396)
- İrfan Alişan (ORCID: https://orcid.org/0000-0002-9473-9476)
- Fatih Necip Arıcı (ORCID: https://orcid.org/0000-0002-1160-2125)
- Cisem Kizildag (ORCID: https://orcid.org/0000-0002-7140-1165)
- Ramazan Azim Okyay (ORCID: https://orcid.org/0000-0001-8767-2771)
- Bektaş Işık (ORCID: https://orcid.org/0000-0001-9553-722X)
- Hüseyin Ali Öztürk (ORCID: https://orcid.org/0000-0001-5713-004X)
- Cahit Dincer (ORCID: https://orcid.org/0000-0001-5428-895X)
Institutions
- Sağlık Bilimleri Üniversitesi (TR)
Publication Details
- Journal
- Medicina
- Published
- 2026-11-18
- DOI
- https://doi.org/10.3390/medicina62112267
- Primary Topic
- Retinal Imaging and Analysis
- Type
- article
- Field-Weighted Citation Impact
- 0.00