Background: Pupil dilatation, access to ophthalmologists, size and cost of fundus cameras, and network connectivity issues are barriers to diabetic retinopathy (DR) screening in primary care. This makes artificial intelligence a potential alternative in DR screening. This study aims to evaluate the performance of an offline AI algorithm - Medios in DR Screening using non mydriatic retinal images taken from the smartphone based Remidio Fundus on Phone (NM-FOP) retinal camera.

Methods: Non mydriatic retinal images from 900 patients with diabetes were captured prospectively with the Remidio NM FOP camera at Diacon Hospital, Bangalore, India. The images were graded by 5 ophthalmologists as per the International Diabetic Retinopathy Classification System and the majority diagnosis was considered as the ground truth. Two images [posterior pole (macula centered), nasal field] were captured per eye for each patient, and were analyzed offline on-device using the Medios AI. The diagnosis of Medios AI was compared to that of the ophthalmologist.

Results: DR was present in 252 patients. Performance of the AI for referable DR (Moderate NPDR or worse, or presence of DME) was - Sensitivity 93% (95% CI 91.3%-94.7%), Specificity: 92.5% (95% CI 90.8%-94.2%). Performance of the AI for all DR (Mild NPDR or worse, or presence of DME) was - Sensitivity: 83.3% (95% CI 80.9%-85.7%), Specificity: 95.5% (95% CI 94.1%-96.8%). Sensitivity for sight threatening DR (STDR) [severe NPDR, proliferative DR (PDR) or presence of DME] was 95.2% (95% CI 88.2%-98.6%). The 3 PDRs missed by the AI were post laser with no active changes visible. This is a common exclusion criteria in other studies, and when excluded, the sensitivity for STDR was found to be 98.7%.

Conclusion: This is one of the largest studies evaluating the performance of AI using non mydriatic retinal images offline. The Medios AI algorithm’s sensitivity and specificity exceeds the FDA mandated superiority end point and can be used for DR screening.


B. Sosale: None. A.R. Sosale: None. H. Murthy: None. S. Narayana: None. U. Sharma: None. S.G.V. Gowda: None. M. Naveenam: None.

Readers may use this article as long as the work is properly cited, the use is educational and not for profit, and the work is not altered. More information is available at