Offline smartphone AI accurately detects DR, glaucoma, and AMD in real-world study

A multi-disease AI system that runs entirely offline on a smartphone-based fundus camera identified diabetic retinopathy, glaucoma, and age-related macular degeneration with high diagnostic accuracy in a real-world cohort, according to a prospective study published in the European Journal of Ophthalmology.¹

Investigators from the National Institute of Ophthalmology in Pune, India, evaluated the Medios-AI (MAI) algorithm, a deep learning system embedded in a handheld fundus camera that generates disease-specific reports at the point of care. The appeal of such a platform is its independence from cloud connectivity—an obstacle that has limited the reach of many AI screening tools in exactly the low-resource environments where retinal disease often goes undetected until vision is lost.

A single image, three diseases

The interest in consolidating screening reflects a broader shift in the field, where AI has moved fastest in retinal imaging and is increasingly framed as a way to extend specialist-level assessment to primary care and community settings.² Most validated autonomous tools, however, have targeted a single condition—typically DR. The MAI study tested whether one algorithm could screen for 3 of the leading causes of irreversible vision loss from the same set of fundus photographs.

In the prospective, cross-sectional study, 193 adults (371 eyes) aged 18 years or older with DR, glaucoma, AMD, or a normal fundus were enrolled between May and December 2024. Dilated fundus imaging was captured using the Remidio Fundus on Phone and Zeiss Clarus 500 cameras, with ungradable images excluded. The offline algorithm's disease-specific reports were then compared against masked grading of the Clarus images by 2 fellowship-trained ophthalmologists, who served as the reference standard. In ambiguous cases, the AI report defaulted to a combined "either DR or AMD" designation.

Strong overall accuracy, with caveats by disease

For detecting any retinal disease, MAI achieved a sensitivity of 99.3% (95% CI, 96-100), a specificity of 95.7% (95% CI, 92-98), and an AUROC of 0.99. Performance varied by condition. Glaucoma screening (n = 109) was the standout, with 98.2% sensitivity, 99.0% specificity, and an AUROC of 0.99. For AMD (n = 56), sensitivity was 88.9% and specificity 97.5% (AUROC, 0.93), while DR (n = 78) showed a sensitivity of 84.6% and a specificity of 99.0% (AUROC, 0.92).

The glaucoma result is notable given that AI has historically translated more readily to DR than to glaucoma, where diagnosis rests on a constellation of structural and functional findings rather than a single image-based signature.³ Here, agreement on vertical cup-to-disc ratio between the algorithm and the human graders fell within a narrow margin of −0.1 to +0.1, with an intergrader intraclass correlation coefficient of 0.97 (P < .001 for all comparisons). The comparatively lower DR sensitivity, by contrast, leaves open the question of how many early or subtle cases a real-world deployment might miss.

Where it could fit in practice

The authors concluded that MAI demonstrated significant diagnostic accuracy across all 3 diseases on an offline, smartphone-based platform, supporting its use for scalable, point-of-care retinal screening in resource-limited settings. The framing aligns with a growing body of real-world AI validation work in ophthalmology, including recent reports that algorithm-based DR screening can match or exceed conventional telemedicine grading.⁴

Several considerations temper the findings. The cohort was modest in size and drawn from a single institution; the design was cross-sectional rather than longitudinal, and ungradable images were excluded from analysis—an exclusion that real-world programs must contend with directly. The lower sensitivities for DR and AMD also underscore that a tool optimized for broad detection of "any disease" may behave differently when the clinical question is a specific, sight-threatening diagnosis.

Still, the prospect of a connectivity-independent device that triages 3 blinding conditions in 1 capture speaks to a practical need. As AI screening continues its push from proof-of-concept toward routine deployment, multi-disease platforms that function at the edge of the health system—where the burden of undiagnosed disease is heaviest—may prove among the more consequential applications.

Reference

1. Kelkar A, Kelkar J, Garg Y, Jain HH, Sengupta S. Smartphone-based offline AI for multi-disease retinal screening: real-world accuracy. Eur J Ophthalmol. Published online June 24, 2026. doi:10.1177/11206721261462321

2. Charters L. How AI is reshaping ophthalmology in 2025 and beyond. Ophthalmology Times. December 23, 2025.Schuman JS,

3. Stevenson S. Glaucoma 2026: which emerging technologies will change practice? Ophthalmology Times. December 26, 2025.

4. Charters L. A real-world study of an AI system for enhanced detection of diabetic retinopathy. Ophthalmology Times Europe. April 7, 2025.