- An offline artificial intelligence (AI) algorithm can detect referable diabetic retinopathy ([R]DR) images taken from a smartphone-based portable camera.
Why this matters
- Current ophthalmology-based screening methods are costly and time-consuming and require training.
- Prospective study of the Medios AI (Remidio) automated system.
- System analyzes retinal images (3 fundus fields) taken with a smartphone-based nonmydriatic retinal camera.
- A minimally trained health care worker imaged 213 patients with diabetes mellitus (DM) at clinics in Mumbai, India.
- RDR: retinopathy more severe than mild DR, with/without macular edema.
- Results compared with direct ophthalmologist grading.
- Funding: Medios Technologies provided the AI software.
- Of 187 patients diagnosed as not having RDR by ophthalmologists, AI correctly diagnosed 92%.
- Of 15 with ophthalmologist-diagnosed RDR, AI correctly diagnosed 100%.
- Of 12 with ophthalmologist-diagnosed mild nonproliferative DR, AI diagnosed 67%.
- AI sensitivity and specificity (95% CIs) were, respectively:
- RDR: 100% (78.2%-100%) and 88.4% (83.16%-92.53%); and
- Any DR: 85.2% (66.3%-95.8%) and 92.0% (87.1%-95.4%).
- Small sample size.
- Current AI version does not allow grading according to International Clinical DR severity scale or National Health Service classification.
- Ungradable images resulting from cataracts and small pupil size necessitated dilated retinal photography screening.
- Local setting, generalizability uncertain.