Abstract
Purpose :
The performance of the offline Medios Artificial Intelligence (AI) screening algorithm deployed on the Remidio NM FOP 10 fundus camera has only been validated for the detection of RDR (defined as more than mild Diabetic Retinopathy). A new AI algorithm further detecting STDR (defined as more than moderate Diabetic Retinopathy or presence of Diabetic Macular Edema (DME)) has been added to the system. This study validates the overall performance of the new system.
Methods :
Non-mydriatic retinal images were captured from 950 individuals with diabetes during routine hospital visits at Diacon Hospital, Bangalore, India. Two images [posterior pole (macula centered), nasal field] were captured per eye using the Remidio NM FOP 10 camera. These were graded individually by 5 retina specialists as per the International Clinical Diabetic Retinopathy Disease Severity Scale. Images were run on the desktop version of the Medios AI grading algorithms, which classifies the patients between healthy, RDR and STDR, and indicates the presence of DME. Diagnosis of the AI was compared to the majority diagnosis of the ophthalmologists (taken to be the ground truth).
Results :
Patients with at least one image deemed ungradable by the AI system were excluded from the analysis (89 patients – 9.37%). Analysis included images from 861 patients of which 184 had RDR. The sensitivity of the AI for RDR detection was 94.56% (95% CI 89.94% - 97.21%) and the specificity 90.89% (95% CI 88.51% - 92.98%). The sensitivity of the AI for STDR detection was 95.59% (95% CI 86.81% - 98.85%) and the specificity 94.96% (95% CI 93.13% -96.33%). The sensitivity of the AI for DME detection was 97.78% (95% CI 86.77% - 99.88%) and the specificity 96.32% (95% CI 94.73% - 97.46%).
Conclusions :
The new AI algorithm that detects STDR cases has high sensitivity and specificity, above the required regulatory mandate. Further work includes integrating and deploying the AI offline on the smart phone device and prospectively validating it.
This is a 2020 ARVO Annual Meeting abstract.