Título

Clinical validation of AirDoc camera and automated system for diabetic retinopathy screening in a Brazilian population

Objetivo

Diabetic retinopathy (DR) stands as the foremost cause of preventable blindness in adults. Despite efforts to expand DR screening coverage in the Brazilian public healthcare system, challenges persist due to various factors including social, medical, and financial constraints. Portable retinal cameras with embedded artificial intelligence (AI) algorithms for automated image classification offer a promising alternative for DR screening. Our objective was to evaluate the performance of Airdoc portable cameras with embedded AI for DR screening, in a Brazilian sample.

Método

Images were captured by two portable retinal devices: AirDoc and Eyer. The included patients had their fundus images obtained in a screening program conducted in Blumenau, Santa Catarina. Two retina specialists independently assessed each image's quality and graded DR classification. The performance of each system for the detection of DR was assessed, and a comparison was performed between both devices regarding image quality and automated diagnostic accuracy for DR.

Resultado

The analysis included 129 patients (mean age of 61 years), with 29 (43.28%) male and an average disease duration of 11.1±8 years. In Ardoc, 21 (16.28%) images were classified as poor quality, with 88 (68%) presenting artifacts; in Eyer, 4 (3.1%) images were classified as poor quality, with 94 (72.87%) presenting artifacts. Eyer's ground-truth images showed a DR prevalence of 16.8%, 14 (11.2%) mild non-proliferative, 7 (5.6%) moderate non-proliferative, and 3 (2.4%) diabetic macular edema (DME). Analysis revealed 87 (82.86%) concordant DR classifications between cameras, with AirDoc underestimating 12 (11.43%) and overestimating 6 (5.71%) cases. For DME, 108 (98.1%) images had concordant reports. The AirDoc AI system showed 100% specificity in detecting referable diabetic retinopathy, but a 33% sensitivity and 66.7% false negative rate.

Conclusão

AirDoc images displayed higher rates of ungradable and low-quality images, affecting the DR grading. Additionally, the AirDoc AI system exhibited low sensitivity and elevated rates of false negatives.