[Accuracy of artificial intelligence compared to trained medical technologists in diabetic retinopathy screening]

The early detection of retinopathy among diabetics is of utmost importance.

To estimate the diagnostic accuracy of two diabetic retinopathy (DR) screening strategies currently used in the Chilean public health system.

Cross-sectional observational study of 371 diabetic patients aged 61 ± 14 years (61% women) who underwent DR screening at a public Hospital between July 1 and August 31, 2019. The mydriatic retinal photographs of all participants were classified using artificial intelligence software (DART) and trained medical technologists, independently. The precision of both strategies was compared with the reference standard, namely the evaluation of the fundus by an ophthalmologist with a slit lamp. Participants with severe non-proliferative DR or worse were considered as positive cases. The ophthalmologist was blind to the results of the screening tests.

Twenty four percent of participants had DR, including 34 (9.2%) who had sight threatening DR in at least one eye. The sensitivity and specificity of DART were 100% (95% confidence intervals (CI): 90-100%) and 55,4% (95% CI: 50-61%), respectively. Medical technologists had a sensitivity of 97,1% (95% CI: 85-100%) and a specificity of 91,7% (95% CI: 88-94%). The only case missed by medical technologists was a patient with unilateral panphotocoagulated DR.

Both strategies had a similar sensitivity to detect cases of sight-threatening DR. However, the specificity of DART was significantly lower compared to medical technologists, which would greatly increase the burden on the health system, a very important aspect to consider in a screening strategy.