Artificial intelligence to detect referable diabetic retinopathy in a Chinese population in Hong Kong

Objective: This study aimed to investigate the sensitivity and specificity of the Deep Fundus, an artificial intelligence program for diabetic retinopathy screening, in Chinese patients with diabetes. Methods: Fundus photographs of Chinese patients with diabetes aged 50 to 90 years who attended Pamela Youde Nethersole Eastern Hospital, Hong Kong between December 2020 and June 2021 during the public diabetic retinopathy screening period were retrospectively retrieved in February 2022. Two ophthalmologists independently graded the fundus photographs for diabetic retinopathy (according to the recommendations in the 2017 International Council of Ophthalmology Guidelines for Diabetic Eye Care) and any concomitant eye pathology such as cataracts or macular diseases. On 27 February 2022, the photographs were uploaded to the Deep Fundus website. Fundus images were classified into referable or non-referable diabetic retinopathy in accordance with the abovementioned Guidelines. Sensitivity and specificity of the Deep Fundus were determined. Results: Of 202 eyes screened, 96 eyes were included in the analysis. There was no discordance between the two ophthalmologists in the diabetic retinopathy gradings. 57 (59.4%) eyes were graded to have positive diabetic retinopathy findings. Using gradings of the two ophthalmologists as the gold standard, sensitivity and specificity of the Deep Fundus were 62.3% and 90.7%, respectively. The positive and negative predictive values were 89.2% and 66.1%, respectively. Conclusions: The sensitivity and specificity of the Deep Fundus for diabetic retinopathy screening in a Chinese population were 62.3% and 90.7%, respectively. Further real-world validation studies are needed to assess the artificial intelligence model, particularly among Chinese patients with cataracts, myopia, diabetic macular edema, or proliferative diabetic retinopathy.