NGHIÊN CỨU ĐẶC ĐIỂM BỆNH VÕNG MẠC ĐÁI THÁO ĐƯỜNG VÀ ỨNG DỤNG TRÍ TUỆ NHÂN TẠO TRONG KHÁM SÀNG LỌC TẠI BỆNH VIỆN ĐA KHOA TỈNH PHÚ THỌ

Abstract

Purpose: To describe the characteristics of diabetic retinopathy (DR) and evaluate the application of artificial intelligence (AI) in screening for DR at Phu Tho General Hospital, Vietnam. Methods: A total of 190 eyes from 95 patients with diabetes mellitus were included. Fundus photographs were taken and analysed for image quality, DR severity, and retinal lesions. AI-based screening was performed using validated algorithms, and its diagnostic performance was assessed using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and overall accuracy. Results: Among the eyes studied, 37.3% had no DR, 53.7% had mild non-proliferative DR, 5.8% had moderate DR, and 3.2% had severe DR. The most common retinal lesion was dot/blot hemorrhages (60.5%), followed by hard exudates (10%), cotton wool spots (4.2%), microaneurysms (5.3%), and IRMA (2.6%). No cases of proliferative DR or pre-retinal hemorrhage were observed. The average image quality score was 75.25 ± 16.26, with 66.32% of images suitable for AI classification. AI screening achieved a sensitivity of 95.8%, specificity of 52.9%, PPV of 77.7%, NPV of 88.1%, and an overall accuracy of 80%. Conclusions: The majority of DR cases were detected at an early stage, highlighting the importance of timely screening. AI demonstrated high sensitivity and acceptable accuracy for DR detection, making it a useful tool for community-level screening, particularly in areas with limited ophthalmology resources. Integration of AI into the routine fundus photography workflow can enhance early detection, reduce ophthalmologists' workload, and improve the efficiency of DR screening programs.