Abstract
BACKGROUND: This study aimed to assess the diagnostic accuracy of an artificial intelligence (AI) system integrated with a portable handheld fundus camera for the detection of diabetic retinopathy (DR) in a community-based screening program. METHODS: A DR screening camp was organized at a tertiary care hospital in India. A cohort of 261 patients with diabetes was screened using a nonmydriatic handheld fundus camera. Retinal images were graded by specialists and compared with the AI system's output. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUC-ROC) were calculated. Subgroup analyses based on image quality was performed. RESULTS: Of the 261 patients screened, 253 had available retinal images, and 243 had gradable images. The AI system achieved a sensitivity of 85.29%, specificity of 99.04%, PPV of 93.55%, and NPV of 97.64% for detecting referable DR. The AUC-ROC was 0.93. The AI system's performance remained robust across all image-quality categories. The AI system showed strong agreement with human graders (κ = 0.86). However, it failed to identify certain non-DR pathologies detected by human graders. CONCLUSIONS: The AI system integrated with a portable handheld fundus camera demonstrated high diagnostic accuracy for referable DR detection in a community-based screening setting. This technology shows promise for expanding DR-screening coverage in resource-limited settings.