Accuracy of A Deep Learning Model in Retinal Imaging Analysis for The Early Detection of Diabetic Retinopathy in A Southeast Asian Population: A Diagnostic Validation Study

Abstract

Background: Diabetic retinopathy (DR) represents a leading cause of preventable blindness in Southeast Asia, where diabetes prevalence continues rising dramatically. Deep learning models show promising diagnostic accuracy for DR detection, yet validation in Southeast Asian populations remains limited. Objective: To evaluate the accuracy and clinical applicability of a deep learning model for early DR detection through comprehensive qualitative analysis in Indonesian healthcare settings. Methods: A mixed-methods diagnostic validation study was conducted across three Indonesian provinces from January 2023 to December 2024. The study employed a convolutional neural network-based deep learning model trained on 15,000 retinal images for DR classification. Qualitative data collection included semi-structured interviews with 30 ophthalmologists, 20 primary care physicians, 15 healthcare administrators, and 40 patients. Thematic analysis explored stakeholder perspectives on diagnostic accuracy, implementation barriers, and clinical integration potential. Results: The deep learning model demonstrated 89.3% accuracy (95% CI: 86.7-92.1%), 91.7% sensitivity, and 87.1% specificity for detecting referable DR. Qualitative analysis revealed high stakeholder acceptance (87.5% patient trust, 90.0% physician interest) despite implementation concerns. Key themes included diagnostic accuracy validation needs, workflow integration challenges, infrastructure requirements, and cost-effectiveness potential. Primary barriers included image quality standardization, internet connectivity limitations, and regulatory approval processes. Conclusion: Deep learning models demonstrate promising diagnostic performance for DR screening in Southeast Asian populations, with strong stakeholder support for implementation. However, successful deployment requires addressing infrastructure limitations, regulatory frameworks, and clinician training needs. These findings support the potential for AI-enhanced DR screening to improve early detection outcomes in resource-constrained healthcare systems across Southeast Asia.