DIRD+: A Browser-Based, Offline-First Clinical Platform for Diabetic Retinopathy Screening Using Edge AI Inference in Low-Resource Settings

BackgroundDiabetic retinopathy (DR) is the leading cause of preventable blindness in working-age adults. In Chile, despite GES coverage since 2006, ophthalmic screening reaches only [~]21% of diabetic patients under control. Real-world evidence from Chilean clinical settings shows that autonomous AI screening platforms have produced heterogeneous field results -- sensitivity ranging from 40.8% to 100% with specificity as low as 55.4% -- while Ophthalmic Medical Technologists (TMOs) consistently achieve sensitivity above 97% in the same studies. This evidence motivates an AI architecture designed to support and maximize the TMOs clinical time, not to replace their judgment. ObjectiveTo develop and describe an open-source, offline-first clinical decision support platform for DR that operates without server infrastructure, preserves patient data sovereignty, and is designed to structure and accelerate the TMOs workflow by providing AI-generated lesion candidates for expert review -- with preliminary technical validation of the detection component. MethodsDIRD+ (Diabetic Integrated Retinal Diagnosis) implements a six-stage on-device inference pipeline using ONNX Runtime -- via WebAssembly in the browser or native runtime in the Tauri/Rust desktop application. The system integrates patient management, bilateral image analysis, a multi-layer annotation canvas, a pluggable clinical guideline engine (ICDR 2024, MINSAL Chile 2017), LLM-assisted report narration, and collaborative dataset contribution. A YOLOv26n detection model was trained on 500 pseudo-labeled APTOS 2019 images using the Annotix framework [20] and evaluated on the IDRiD test set (n=81 images). ResultsOptic disc detection -- the spatial calibration landmark -- achieved AP=1.000 on IDRiD (IoU threshold=0.1, F1=1.000). Soft exudate detection achieved AP=0.243 (F1=0.364). Internal validation mAP50=0.578 globally. On-device inference in the Tauri desktop application averaged 0.297 s/image (3.4 images/second) on CPU without GPU. Detection performance reflects a first-generation model trained on 500 images; progressive improvement through collaborative annotation is ongoing. ConclusionsDIRD+ demonstrates that a complete offline-first DR clinical workflow can be deployed at zero cost without server infrastructure or GPU, in both web browser and native desktop environments. The human-in-the-loop architecture -- where AI structures findings and the specialist decides -- is grounded in Chilean clinical evidence and offers a viable pathway for TMO-assisted DR screening in connectivity-limited primary care settings.