Validation study reveals high diagnostic accuracy for smartphone-based AI glaucoma screening in adults, suggesting viable point-of-care detection in low-resource settings.
Key Points
To develop and clinically evaluate Glaucoma Screening on Phone (GSoP), an ultra-low-cost, fully offline smartphone-based artificial intelligence platform for glaucoma screening in low-resource clinical settings.
Enrolled a case-control cohort of 208 adults across two clinical sites (JUMC and UZ Leuven), capturing 345 short optic disc videos after pharmacological pupil dilation using a 3D-printed optical smartphone adaptor.
Implemented a fully local, on-device machine learning pipeline combining automated frame selection, optic disc localization via YOLOv8n, and glaucoma classification using fine-tuned EfficientNetV2–B0 models.
YOLOv8n-based automated localization isolated the optic disc in 100% of evaluated clinical test frames, with a complete workflow time of 2–3 minutes per eye.
Achieved an AUC of 0.99, 96.1% accuracy, 96.2% sensitivity, and 96.0% specificity in 5-fold cross-validation, with 100% accuracy, sensitivity, and specificity on the primary-site holdout set.
Maintained 100% sensitivity with 76.9% accuracy and 71.9% specificity when tested against cross-site demographic and operational domain shifts at the second independent clinical site.