Clinical validation of an ultra-low-cost smartphone-based offline AI platform for glaucoma screening in low-resource settings

Glaucoma is one of the leading causes of irreversible blindness worldwide, affecting more than 76 million people, with prevalence expected to increase substantially in the coming decades. Early detection is critical to prevent vision loss; however, access to ophthalmologists and fundus imaging devices remains limited in many low- and middle-income countries. This study aimed to develop and evaluate Glaucoma Screening on Phone (GSoP), a low-cost, fully offline, smartphone-based platform as a semi-automated proof-of-concept for glaucoma screening by trained non-specialists in resource-limited settings. The GSoP platform integrates a 3D-printed optical adaptor with an on-device machine learning pipeline deployed on a standard smartphone. Short optic disc videos ( n = 345) were captured from a case-control cohort of 208 adults following standard pharmacological dilation at two clinical sites. Data was processed entirely locally for automated frame selection, optic disc localization using a previously validated YOLOv8n-based module, and glaucoma classification (EfficientNetV2–B0). The classification model was fine-tuned using GSoP-acquired frames to account for device-specific optical characteristics. During on-device inference, the automated localization module successfully isolated the optic disc in 100% of the evaluated clinical test frames. In five-fold participant-level group cross-validation, the classification model achieved an area under the receiver operating characteristic curve (AUC) of 0.99, with 96.1% accuracy, 96.2% sensitivity, and 96% specificity. During independent testing, the model correctly classified all cases in the primary-site holdout dataset (JUMC), achieving 100% accuracy, sensitivity, and specificity. When evaluated on the cross-site independent test set (UZ Leuven), despite the domain shift associated with differences in demographic and operational conditions, the model maintained 100% sensitivity, with 76.9% accuracy and 71.9% specificity. The complete semi-automated screening workflow was completed within 2–3 minutes per eye using a standard smartphone. The GSoP platform demonstrates the feasibility of a fully offline, ultra-low-cost, semi-automated glaucoma screening proof-of-concept. While future prospective validation on unselected, non-mydriatic populations is required to realize true autonomous community deployment, the platform’s capability to maintain high sensitivity despite cross-site domain shifts highlights its potential as an accessible point-of-care screening tool for resource-constrained clinical environments.