TITLE:
H-UQ-MFF: Hybrid Uncertainty-Aware Multi-Feature Fusion for Clinically-Translatable Glaucoma Detection with FDA-Compliant Validation
AUTHORS:
Venkata Akhil Mettu, Sree Charitha Obiliachigari, Siri Pranitha Mandali, Sai Charan Reddy Obiliachigari
KEYWORDS:
Glaucoma AI, EyePACS-AIROGS-Light-V2, Uncertainty Quantification, Clinical Translation, Deep Learning, Ophthalmology
JOURNAL NAME:
Open Journal of Ophthalmology,
Vol.16 No.2,
March
24,
2026
ABSTRACT: One of the biggest causes of permanent blindness in the world today is glaucoma, and effective treatment depends on early detection. Although artificial intelligence (AI) has demonstrated potential in automating glaucoma screening, there is still a significant obstacle in transferring research datasets to actual clinical settings. When applied to clinical data, current models show an 8% - 18% performance loss, which is mostly caused by out-of-distribution samples, demographic bias, and poor imaging quality. We provide a clinically-translatable glaucoma detection paradigm that makes use of multi-modal fusion, uncertainty quantification, and the EyePACS-AIROGS-light-V2 dataset in order to overcome these difficulties. Our method, called H-UQ-MFF (Hybrid Uncertainty-Aware Multi-Feature Fusion), combines structural and texture characteristics from optic disc analysis with deep features from ResNet50 while dynamically weighting contributions according to prediction uncertainty. With an AUC of 0.9969, sensitivity of 0.9811, and specificity of 0.9717, internal validation outperforms ResNet50, EfficientNet-B0, and Deep Ensemble baselines. Generalizability is confirmed by external validation on REFUGE and PAPILA datasets, where H-UQ-MFF outperforms cutting-edge models and lowers calibration error. Beyond technical performance, the framework ensures clinical safety and regulatory preparedness by incorporating drift tracking techniques, bias analysis, and FDA-compliant evaluation processes. This work bridges the gap between research innovation and clinical deployment by establishing a repeatable baseline for AI translation in ophthalmology.