Validating a Deep Learning Algorithm to Identify Patients with Glaucoma using Systemic Electronic Health Records

Researchers at Stanford Medicine led by John Xiang, Rohith Ravindranath, and Sophia Y. Wang validated a deep learning model—the glaucoma risk assessment (GRA)—trained on the national All of Us dataset, then fine-tuned on 20,636 Stanford patients (15% with glaucoma) seen between November 2013 and January 2024. The model uses only systemic EHR data: demographics, diagnoses, medications, lab results, and physical exam measurements—no eye imaging required. On a held-out test set, the best configuration achieved an AUROC of 0.883 and a positive predictive value (PPV) of 0.657. Calibration aligned closely with clinical risk: patients in the highest prediction decile had a 65.7% glaucoma diagnosis rate and a 57.0% treatment rate. Performance improved with up to 15 trainable layers and larger data volumes.

The study, submitted to the AMIA Annual Symposium 2026 and posted on arXiv (2604.20921), highlights a scalable, accessible pre-screening method that could reduce reliance on specialized ophthalmic equipment. By flagging high-risk patients from routine EHR data, the GRA model could enable earlier referral and intervention, particularly in underserved or primary care settings. The authors note that while the model's PPV is moderate, its calibration and AUROC suggest strong discriminative power for population-level screening. Future work may explore integration into clinical workflows and validation across diverse health systems.