ARVO 2026: AI screening for psychological distress in glaucoma clinics

Sam Burchuck, PhD, an assistant professor in biostatistics and bioinformatics at Duke University, discussed sequential screening for psychological distress in glaucoma clinics using an AI-assisted method at ARVO 2026.

He began by emphasizing that psychological distress is highly prevalent in glaucoma, with some studies reporting over 40% of patients experiencing clinically meaningful symptoms, including anxiety, depression, and fears of going blind. This distress is not just a parallel issue; it is strongly linked to key glaucoma outcomes such as worse medication adherence, poor follow-up adherence (missed appointments), reduced vision-related quality of life, increased disease progression, and higher healthcare costs. Despite its importance, routine distress screening is rarely done in ophthalmology clinics, primarily due to time, resource, and workflow constraints in already busy environments.

Burchuck then connected this work to a broader history of distress screening in medicine, referencing the American Heart Association’s 2008 guidelines recommending two-stage depression screening in cardiology clinics: initial screening with the PHQ-2, followed by the PHQ-9 for those who screen positive. Even that relatively streamlined approach is often impractical in ophthalmology clinics, which motivated his team to explore whether AI and EHR data could automate the first stage.

He described earlier work using the Duke Ophthalmic Registry (40,000+ general ophthalmology patients), where their AI framework predicted an EHR-based distress phenotype with strong performance. However, that earlier outcome was not based on patient-reported measures.

The current study addresses this limitation through a prospective cohort of 300 patients with primary open-angle glaucoma, where patient-reported distress was collected using the PHQ-8, a validated gold-standard instrument. They implemented a sequential approach:

1. Use an AI model applied to rich EHR data (diagnoses, procedures, medications, labs, problem lists, clinical notes, etc.) to screen all patients automatically.
2. Administer a brief screener (e.g., PHQ-2) only to a high-risk subset flagged by the model.

This approach reduces manual screening burden to about one in four patients, while improving key test characteristics compared with universal brief screening. Specifically, the positive predictive value (PPV) increased by nearly 33% relative to manual universal screening, meaning that a positive screen is much more likely to reflect true distress.

Burchuck also noted that the model identifies meaningful predictors such as prior mental health treatment (ie, therapy, antidepressants) and more general systemic factors like headache/migraine and cancer diagnoses, reflecting broader correlates of distress. Importantly, he frames the clinical value of this AI-assisted screening as enabling targeted behavioral health interventions that could improve adherence, potentially impact intraocular pressure (via stress reduction), and ultimately reduce glaucoma progression risk.