Computable phenotyping for primary open-angle glaucoma using electronic health records in the Million Veteran Program

Abstract

Purpose : Primary open-angle glaucoma (POAG), a complex disorder leading to vision loss and blindness, has a substantial genetic component yet relatively few identified genetic associations. Here, we aim to identify POAG cases and controls for genome-wide association studies (GWAS) in the diverse Veteran’s Affairs (VA) Million Veteran Program (MVP). The MVP provides array data in tandem with electronic health records (EHR) and has to date enrolled more than 850,000 participants; upwards of 600,000 have been genotyped on a custom Affymetrix Axiom Biobank array. We developed a multi-stage, rules-based algorithm targeting structured data contained within EHR and validated at two study sites.

Methods : We developed an initial algorithm utilizing structured (ICD-9-CM; ICD-10-CM; CPT) codes in the EHR. We developed test sets of cases and controls and iteratively refined the algorithm based on detailed chart and imaging reviews by eye providers. The current algorithm requires POAG cases to have at least six POAG diagnosis codes received at (i) 30+ years of age, (ii) separate eye clinic visits, and (iii) during eye clinic visits only; POAG cases must also have medication codes for at least three separate intraocular pressure (IOP) lowering prescriptions, and be free from diagnostic codes for conditions that can cause other types of optic neuropathy, visual field defects, or secondary glaucomas, and be free from all exclusion codes. We required controls to have at least two eye clinic visits without POAG diagnosis codes or any exclusion codes.

Results : Chart reviews of 75 cases and 88 controls of at two VA Medical Centers (Cleveland, OH Providence, RI) indicate that the current algorithm performs with a positive predictive value of 85% and a negative predictive value of 100%. Applying this algorithm to the current MVP dataset identifies 18,975 cases and 195,624 controls.

Conclusions : POAG cases and controls can be identified from the VA EHR using diagnostic codes in combination with medication usage. Accuracy was reduced among cases due to several having physiological cupping, or suspicious optic nerves without evidence of glaucoma. Physiological cupping is a clinical term without a corresponding ICD-9-CM or ICD-10-CM, therefore we plan to add natural language processing to improve case accuracy.

This is a 2020 ARVO Annual Meeting abstract.