Artificial intelligence deployed on Electronic Health Record data: Machine learning methods for identification and referral of at-risk patients from primary care physicians to eye care specialists. (Preprint)

Abstract

Identification and referral of at-risk patients from primary care practitioners (PCPs) to eye care professionals remains a challenge. Approximately 1.9 million Americans suffer from vision loss as a result of undiagnosed or untreated ophthalmic conditions. Artificial intelligence (AI) modeling of 1,486,078 patients identifies individuals at higher risk for five leading vision conditions: glaucoma, age-related macular degeneration, diabetic retinopathy, visually significant cataracts, and ocular surface disease.

To build and compare machine learning (ML) methods, applicable to Electronic Health Records of PCPs, capable of triaging patients for referral to eye care specialists.

Accessing the Optum® de-identified Electronic Health Record dataset, 743,039 patients with age-related macular degeneration (AMD), visually significant cataract, diabetic retinopathy, glaucoma, or ocular surface disease (OSD) were exact matched on age and gender to 743,039 controls without eye conditions. Between 142-182 non-ophthalmic parameters per patient were input into five ML methods: Generalized Linear Model (GLM), L1-regularized logistic regression, random forest, XGBoost, and J-48 decision trees. Model performance was compared for each pathology to select the most predictive algorithm. Area under the curve (AUC) was assessed for all algorithms for each outcome.

XGBoost demonstrated the best performance, showing, respectively, prediction accuracy and AUC of 78.6% and 0.878 for visually significant cataract, 77.4% and 0.858 for exudative AMD, 79.2% and 0.879 for non-exudative AMD, 72.2% and 0.803 for OSD requiring medication, 70.8% and 0.785 for glaucoma, 85.0% and 0.924 for type 1 nonproliferative diabetic retinopathy (NPDR), 82.2% and 0.911 for type 1 proliferative diabetic retinopathy (PDR), 81.3% and 0.891 for type 2 NPDR, and 82.1% and 0.900 for type 2 PDR.

The five ML methods deployed were able to successfully identify patients with elevated odds ratios (ORs [95% CI]), thus capable of patient triage, for ocular pathology ranging from 2.4 [2.4-2.5] (glaucoma) to 5.7 [5.0-6.4] (type 1 NPDR), with an average OR of 3.9. The application of these models could enable PCPs to better identify and triage patients at risk for treatable ophthalmic pathology. Early identification of patients with unrecognized site threatening conditions may lead to earlier treatment and a reduced economic burden. More importantly, such triage may improve patients’ lives.