Artificial Intelligence Models to Identify Patients with High Probability of Glaucoma Using Electronic Health Records

Purpose: Early detection of glaucoma allows for timely treatment to prevent severe vision loss, but screening requires resource-intensive examinations and imaging, which are challenging for large-scale implementation and evaluation. The purpose of this study was to develop artificial intelligence models that can utilize the wealth of data stored in electronic health records (EHRs) to identify patients who have high probability of developing glaucoma, without the use of any dedicated ophthalmic imaging or clinical data. Design: Cohort study. Participants: A total of 64 735 participants who were >18 years of age and had >2 separate encounters with eye- related diagnoses recorded in their EHR records in the All of Us Research Program, a national multicenter cohort of patients contributing EHR and survey data, and who were enrolled from May 1, 2018, to July 1, 2022. Methods: We developed models to predict which patients had a diagnosis of glaucoma, using the following machine learning approaches: (1) penalized logistic regression, (2) XGBoost, and (3) a deep learning architecture that included a 1-dimensional convolutional neural network (1D-CNN) and stacked autoencoders. Model input features included demographics and only the nonophthalmic lab results, measurements, medications, and diagnoses available from structured EHR data. Main Outcome Measures: Evaluation metrics included area under the receiver operating characteristic curve (AUROC). Results: Of 64 735 patients, 7268 (11.22%) had a glaucoma diagnosis. Overall, AUROC ranged from 0.796 to 0.863. The 1D-CNN model achieved the highest performance with an AUROC score of 0.863 (95% confidence interval [CI], 0.862-0.864). Investigation of 1D-CNN model performance stratified by race/ethnicity showed that AUROC ranged from 0.825 to 0.869 by subpopulation, with the highest performance of 0.869 (95% CI, 0.868-0.870) among the non- Hispanic White subpopulation. Conclusions: Machine and deep learning models were able to use the extensive systematic data within EHR to identify individuals with glaucoma, without the need for ophthalmic imaging or clinical data. These models could potentially automate identifying high-risk glaucoma patients in EHRs, aiding targeted screening referrals. Additional research is needed to investigate the impact of protected class characteristics such as race/ethnicity on model performance and fairness.