Abstract
Purpose :
The US Preventive Services Task Force has highlighted a critical need for research to create algorithms identifying high-risk glaucoma patients who might benefit from screening. This study aimed to develop AI models using electronic health records (EHR) data to identify high-risk glaucoma patients, facilitating the identification of the most high-risk candidates for glaucoma screening.
Methods :
We identified all participants with at least one ophthalmic diagnosis in the All of Us Research Program, a national multicenter cohort of patients contributing electronic health records (EHR) and survey data. Participants were divided into those who were diagnosed with glaucoma and those who did not have glaucoma. To become a tool for identifying glaucoma risk in patients without prior ophthalmic data, the model was constrained to use demographic and non-ophthalmic EHR inputs, encompassing diagnoses, medications, physical exams, and basic laboratory test results. We developed models predicting whether participants developed glaucoma using the following approaches: 1) penalized logistic regression; 2) XGBoost, and 3) a custom 1D-CNN with stacked autoencoders to create dense feature matrices for diagnoses and medications. Evaluation metrics included area under the receiver operating characteristic curve (AUROC) and balanced accuracy on a held-out test set.
Results :
Of 64,735 participants, 7,268 (11.31%) were diagnosed with glaucoma. Mean age was 63.0yrs, and 39,913 (61.7%) participants were female. Overall, the AUROC ranged from 0.69-0.87, with 1D-CNN performing the best with an AUROC of 0.87 and a balanced accuracy of 75.4%. This model also has an AUROC range of 0.80 - 0.89 across stratified groups of race/ethnicity. Explainability analyses suggested that age, race, heart rate, body mass index, and hemoglobin A1c contributed most to the model predictions.
Conclusions :
We developed models to predict glaucoma risk using non-ophthalmic EHR and demographic data. Possible future uses include identifying patients without prior ophthalmic care who would benefit most from glaucoma screening. Further research is needed to ensure that such models are fair and trustworthy across different demographic subpopulations.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.