Predictive Analytics for Glaucoma using Data from the All of Us Research Program

Abstract

Purpose : The All of Us Research Program is a nationwide prospective cohort study. As the only alpha phase demonstration project related to ophthalmology, our aims were to (1) externally validate a previously published model predicting need for surgery among individuals with glaucoma, (2) develop new models using All of Us data, and (3) share insights regarding the use of this data for ophthalmic research.

Methods : Electronic health record data were extracted for 1231 adult participants in All of Us diagnosed with primary-open angle glaucoma. We compared a previously published single-center cohort and the All ofUs cohort with respect to demographics and need for glaucoma surgery. The performance of the single-center model was evaluated on All of Us data, based on area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. All of Us data were then used to train new models using multivariable logistic regression (LR), artificial neural networks (ANN), and random forests (RF). These were cross-validated on All of Us data. Performance was evaluated based on AUC, accuracy, precision, and recall.

Results : The mean (standard deviation) age of the All of Us cohort was 69.1 (10.5) years, with 57.3% women and 33.5% Black or African American, both significantly exceeding representation in the single-center cohort (p=0.04 and p<0.001, respectively). Of 1231 participants, 286 (23.2%) needed glaucoma surgery. When applying the single-center model to All of Us data, accuracy was 0.69, and AUC was 0.49, indicating that the model was not generalizable to All of Us data. Using All of Us data to train new models resulted in superior performance: AUCs ranged from 0.80 (LR) to 0.99 (RF). Blood pressure (BP) had the highest relative importance for driving predictions in RF models.

Conclusions : Models trained with national All of Us data achieved superior performance compared to using single-center data. The relationship between BP and glaucoma warrants ongoing investigation. Novel big-data sources such as All of Us offer numerous opportunities for ophthalmic research.

This is a 2021 ARVO Annual Meeting abstract.

 

Figure 1. Composite receiver operating characteristic curves for multivariable LR model (area under the curve, AUC=0.80), ANN (AUC=0.93), and RF (AUC=0.99) predicting need for surgery among All of Us research participants with glaucoma.

 

Figure 2. Strengths and limitations of All of Us for ophthalmic research.