Abstract
Purpose :
Glaucoma is a chronic, degenerative optic neuropathy, and the leading cause of irreversible blindness. It affects about 70-90 million people worldwide. Primary open-angle glaucoma is the most common form of glaucoma. At present, there is no cure for glaucoma. Early detection and treatment are crucial for preventing vision loss from this blinding disease. Polygenic risk scores (PRSs) enable disease early prediction based on an individual’s genetic makeup. We constructed PRSs and utilized advanced machine learning for glaucoma prediction.
Methods :
We conducted this study using participants from the UK Biobank cohort, a population-based prospective study of 500,000 individuals. All study subjects were 40 years of age and older. The participants were genotyped and further imputed to 92 million genetic markers. About 118,000 study subjects participated in the eye and vision component of the study, where numerous ocular measurements were obtained, including intraocular pressure (IOP). Glaucoma information was extracted from electronic medical records and self-reported medical history (cases n = 13,712). We constructed PRSs for multiple diseases/traits, such as glaucoma and IOP, and evaluated their predictive power using XGBoost (an efficient machine learning algorithm) and cross-validation. To avoid overfitting in predictive models, we used independent datasets for model training and testing.
Results :
The glaucoma PRS was significantly associated with glaucoma disease risk (P = 4.69 x 10-24) in our testing dataset, after adjusting for covariates, such as age and sex. Subjects in the top 10% of PRSs were 10.28 times more likely to have glaucoma, compared to those in the bottom 10% category (P = 7.5 x 10-27). The area under the receiver operating characteristic curve (AUC) for glaucoma risk prediction was 0.82. Furthermore, the XGBoost method gave better prediction accuracy than traditional logistic regression.
Conclusions :
We determined that an ensemble of PRSs and advanced machine learning improve the prediction of glaucoma.
This is a 2021 ARVO Annual Meeting abstract.