Databases as Tools in Ophthalmic Research: Examples and a Machine-Learning Advancement

Abstract

Databases are an increasingly important and frequent tool for research in Ophthalmology. Carefully considering potential sources of bias and appropriate methodology is paramount to drawing accurate conclusions. This thesis aims to demonstrate examples of study designs applied to databases in Ophthalmology and the potential for machine-learning extension. We present the analysis of a small academic epiphora database, the large academic UK Biobank database, and a large national administrative database of vitreoretinal procedures. In addition, we demonstrate the utility of a low-code named entity recognition workflow for constructing an ophthalmic disease registry from free-text electronic clinical records. Using the small academic epiphora database, we examined the correlation of dacryocystography (DCG) and dacryoscintigraphy (DSG) findings in fellow asymptomatic eyes. We found a high rate of DSG abnormalities compared to DCG in asymptomatic eyes. This high rate has important implications for using control eyes in lacrimal imaging studies of functional epiphora. In the UK Biobank, we found systolic blood pressure and pulse pressure were associated with incident primary open-angle glaucoma. In the administrative database study, we found populationwide decreases in the rates of scleral buckle use and increases in rates of vitrectomy for retinal detachment repair in Australia. Finally, we created a machine learning registry database of ophthalmic diseases from free-text electronic clinical records. In conclusion, study designs must be adapted to the structure of pre-existing databases when used for research. This approach contrasts with conventional prospective data collection and requires careful consideration of bias and limitations when designing analyses and interpreting results.