The project, Medical Advice from Glaucoma Informatics (MAGI), seeks to improve glaucoma diagnosis and management with state-of-the-art machine learning classifiers. These classifiers will automate the interpretation of standard automated perimetry (SAP), newer visual field tests, and structural tests for glaucoma in the general population and in stratified glaucoma populations. Phase 1 will complete the feasibility testing already underway. Phase 2 will apply the refined methods to a wider set of glaucoma testing problems.The management of glaucoma depends on a series of classifications. The glaucoma provider classifies tests as normal or indicative of glaucoma. The clinician then determines whether an eye has glaucoma or has had progression. Assembling these classifications, the provider makes decisions about management. Automated test interpreters, either as part of the testing machine or as a computer-based resource, can aid glaucoma providers with real-time interpretations. The research we propose takes advantage of our extensive data sets and builds on the ongoing research in our laboratories.Statistical classifiers, Bayesian nets, machine learning classifiers, and expert systems represent different types of classifiers with diverse properties. Machine learning classifiers can perform exceptionally well at identifying classes, even when the data are complex and have dependencies. We will test and select the optimal machine learning classifier for diagnosis. We will further improve classifier performance and determine feature utility by optimizing the feature set visual field tests are time consuming and stressful. We will streamline the tests by removing unimportant test points.Even with decades of experience, there is uncertainty with regard to the evaluation of the SAP. There is less accumulated knowledge about non-standard tests, such as short-wavelength automated perimetry, nerve fiber layer thickness, or optic nerve head topography. Machine classifiers may learn how to interpret nonstandard tests better. We will go beyond STATPAC's capabilities with classifiers that have learned to interpret SAP, nonstandard visual field tests, structural glaucoma tests, and STATPAC plots in the general population and in patients stratified by race, family history, and other information available at the time of the test.Conversion of suspects to glaucoma and progression of glaucoma cannot yet be predicted from tests. We will develop classifiers for these predictions. Classifiers will be designed to diagnose early glaucoma, detect early progression, and identify glaucomatous eyes at risk of progression.Unsupervised learning provides cluster analysis that can determine distinct groups with members in some way similar from the test data. In an effort to discover new and use useful information with unsupervised learning, we will mine our data in visual function and structural tests for glaucoma and in specific combinations of population groups.
