Computational Investigation of Glaucoma Progression

Tobias Elze, PhD The Schepens Eye Research Institute, Massachusetts Eye & Ear, Harvard Medical School


Michael Boland, MD, PhD Johns Hopkins University
Louis R Pasquale, MD Massachusetts Eye and Ear, Harvard Medical School
Lucy Q Shen, MD Massachusetts Eye and Ear, Harvard Medical School
Gadi Wollenstein, MD University of Pittsburgh


Glaucoma typically is accompanied by functional vision loss, called visual field defects, which may progress over time. That’s the reason why visual fields of glaucoma patients are regularly measured; however, these measurements are influenced by a number of factors and it is often hard for clinicians to decide whether changes over time reflect true changes of functional vision or are the result of normal measurement variations or artifacts (ie, one-time results). The aim of this project is to investigate the spatial configuration of glaucomatous visual field defects by a combination of mathematical algorithms and clinical expertise in order to identify patterns of disease progression. The resulting quantitative models will be implemented as a software which calculates for a given patient measurement the probability of if, how, and how quickly vision loss is expected to progress.

Project Details

In this project, we identify representative patterns of the progression of glaucomatous vision loss and of glaucomatous nerve fiber damage by computational statistical learning procedures.

The first part of the project focuses on the problem that many glaucoma patients develop initially, which are small scotomas ("blind spots" in the field of vision) that may enlarge over time up to the stage of total blindness. However, the onset of individual vision loss, and the respective course of progression differs considerably from patient to patient. This substantial inter-individual variety makes the patient-specific prediction of vision loss progression extremely challenging. In previous studies, our lab combined a bioinformatical machine learning procedure (archetypal analysis) with ophthalmic background knowledge and identified representative patterns ("archetypes") of glaucomatous vision loss. As a first step of this project, we identify how these patterns change over time and determine the archetypes of vision loss progression.

In the second part of this project, we apply machine learning algorithms to determine the patterns of glaucomatous damage to the retina, in particular the thinning of the retinal nerve fiber layer. Subsequently, we relate these nerve fiber damage patterns to the patterns of vision loss progression by multivariate statistical procedures and develop a structure-function model of glaucoma progression. Finally, we implement this structure-function model as a software that reads retinal imaging scans and visual field measurements of patients and assists clinicians with diagnosis and prognosis of glaucoma.

While many previous studies investigated structure-function relationships in glaucoma and the progression of the disease, to date, the existing models are limited by the complexity of the different subtypes of the disease, each of which must be studied using dedicated bioinformatical methodology. The members of our lab combine skills from engineering and computer science with specific expertise in ophthalmology. Therefore, our group has the skills to process extraordinarily large and high-dimensional data sets to attack the problem of glaucoma progression at a unique level of detail in the course of this project.

Glaucoma is one of the leading causes of blindness worldwide. Once accomplished, this project will provide a clinical assistance software to predict patient-specific disease progression. We believe that our structure-function based software will help practitioners with clinical decisions about initiation or changes in ocular therapy. Thereby, this project may help to prevent the onset of functional vision loss, or progression of glaucoma, which poses a major threat of vision loss as a potentially blinding eye disease.