Tobias Elze - US grants

Glaucoma is a complex neurodegenerative disease that results in degeneration of retinal ganglion cells and their axons. With older people making up the fastest growing part of the US population, glaucoma will become even more prevalent in the US in the coming decades. Due to the complex interaction of multiple factors in glaucoma, better structural and functional predictors are needed for its progression. The main impediments are massive health record data and sophisticated computational models. Our overall goal is to leverage the power of big data and rapidly evolving machine learning approaches. The NEI's ?Big Data to Knowledge (BD2K)? initiative and the American Academy of Ophthalmology Intelligent Research in Sight (IRIS) registry are all efforts to exploit the power of data and to better understand diseases and to provide improved prevention and treatment. In this multi-PI proposal, we offer to assemble over 1 million optical coherence tomography (OCT) and visual fields (VFs) from the glaucoma research network (GRN). We propose to develop a hybrid artificial intelligence (AI) algorithm that synthesizes Gaussian mixture model expectation maximization (GEM) and archetypal machine learning approach to identify glaucoma progression and its monitoring using VFs and retinal nerve fiber layer (RNFL) thickness measurements. We will make these tools openly available to the vision and ophthalmology research communities. Our proposed studies could offer substantial improvements in the prognosis of glaucoma as well as potentially providing OCT and joint VF/OCT surrogate endpoints to be used in glaucoma clinical trials.

Project Summary/Abstract Type 2 diabetes mellitus (T2DM), a metabolic disease that affects over 300 million people worldwide and that can be accompanied by serious health complications such as heart disease, kidney failure, stroke, and damage to the eyes, in particular diabetic retinopathy (DR), which is diagnosed in a third of people with diabetes and which is the leading cause of blindness within the age group between 20 and 64 years. T2DM is clinically diagnosed by parameters related to glucose metabolism obtained by blood tests. Due to its long pre- symptomatic phase, an estimate of 25% of diabetics in the US are undiagnosed. In this project, the relationship between spatial patterns of retinal nerve fiber layer (RNFL) thickness (RNFLT), measured by spectral-domain optical coherence tomography (OCT), and blood test levels as well as levels of DR severity is investigated in 9,261 participants of a population based study. In a first step, OCT RNFLT measurements of the macular and the circumpapillary area around optic nerve head are segmented into spatial sectors, and representative spatial patterns of RNFLT are calculated by an unsupervised machine learning method. Afterwards, a multivariate linear model comparison is performed with the coefficients of the spatial RNFLT patterns as regressors and diagnostic blood test results as dependent variable. The optimal combination of the RNFLT patterns, determined by an established model selection criterion (Bayes Factor), is expected to reveal insight into the association between the specific retinal locations of RNFL thinning accompanying the change in parameters related glucose metabolism during the development and progression of T2DM. Furthermore, fundus images are graded by DR severity following a nine-step scale derived from the Early Treatment Diabetic Retinopathy Study from no DR to severe proliferative DR. The spatial RNFLT patterns and metabolic blood test scores are then compared with respect to modeling DR severity by linear regression. An optimal model of DR severity combining glucose metabolism parameters and RNFLT patterns is developed. Finally, in an analogous procedure, DR severity of the follow-up measurement, five years after baseline, is statistically predicted from RNFLT and metabolic blood parameters and from their change over time. To summarize, the proposed research identifies spatial patterns of RNFLT associated with parameters of glucose metabolism and their development over DR severity. Once accomplished, the proposed project would provide the details to establish RNFLT as an alternative manifestation of T2DM that complements diagnostic blood tests and thereby, for instance, lay the foundations for the development of novel and more accurate T2DM progression monitoring or the prediction of the onset of DR.

Project Summary/Abstract Glaucoma is a disease of the optic nerve which is accompanied by visual ?eld (VF) loss. While accurate VF loss diagnosis and the detection of its progression over time is of high relevance to clinical practitioners as it indicates the initiation of or change in ocular therapy, there is no consensus on objective measures for this purpose, and VF measurements are known to be often unreliable. The main objective of this project is to develop clinically applicable measures to improve the diagnosis of glaucomatous VF loss and of its progression by two approaches: First, the identi?cation of representative loss patterns and their progression, achieved by large-scale, customized bioinformatical procedures applied to data from glaucoma patients from nine clinical centers and second, the inclusion of eye and patient speci?c personalized parameters. In total, 480,486 VFs, are available for this project. One major aim is to develop novel diagnostic indices based on computationally identi?ed evolution patterns of VF loss, particularly (1) an index that denotes the probability of glaucomatous vision loss and (2) an index that assigns probabilities to a VF that follow-up measurements will be in a certain defect class. The indices will be statistically evaluated on separate VF samples and compared to existing approaches. Routinely available patient speci?c parameters which are candidates to impact glaucomatous vision loss are patient ethnicity, type of glaucoma, spherical equivalent (SE) of refractive error and the location of the blind spot relative to ?xation. The effect of these parameters on the vision loss patterns will be systematically studied. The impact of their inclusion in the novel diagnostic indices and their potential improvement on glaucoma diagnosis will be quanti?ed on a separate data set. A further aim is the calculation of a spatial map speci?c to a measured VF that represents the preferred VF locations of future defects as well as their reliability as an aid to event-based progression diagnosis. A second major objective is the investigation of the relationship of VF loss and individual parameters related to retinal structure, based on retinal nerve ?ber layer thickness (RNFLT) measurements around the optic disc. The inter-relationship of representative patterns of RNFLT and its decrease over time with trajectories of major retinal arteries, SE, and blind spot location is systematically studied, and the impact on patterns of VF loss is quantitatively analyzed with the goal to improve the interpretation of existing VF loss and to predict future glaucomatous vision loss. Main contributions of the project with relevance to clinical practice are publicly available open-source software implementations of new diagnostic indices and maps, enhanced by individual functional and structural parameters, and a detailed and personalized model for the relationship between retinal structure and glaucomatous vision loss.