Large health datasets that contain diverse data from large, heterogeneous populations can provide increased statistical power and potentially allow for more complex analyses. However, this increased statistical power is a double-edged sword, as it becomes possible to find statistically significant results that have a small effect size, and worse may be driven by confounds or systematic errors. In contrast, smaller datasets allow for detailed, manual data handling and specific experimental designs, offering precision and targeted insights. In this study, we employed a methodology that leveraged the strengths and mitigated the limitations of large datasets. This enabled us to find and explore relationships that could not be found in smaller datasets.

Specifically, our group takes a novel approach to understanding glaucoma: through the lens of brain imaging. Glaucoma is a common eye condition where increased intraocular pressure leads to optic nerve damage and potential vision loss. Existing research focuses on the resulting changes in the eye itself. However, this research aims to explore how glaucoma might also affect the brain, particularly the pathways involved in processing visual information.

Leveraging data from the UK Biobank, which includes health information from half a million people, we examined the brain scans of 905 individuals diagnosed with glaucoma. We compared these scans with those from 5,292 healthy individuals, using diffusion MRI. This imaging modality gives clues about the structure and health of the white matter of the brain. The white matter is the part of the brain which consists of bundles of myelinated axons that connect different brain regions. We specifically focus on a part of the white matter called the optic radiations, which is not directly affected by glaucoma, but is downstream in the visual pathway.

We leveraged our large control population to find a group of healthy participants that matched the glaucoma patients in several important ways: age, sex, ethnicity, and socioeconomic status. This careful matching helps ensure that any differences found in the brain scans are more likely due to glaucoma and not other unrelated factors. After this matching, we used machine learning to see if we could predict glaucoma status from the optic radiation tissue properties. Specifically, we extracted one-dimensional profiles of tissue properties along the optic radiations, then used a 1-dimensional convolutional neural network (CNN) with these profiles as input to predict glaucoma status. We found that the CNNs were better at identifying people with glaucoma when they analyzed the optic radiations compared to when they used profiles from other, non-visual parts of the brain. This finding is significant because it suggests that glaucoma might have a specific impact on the visual white matter.

We also discovered that linear regression models were unable to find these changes, suggesting they could be non-linear. Interestingly, even the neural net based classification models developed in the study did not effectively predict age-related conditions or age-related macular degeneration, pointing to a specificity in the way glaucoma affects optic radiation tissues.

Previously, studying glaucoma with diffusion MRI was limited to examining available subjects (both tests and controls) from carefully selected nearby populations. In this paper, our statistical matching approach allows us to assemble a dataset that has many of the benefits of large datasets (such as a large number of subjects and a potentially more heterogeneous population) while retaining some of the benefits of collecting subjects oneself (a matched control and subject pool). Of course, we are still limited in our choice of subjects and diseases. The prevalence of glaucoma is what allowed us to construct such a large dataset, and we could not do this with less prevalent diseases.

The increased number of subjects also enables the use of sophisticated machine learning models like CNNs. Our research indicates that glaucoma's impact on the optic radiations presents as a non-linear change in tissue properties, a phenomenon that linear models failed to detect but which CNNs could identify. Furthermore, our generalization analysis (where we compared the effects of glaucoma to the effects of age and age-related macular degeneration) was only possible because our large dataset also contained subjects from those populations.

This methodology not only provides valuable insights into the relationship between glaucoma and brain structure but also opens avenues for further exploration into the effects of diseases that impair sensory input on brain aging. We demonstrate that, through the use of neuroimaging, large datasets, statistical matching, and machine learning in concert, we can open up new avenues for understanding diseases, not just as an isolated issue but as a part of the body's interconnected systems. This approach could be further expanded to other sensory systems or other diseases, given appropriate datasets. Taken together, this demonstrates the utility of large datasets for finding complex phenomena in heterogeneous populations. We note that increased statistical power is not the only advantage the large datasets provide, and in fact, that this increased statistical power can even be a problem. Instead, careful handling of large datasets allows us to be both more confident, and more constrained, in our interpretations of the effects we observe.