Principal Investigator: Dr Benjamin Goudey
Department: IBM Australia LimitedTags: 51064, deep learning, differential diagnosis, Genomic risk prediction, polygenic risk score, retinal-imaging
It is increasingly understood that many neurodegenerative conditions begin decades before cognitive symptoms appear. If these conditions could be detected earlier, before neurodegeneration has occurred, there may be greater chance of successful intervention to slow down or stop disease progression. Here, we focus on a broad definition of neurodegeneration that not only encompasses dementia and cognitive decline related conditions, but also consider conditions of the eye including age-macular degeneration (AMD), glaucoma and diabetic retinopathy.
This project seeks to develop predictive models of neurodegeneration related phenotypes through the combined analysis of genomic data and retinal imaging. Historically, these two distinct types of data have been analysed in a separately with little analysis exploring their combined utility, due to the difficulty in bringing together teams with the skills and domain knowledge to best make use of these modalities. Recent analyses of brain imaging and genomic data showed the impact that such joint analysis can bring (Elliott et al, Nature 2018).
While there have been genome wide association studies conducted looking at the association of genetic and neurodegenerative disease. As well as preliminary work exploring the link between retinal imaging and brain-related phenotypes. There has been very little work exploring how the combination of these modalities can be used to detect disease, especially when considering detection of multiple conditions. We believe that joint analysis of retinal imaging and genomics can help increase our ability to distinguish diseases that may present similarly in the eye or may have similar genomic risk profiles.
The large number of individuals available in the UK Biobank also provides an opportunity to explore the utility of deep learning to this task. While deep learning has shown great utility for imaging modalities, there remains little exploration of how it can be used to combine genomic and imaging modalities together.
Given the large number of data modalities and the range of modelling options, we envisage that will take 12 months to develop, with a range of ongoing explorations in model robustness and understanding the biological underpinnings for a further 12 months.
The discovery of retinal markers and genomic profiles associated with neurodegenerative diseases, is likely to have two main public benefits. One, in principle, enable cheap and accessible population screening for neurodegenerative diseases (enabling improved health management). Two, enable improved identification of persons for preventive and/or interventional clinical trials.