World first collaboration uses machine learning to enable faster and more accurate analysis of how genes impact glaucoma, a globally debilitating eye disease.
Glaucoma occurs when pressure from fluid in the human eye builds up over time, causing damage to the optic nerve and eventually leading to partial or full vision loss. Affecting more than 60 million people worldwide and 300,000 Australians, glaucoma is expected to cost the Australian economy more than $870 million by 2025.
In collaboration with researchers from QIMR Berghofer Medical Research Institute, Max Kelsen applied machine learning to automatically analyse extremely large eye imaging data sets to detect glaucoma.
Prior to working with Max Kelsen, the glaucoma research team at QIMR Berghofer had already identified about 100 genes considered likely to play a role in the predisposition for glaucoma with a sample of around 70,000 individuals. However, this work was done ‘manually’ and painstakingly over a number of months. To more accurately detect which individuals are at risk, QIMR Berghofer’s researchers realised they would need a much larger sample size.
Such large sample sizes were available, but the work required to correlate neural deterioration with DNA would have taken hundreds of hours and would have been prone to estimation error derived from fatigue and insatiable image repetition, particularly when scoring large volumes of images.
To overcome this limitation, Max Kelsen developed machine learning algorithms automating the process of clinical rating in a cohort of 280,000 images of eyes with various states of the disease from around the world.
Working closely in an iterative approach with QIMR Berghofer experts, Max Kelsen’s team built and trained an algorithm over six weeks to accurately measure the level of deterioration in a human eye - using a standard system of measurement on a scale from 0 to 1, with 0.7 or above indicating glaucoma.
By applying machine learning, Max Kelsen researchers can now deliver clinical ratings for hundreds of thousands of images in less than an hour. They have also substantially reduced the estimation error of these ratings. As a consequence, Max Kelsen researchers enabled the QIMR Berghofer team to increase genetic discovery from the previously identified cohort of 100 ‘risk increasing’ genes to over 200.
This collaboration has opened up new paths for researchers around the world to apply machine learning to multiple ethnic and longitudinal studies, generating potential for increased genetic discovery of this complex disorder. This demonstrates the real world, practical ways in which Max Kelsen can apply machine learning to enable precision medicine.
