Scientists at Google and health technology company Verily, a Google subsidiary, have developed a new way to derive data from scanning the back of the eye that can predict the risk of cardiovascular diseases such as a heart attack.
By scanning the retina, data such as age, gender, blood pressure and whether the patient is a smoker can be derived. This can then be used to predict the risk of certain cardiovascular diseases. This is a picture of the many blood vessels at the back of the eye.
The study used, among other things, two retinal photos of two different patients, where one patient developed a heart problem in the five years after the photos were taken. Google’s algorithm was able to attribute the photo to the correct patient in 70 percent of the cases, without using other data points. A doctor from the University of Adelaide told The Verge that it is a solid study that makes it possible to get more out of the available data than has been possible until now.
According to the researchers at Verily, one of the interesting aspects of this research is that heatmaps have been created that highlight certain areas of the retina that have contributed most to the algorithm’s predictions. The researchers argue that this could provide new insights and give physicians more confidence in the neural network model.
Google’s deep learning models are trained on data from more than 284,000 patients. The data for these analyzes included not only eye scans but also general medical data. Patterns were recognized through a neural network, whereby data from the eye scans could be associated with risk factors for cardiovascular problems. Validated based on two independent datasets of 12,026 and 999 patients, according to Google.
The researchers do acknowledge that their study has a number of limitations. First of all, only images with an angle of view of 45 degrees were used. In addition, the scientists say that the size of the dataset used is relatively small for deep learning. Also, certain important data for cardiovascular disease, such as the amount of lipids in a person’s blood, were not included in the datasets used. In addition, the data on whether or not someone smoked came from the patient themselves, so they may not be completely reliable.
The research has been published in the scientific journal Nature Biomedical Engineering, under the title Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning.
An image of the same retina, with the photo on the left being the default image. The second image shows a heat map with the blood vessels highlighted in green that the algorithm used to predict high blood pressure.