The retina, a window to the heart

Retinal fundus images – photographs of the back of the eye – have traditionally been used to monitor the progression of eye diseases. Increasing research, however, suggests that the network of blood vessels visible in fundus images may be useful for evaluating and even clinically diagnosing other complications such as cardiovascular health.

In a study published in Nature Biomedical Engineering1 early this year, researchers from Google and Verily Life Sciences used deep learning, a form of artificial intelligence that uses algorithms to “learn” from available data, to train computer models using fundus images and variables relevant to cardiovascular disease from close to 300,000 patients. They then used these models to make predictions using only fundus images from another 13,000 patients.

Their models reliably predicted the age, gender, smoking status, systolic blood pressure, and body mass index of the 13,000 patients. The researchers noted that these are major variables used in several cardiovascular risk calculators, but cannot necessarily be determined from retinal images themselves.

The researchers then trained a model to predict patients’ risk of major adverse cardiovascular events in the next five years. Using only fundus images, the model predicted this outcome with 70% accuracy, which is comparable to a standard risk calculator that uses data from various sources. The researchers took it one step further and examined how the models made these predictions. Encouragingly, they found that the models used blood vessels for predicting risk factors such as age, smoking status and systolic blood pressure, and general information across the image to predict other variables like body mass index.

Although their findings are promising, the researchers caution that their models need further validation using larger and more varied datasets.

The requirement of only a fundus photograph, which is non-invasive, cheap and quick, for identifying or predicting cardiovascular disease, the leading cause of death globally, shows the promise of this emerging research. Similar deep learning methods are being applied to an increasing number of medical conditions, including cancer diagnosis2, and is exciting for the future of diagnostics and precision medicine.

Reference
1Poplin et al. Predicting Cardiovascular Risk Factors from Retinal Fundus Photographs using Deep Learning. Nature Biomedical Engineering. 2(3), 158 (2018).
(https://www.nature.com/articles/s41551-018-0195-0.epdf?author_access_token=YWBi0EzCgfAVb_S540xl-tRgN0jAjWel9jnR3ZoTv0OMsbBDq-7d5VZef-dAA8S4kHGY_hXONc93gwXXjuO908b_ruUDVkgB5jW3RnvvRdLFLmvpTsPku5cXZoTEtr09fPvTK40ZbWzpoOGfLab-NA%3D%3D)

2Esteva et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 42, 115–118 (2017)
(https://www.nature.com/articles/nature21056#figures)