A neural network developed by Google and Verily sets its sights on predicting cardiovascular disease
Scientists from Google and its subsidiary research organization, Verily Life Sciences, have developed an AI algorithm that could help to assess a person’s risk of developing heart disease by analyzing retina scans.
Ophthalmologists have been able to identify conditions in patients by observing their eyes for eons. This AI seeks to sharpen the process.
It has long been known that the human eye is a window into a wide variety of health conditions, including diabetes and cardiovascular disease. High cholesterol levels or high blood pressure can leave signs on the blood vessels in the retina, helping to identify related medical issues.
Heart disease, also referred to as cardiovascular disease (CVD), includes a number of medical conditions in which patients’ blood vessels are narrowed or blocked, leading to chest pain or strokes.
CVD is the leading cause of death globally and it is vitally important to find ways to identify those who may be affected by it, encouraging early intervention.
But predicting the risk of developing heart disease is not always easy, as multiple factors are generally involved including exercise levels, age, BMI, lifestyle, diet, genetic factors, and much more.
Artificial Intelligence (AI) tools trained to analyze medical images could improve currently employed methods of determining these risks.
In a blog post, Michael V. McConnell, Head of Cardiovascular Health Innovations at Verily said:
“Discovering strong predictive technologies will enable the preventive approach to healthcare that is core to our mission”.
McConnell and other researchers from Google and Verily have recently trained a machine learning algorithm to assess an individual’s risk of developing heart disease by analyzing his/her retinal scans.
Using the eye as a window to the heart
The scientists at Google and Verily trained their algorithm on 284,335 patient images tagged with information that was found to be relevant to heart disease, such as age, smoking status, blood pressure, and BMI.
Once the training was completed, the AI tool was tested on another 12,026 images and the results were published on the journal Nature Biomedical Engineering.
The neural network was able to estimate the risk of developing CVD in patients, with results that matched calculations using heart disease-related factors.
Merely by analyzing retinal fundus images, which show the rear interior wall of the eye and its blood vessels, the algorithm was able to get within 3.26 years of a person’s actual age. It could also predict individuals’ smoking status with 71% accuracy and their blood pressure within 11 units of the highest number of the actual measurement.
Out of the 12,026 images that the AI tool was tested on, 150 were of patients that were known to have suffered from a major cardiovascular condition within five years after the scan was made.
The neural network was presented with two retina images and asked to predict which one indicated a risk of major cardiac event or stroke and it identified the correct scan 70% of the time.
These results are very promising, as tested on the same data set, the European SCORE risk calculator, an assessment tool that is currently used to predict the risk of CVD, predicted the correct scan at a similar rate- 72% of the time.
Said Michael V. McConnell, Head of Cardiovascular Health Innovations at Verily, in a blog post about the study.
“One of the exciting aspects of this study is the generation of “attention maps” to show which aspects of the retina contributed most to the algorithm, thus providing a window into the “black box” often associated with machine learning. This can give clinicians greater confidence in the algorithm, and potentially provide new insights into retinal features not previously associated with cardiovascular risk factors or future risk.”
McConnell highlighted that while these results seem promising, they are still preliminary and more research is necessary before the tool can be introduced within healthcare settings.
A dataset of 300,000 images is quite limited when training a deep-learning algorithm and the researchers feel that access to more data could improve its performance.
“This is promising, but early research,” said McConnell. “More work must be done to develop and validate these findings on larger patient cohorts before this can arrive in clinical settings.”
Approaching the era of AI diagnostic tools
The neural network developed by Google and Verily is merely the latest example of how researchers worldwide are applying machine learning for faster and more accurate diagnoses of medical conditions.
Google, one of the pioneers in this field, has been working on AI diagnostic tools for some years now.
Just a few weeks ago, we published an article about an algorithm developed by Google DeepMind in collaboration with Moorfields Eye Hospital in London, which was trained to diagnose severe eye conditions by analyzing retinal scans.
The possibility that AI will speed up diagnoses and help to identify the risks of developing health conditions in future is becoming increasingly plausible.
By analyzing large sets of data at speeds that humans could never achieve, machine learning tools could revolutionize a variety of scientific practices.
AI may soon become a key tool for scientific research, development, and discovery, assisting scientists worldwide in their daily tasks and simplifying the work of many healthcare professionals.