Google parent company Alphabet claims that one of the technology initiatives under its X experimental research divisions can accurately predict heart disease simply by scanning a person's eyes.
It claims that its technology can evaluate scans of the back of the eye and accurately deduce a person's risk of cardiovascular problems against other data, including their age, blood pressure and gender.
The AI algorithms need a good dose of machine learning before they can do this well, though, which involved analysing a medical dataset of around 300,000 patients and crunching both eye scans and more general medical data.
Deep-learning neural networks of the type implemented by Verily in this project essentially pick apart data similar to, but much faster then, human brains. They can identify patterns in otherwise disparate data sets, in this case potentially helping to identify indicators of cardiovascular problems.
Once the system is perfected, Verily claims that the results could be as accurate as current blood testing methods. And scanning the eye is also a lot easier than extracting blood, especially if the patient has an aversion to needles.
There are limitations to the AI tech, though. It was only trained on images with a 45-degree field of view, meaning the AI model would need to be retrained for other eye images, and larger datasets would be needed before such an AI system is ready for use in clinical trials.
And, while trained on a large sample of people with cardiovascular problems there is also a question mark over the level of false positives it may generate.
The work undertaken by Verily is similar to the project's that Google's DeepMind division has sought to pursue in the UK, but with many campaigners questioning whether the health records used can and will remain private.
Use the same password for every website? It might be time to change them all
Applicants for parking bay suspensions put at risk of credit card fraud by Islington Council
Robert Swan appointed interim CEO after Brian Krzanich's departure
Should you link your data sets to add value, or leave them separate to reduce risk?