It’s that age old, dinner party question: If you could know when you were going to die, would you want to know?
Luckily for the morbidly-inclined, it’s quickly becoming an option, with new research recently published by the University of Adelaide in Australia.
A team of radiologists and computer scientists have developed a technique in which AI can – with 69% accuracy – predict when patients would die within a five-year timespan.
That’s a prediction with about the same accuracy that a clinician could make.
“We think the limit is probably around 85% accuracy for the task of predicting mortality in older patients, but it is worth noting that the prediction itself is not the goal,” says Dr Luke Oakden-Rayner, lead investigator on the team.
“Predicting mortality is a way to assess health, and we think this sort of ‘health score’ will be useful, even if we can’t get perfect predictions.”
Making predictions about complex systems – like the human body – will always be limited by randomness, says Oakden-Rayner.
“We see this in clinical practice, as some people with severe diseases live long lives and others who are apparently healthy can unfortunately die.”
All it needs are images of your organs
The team analyzed routinely collected CT chest scans from adults over the age of 60 to predict mortality within a five-year timeframe.
Mortality, says Oakden-Rayner, ‘represents an easily obtained and well-defined outcome.’
In other words: You’re either dead, or you’re not.
Using convolutional neural networks, the team taught the network to recognize known biomarkers indicative of disease, feeding the system more than 16,000 images.
A clinical expert then reviewed mortality predictions made by the network, and compared them to predictions made by radiologists.
Their results suggest that the computer has learnt to identify what disease looks like, and in a much shorter timeframe than is required to teach human experts.
Deep convolutional neural networks are the best approach we currently have at teaching machines to recognize objects in images, but analyzing complex medical imagery presents a new problem.
“Medical images present some difficulties that we don’t see in other areas of machine learning,” says Oakden-Rayner.
“The images are three dimensional, and are often several orders of magnitude larger than the photographs that might be used for more general image analysis tasks.”
Another open issue being faced by computer science is the computation burden of using such large images.
Some research groups have approached this challenge by selecting 2D images from overall volumes, but the team at the University of Adelaide approached it from a different direction.
“In our research we used downsampling to shrink the images, but chose to retain the 3d structure of the data. Other groups have tried other approaches, but it isn’t clear yet which are best for a given clinical task.”
And, to add yet another hurdle, there are strict ethical guidelines directing the design of medical trials.
“We’re not just doing machine learning here. Luckily we have a strong team that contains experts in medical research and epidemiology as well as machine learning. I know that groups made up of computer scientists can struggle with these challenges.”
There are obvious benefits to this technology: earlier diagnosis and treatment, which is often associated with more favorable clinical outcomes; another set of ‘eyes,’ spotting things clinicians may have missed; more tailored approaches to treatment; and insight into the patients longevity, which is, obviously, limited to what doctors are able to see.
But it’s important to note that this tool, when fully developed, is meant to work in conjunction with clinicians – not instead of them.
“Clinicians are not really trained to do this task, so it is something we can offer in addition to normal clinical practice. Clinicians could incorporate this tool into their normal practice to help guide treatment choices.”
Of course, any technology can be misused, and a high priority for the team was managing data privacy.
“I don’t think this is any more problematic with our work than with normal medical data or genomic information, all of which could be used by insurance companies and other businesses to determine risk.”
At the end of the day, the advantages of this innovative new technology and data are enormous, and, in Dr Oakden-Rayner’s opinion, far out-weight the risks.
“We have to remember that the benefits of gathering this sort of data is enormous. Clinical medicine would not function without it. That said, we do need to be careful of these risks, and we rely on appropriate legislation to protect the rights of patients.”
Where to from here?
It’s still a young body of research, but such promising results – despite the small patient numbers – are encouraging.
“We are expanding our research into much larger datasets, with tens of thousands of patients. We also intend to incorporate readily available predictive information such as age and sex.”
Research is already underway, and so far, says Luke, the results are looking good.