Will machine learning replace radiologists

Does big data make medical professionals unemployed?

In the recent debate about the future role of information technology in medicine, Ezekiel Emanuel from the University of Pennsylvania (Philadelphia) has spoken out several times.

First of all, in the "Journal of the American College of Radiology", he prophesied that radiology would end as a prosperous medical discipline (J Am Radiol 2016, online September 18). He declared machine learning to be the ultimate threat that no radiologist would be able to cope with in five to ten years.

Machines learn ... and could replace radiologists

Together with Ziad Obermeyer from Harvard Medical School, Emanuel is now following it up in the "New England Journal" (N Engl J Med 2016; 375: 1216-19). Again, it's primarily about machine learning.

Algorithm-based expert systems, such as those used to find drug interactions, would apply rules only to data. Learning machines, on the other hand, derived rules from data.

"Machine learning will replace much of the work that radiologists and anatomical pathologists do," they write. After all, digitized images can easily be thrown into algorithmic consumption by computers.

Are Computers Better Doctors?

The large amount of image data associated with advances in the image recognition capabilities of computers will mean that the accuracy of computers will soon surpass that of humans.

There will be a further boost from efforts to ensure patient safety - "after all, algorithms don't need sleep, and they are as awake at two in the morning as they are at nine in the morning".

Emanuel and Obermeyer have no doubts about how quickly this will all happen: "The time frame is years, not decades."

No more radiologists in 20 years

As Obermeyer confided in an interview with the Internet news service STAT of the Boston Globe Media, radiologists will mutate into cyborgs afterwards, supervise algorithms that read thousands of images per minute and only intervene when in doubt. "In 20 years there will be no more radiologists in their present form anywhere."

Of course, there are also voices that recommend a little more humility. One of them belongs to Stephen Holloway. He is the director of Signify Research, a company in Cranfield, UK, which specializes in information technology applications in the healthcare sector.

In the radiological news portal AuntMinnie, he pointed out that IT in medicine has already made considerable headway in analyzing large amounts of data - machine learning, which Obermeyer and Emanuel are focusing on, but just not.

Contradiction from IT experts

Holloway admits that there is certainly potential here. But: "It makes a big difference whether artificial intelligence should detect abnormalities or whether it should make diagnoses." Systems for the former are likely to be available in five years. A diagnostic use on a larger scale is not to be expected anytime soon.

Emanuel and Obermeyer's approach, however, reaches far beyond radiology. They see machine learning as an instrument with which the accuracy of medical diagnostics can generally be improved. At least they admit that this will happen more slowly and only in the course of the next decade.

The diagnostic standards for many diseases are unclear, and many things cannot be reduced to simple binary decisions as in radiology - benign or malignant. That complicates the algorithm training.

They cite the unstructured data in electronic medical records and the need to construct validation models as further reasons for the delay.

A recent study in which artificial and medical intelligence measured their diagnostic powers (JAMA Intern Med 2016, online October 10, reported by the "Ärzte Zeitung") has recently shown how far the path is there. So-called symptom checkers were clearly inferior to the doctors.

Algorithms also have limits

The optimism with which Obermeyer and Emanuel are approaching the use of big data algorithms in medicine still seems limitless. However, even the most powerful algorithms have limits: They are and will remain tied to the realm of the calculable.

This area is not all-encompassing, as Kurt Gödel and Alan Turing already demonstrated in the 1930s. Whether algorithms generally arrive at a result in the end cannot be decided algorithmically.

Consequently, not everything can be calculated, and there is not a calculable solution for every problem. So the medical intelligence may be limited - the artificial intelligence has been proven to be too.