Evaluating the impending impacts of the machine-learning economy
Karl Benz is credited with inventing the automobile in Germany, but Henry Ford introduced the concept of mass-produced vehicles that were economical for the everyday consumer.
Many consider machine learning as revolutionary to medicine as the automobile was to transportation, but what will it take to enter the mainstream? What are the economic shifts we can expect? And how can radiologists prepare for a machine-learning future?
For answers, let’s consider some basic lessons from Economics 101.
Lesson #1. When the cost of something falls, more people demand it. Machine-learning algorithms take available information and use it to fill in — or predict — missing information or something that is unknown (such as the presence of a disease). In this way, machine learning is a prediction technology, explains Avi Goldfarb, PhD, the Ellison Professor of Marketing at the Rotman School of Management at the University of Toronto. “As artificial intelligence improves, it will lower the cost of machine prediction,” he says. “And as the cost of prediction drops, we’ll begin using machine learning for many more tasks — including those that were never before framed as prediction problems, like medical diagnosis.”
Lesson #2. When the cost of something falls, the value of substitutes drops along with it. The substitute for machine prediction is human prediction. As the cost of machine learning falls and it starts getting used more often, the value of human prediction skills will also fall. So, if diagnostic radiology were considered only as a prediction problem, the value of radiologists would fall.
Lesson #3. When the cost of something falls, the value of complements rises. “When coffee becomes cheaper, people buy more sugar and milk, and their value rises,” says Goldfarb. “In the machine-learning economy, we believe human judgment is going to become increasingly valuable as a complement to diagnostic prediction.”
Implications for Radiology
These days, everybody is talking about IBM’s Watson, but machine learning isn’t new. In fact, Arthur Samuel, a pioneer in artificial intelligence at IBM, built a checkers-playing program as far back as 1959. What is new is the ability to apply machine learning to radiology, says Adam C. Powell, PhD, president of Payer+Provider Syndicate, a management advisory and operational consulting firm focused on the managed care and health care delivery industries.
“For machine learning to work, you need digital inputs and outputs,” says Powell. “In radiology, we are now capturing images and information digitally and thus have a digital input. Thanks to electronic medical records, we now have a digital output from the diagnostic process. We can develop algorithms to map inputs to outputs, using archives of digital images and recorded diagnoses as training data.”
So can a computer learn to do a better job of diagnosing medical problems than the radiologists who programmed it? And will the laws of economics put radiologists out of their jobs?
Not according to Howard B. Fleishon, MD, MMM, FACR, secretarytreasurer of the ACR’s Board of Chancellors. “Right now, there’s paranoia about radiologists being displaced by machine learning,” he says. “But radiology is much more than just routine image interpretation. There are complexities that require human judgment. There’s also the important role of patient and physician interaction. Clearly, our profession will change as machine learning becomes better and cheaper. But radiologists will also become more effective as a result, and the value of our judgment will continue to rise.”
New Roles for Radiology
Powell predicts that the advent of machine learning in medicine will create new roles for radiologists. “Bread-and-butter services may become more automated, but there may need to be more radiologists who have specialized knowledge in particular niches. There may also need to be ‘algorithm curators’ that help decide which algorithms best achieve specific objectives. And we may need more medical ethicists to help make judgments,” he says.
The need for patient interaction and shared decision-making may also expand. “The patient will still need someone to explain the implications of the findings,” says Powell. “How can they visualize the issue? How can they make intelligent decisions about the care pathway they wish to follow based upon the probabilities that are presented by the data? Machine learning may drive the need for more clinical radiologists who take a holistic approach.”
What will it take for machine learning to become mainstream in medicine? The answer, says Goldfarb, is surprisingly simple: “It just has to be better than a person. Are we there yet in radiology? Not quite. But as more data accumulates, as the research progresses, there will be an increasing number of situations in which machines are able to automatically make diagnostic predictions with a greater degree of accuracy than a person would.”
Powell agrees. “It’s not going to be all or nothing, and there are many different types of problems to solve in radiology,” he says. “It may be easier for a machine to perform better than a person at evaluating some very constrained, frequently seen clinical situations. It will probably be much harder for a machine to be better at irregular, complex, or rare problems.”
What do radiologists need to know now to prepare for the machine-learning future? The most important takeaway, says Goldfarb, is that these are prediction technologies, and so the parts of the job that involve prediction will increasingly be done by machine. “But don’t panic,” he advises. “While machines may consume some aspects of your job, we’ll start to see the rise of a whole lot of complementary things that still require human judgment. As a profession, radiologists should identify and invest in the skills that are complementary to better interpretation of images.”
Fleishon believes that machine learning will pave the way for incredible opportunities for radiologists. “Not only will it supplement our interpretive skills, but applying algorithms to personalized medicine and using massive data sets to drive new research initiatives could dramatically improve the future of patient care and our profession.”
By Linda G. Sowers, freelance writer for the ACR Bulletin