Eminent academicians are editorializing in the New England Journal of Medicine (1) (and elsewhere) that artificial intelligence computer systems are about to disrupt both the practice and the profession of medicine, to the point where some specialties, such as radiology, may vanish entirely.
Such prognostications are symptoms of an unfortunate condition I call “techno-hypnosis:” entrancement with an emerging technology while ignoring lessons of history from the wider world.
Let’s look at the radiology claim. It stems from recent breakthroughs in machine learning techniques – a type of artificial intelligence – that now allow computer scientists to develop, with relative ease, software that recognizes disease in medical images, for example, signs of a fracture in a leg bone’s x-ray, or signs of a cancer in a mammogram.
Moreover, formal studies have demonstrated that such software can detect certain features with the same accuracy as humans, and sometimes better.
But this does not mean the sky is falling for radiologists, for several reasons.
First, we have been here before. Computers started reading the squiggly lines on electrocardiograms in the 1970s, and quickly became better at it than 99% of physicians. Yet, cardiologists today still read EKGs – in vast numbers – and they read them faster because the computer does the legwork of making basic measurements and calling out specific features onto which the cardiologist can apply his or her deeper skills.
Second, success in a few trials does not guarantee success in the wider world, a problem widely known in the artificial intelligence field for decades. All software systems have limitations, and – especially with complex systems – these limitations are often not apparent until they leave the test bed and encounter the unrestrained chaos of the real world. Maybe the new AI techniques have licked this problem, but I doubt it. Witness last month’s report that gene sequencing laboratories have for 10 years been incorrectly classifying certain gene variants as harmful (2). The labs were violating statistical assumptions that weren’t even recognized as being assumptions, but are now obvious in retrospect. AI systems in the real world would be subject to exactly the same type of error.
Third, the new AI technologies have a huge problem: they cannot explain themselves to a human, and a sub-genius human cannot understand how they work. Humans discovered the peril of this situation 2500 years ago, when the ancient Greeks wrote stories of wars started by cryptic pronouncements from the oracle at Delphi, pronouncements the oracle refused to explain. Fortunately, DARPA (the Defense Advanced Research Projects Agency) knows its classical literature and is launching an initiative to add explanatory features to the new AI. It won’t be easy.
Finally, even if AI systems are spectacularly successful, human radiologists are not going to go away. For the sake of argument, let’s define “spectacular” as meaning that the cost of an MRI scan, including interpretation, drops to $1 because AI software has superhuman accuracy. If that happened, then every physician visit by every patient would include an MRI scan. (Why guess if your headache is a sinus headache, when it only costs $1 to see?) The resulting explosion of MRI scans would generate large workloads for human radiologists, even if they had to review only 1 in 100 of the AI system’s interpretations.
This principle, well-known in economics, is called the Jevons paradox, after William Stanley Jevons. In 1865 Jevons noted that the greater fuel efficiency of the new Watt steam engine, compared to the then-standard Newcomen engine, would not lower coal consumption, but would increase it because the lower price per unit of energy would allow more people and companies to consume coal (3).
Clearly, major changes are coming to cognitive medical specialties, but, given healthcare’s tangled economics, it is premature to affirm that computerized image interpretation will decimate radiology workloads.
(1) Obermeyer Z, Emanuel EJ. Predicting the future -- big data, machine learning, and clinical medicine. N Engl J Med. 2016; 375: 1216-1219.
(2) Manrai AK, et al. Genetic misdiagnoses and the potential for health disparities. N Engl J Med. 2016; 375: 655-665. PubMed 27532831
(3) Jevons WS. The Coal Question. London: Macmillan and Co., 1865. Pages 102-104. https://archive.org/details/coalquestionani00jevogoog/page/n8 https://en.wikipedia.org/wiki/The_Coal_Question