everythingpossible - Fotolia
Starting around 2013, IBM began making some bold claims about Watson's role in cancer care. Executives said it would revolutionize diagnosis and treatment of one of the world's most stubborn problems. At one point, the company even said in a press release that its technology would "power 'Moon Shots' mission aimed at ending cancer."
Ending cancer is a big goal. Unfortunately, Watson has consistently fallen short. Over the summer, STAT News and The Wall Street Journal published exposes showing how doctors have little trust in Watson because it consistently delivers bad recommendations. Watson is giving AI a bad name in healthcare.
That's too bad because healthcare and AI could be a natural fit. Machine learning tools could have an important role to play in reshaping a chronically overpriced, underperforming industry. For the technology to have this impact, those selling it are going to need to exercise a bit more modesty.
The more methodical approach
Google has been busy in recent years developing powerful machine learning tools mostly aimed at consumer applications, like search and voice recognition. The tech giant has been quieter about its efforts in healthcare, which have been among its most impressive. Just recently, Google AI published a pair of journal articles describing progress in using a deep learning algorithm to identify signs of metastatic breast cancer in imaging results.
This is a long way from IBM's efforts at applying AI in healthcare. IBM started with a lofty promise to create a big marketing splash and then rushed its technology into the hands of providers, encouraging them to use it as broadly as possible, in everything from diagnosis to treatment. Google, on the other hand, leads with trials showing measurable impact in one specific area of cancer diagnosis.
Haven't heard about Google's results? It's probably because the company hasn't taken out any prime-time television advertising spots to crow about how its technology is revolutionizing medicine. Even the company's blog post announcing the success points out that the findings have substantial limitations and more research is needed before any clinical applications are conceivable.
Targeted trials trump broad claims in healthcare and AI
Healthcare is incredibly complex, and cancer treatment especially is tricky. It was never realistic for IBM to think it was going to set loose an AI tool to fix every problem in cancer care. The more targeted approach that Google is taking right now makes more sense.
Beyond cancer care, there are all kinds of inefficiencies in healthcare. There's huge variation in treatment across diseases -- sometimes because patients present unique symptoms, but sometimes because doctors have idiosyncratic ideas of how to deliver care. Pricing varies tremendously from hospital to hospital, patient to patient. Staffing levels don't always match patient volumes.
These problems are all potential targets for AI applications. But it's hard to see how a general-purpose AI tool could address all such problems. Each issue is driven by a unique set of processes -- which are reflected in unique data sets -- and a machine learning algorithm that develops proficiency in pricing care will have no ability to assist in disease diagnosis. Pitching an AI tool for healthcare broadly makes no sense. Even aiming a tool at a single disease that has as many variations as cancer, which is what IBM did, is unlikely to deliver satisfactory results.
For healthcare and AI to work well together, we're going to need to see a number of different tools aimed specifically at narrow areas of the system, not broad approaches. Any AI vendor that's so bold as to promise to cure cancer or any other complicated disease should be recognized for what it is -- a 21st century snake oil salesman.
But we all know there's money to be made from AI, and the healthcare sector certainly has money to spend. The question will be whether providers see value in AI tools. The recent shortcomings of Watson -- particularly in light of IBM's aggressive marketing of the technology -- may lead some providers to doubt the role of machine learning. It's up to those developing and selling these tools to prove their worth and exercise a little modesty when pitching them.