Next up for Watson, I.B.M.’s clever question-answering computer? A stint as a medical student at the Cleveland Clinic Lerner College of Medicine of Case Western Reserve University.
The collaboration, announced on Tuesday, includes a bit of controlled crowdsourcing, with the Cleveland clinicians and medical school students answering Watson’s questions and correcting its mistakes.
“Hopefully, we can contribute to the training of this technology,” said Dr. James K. Stoller, chairman of the Education Institute at Cleveland Clinic. The goal, he added, was for Watson to become a “very smart assistant.”
Part of Watson’s training will be to feed it test questions from the United States Medical Licensing Exam, which every human student must pass to become a practicing physician. The benefit for Watson should be to have a difficult but measurable set of questions on which to measure the progress of its machine-learning technology.
Yet Watson, it seems, is not an answer to the nation’s looming shortage of physicians. “We’re not going to roll Watson in and certify it as a doctor,” said Dr. David Ferrucci, the I.B.M. scientist who is the principal investigator for the Watson project.
Once trained, Watson ought to be able to help physicians cope better with the rapid pace of incoming new research. Dr. Stoller estimates that the “half-life of existing knowledge” in medicine is probably down to four to eight years on most topics. After that, it’s obsolete, or partly so.
“What we learned in medical school loses accuracy over time, and is doing so at an accelerating rate,” Dr. Stoller observed.
For Watson, medical knowledge offers a complex knowledge terrain with deep and rich information, and a multitude of interrelationships between possible causes of a patient’s symptom. Highly probable answers exist, even correct ones, but there are many possibilities and paths to a decision. The human body is not a standardized, multiple-choice test.
The goal for Watson’s artificial intelligence technology, said Dr. Ferrucci, is to move beyond simply presenting Watson with questions, and getting answers back — which it did so impressively in defeating human Jeopardy champions last year.
In medicine, he said, you have a problem with many variables. For example, a 69-year-old female with certain symptoms, vital signs, family history, medications taken, genetic makeup, diet and exercise regimens.
Someday, Dr. Ferrucci said, Watson should be able to collect and assess all that patient data, and then construct “inference paths” toward a probable diagnosis – digesting information, missing nothing and winnowing choices for a human doctor.
“That’s where we want to go with Watson,” Dr. Ferrucci said.