Like the famous HAL computer in the movie 2001: A Space Odyssey, Watson has had to learn by starting with basic stuff. When its objective was to beat human Jeopardy champions, Watson marshaled 12 terabytes of stored data, including questions and answers from previous Jeopardy games, to help decide which information was most relevant in forming an answer to each new Jeopardy challenge.
In healthcare, sometimes there is no one right answer, so Watson is being trained to return a list of possible diagnoses based on input ranging from the patient's own story, ordered tests, experiences of similar patients, and peer-reviewed medical literature, which is weighted most heavily.
To date, Watson has ingested more than 600,000+ pieces of medical evidence, two million pages of text from 42 medical journals, and clinical trial data in the area of oncology research.
Like a good journalist, Watson also learns which sources are reliable, favoring those sources' results over time, and discounting those results that come from less reliable sources. The term "artificial intelligence" is a bit of a misnomer, because Watson doesn't generate new ideas itself, but instead relies upon accessing existing information.
What makes the Memorial Sloan-Kettering test fascinating is that the institution already has 15 years of electronic health records on five million encounters, Kohn says. Over time, this kind of data will become some of the most valuable assets of any healthcare system, and for the first time, it will be powering an AI system to help make decisions.