A recent presentation by Martin Kohn, IBM's chief medical scientist, and Pat Skarulis, CIO at New York's Memorial Sloan-Kettering Cancer Center, suggests Watson's up to the challenge.
During last week's Digital Health Conference, sponsored by the New York eHealth Collaborative, Kohn and Skarulis outlined an impressive initiative their two organizations have embarked upon to use Watson "in the trenches" to treat oncology patients at Sloan-Kettering. Kohn outlined some of the basics of the project.
He was quick to point out that the supercomputer isn't just a "search engine on steroids," or even a massive database. It relies on parallel probabilistic algorithms to analyze millions of pages of unstructured text in patient records and the medical literature to locate the most relevant answers to diagnostic and treatment-related questions.
[ Is it time to re-engineer your Clinical Decision Support system? See 10 Innovative Clinical Decision Support Programs. ]
Kohn explained that 90% of the world's data has been created in the last two years, and 80% of that data is unstructured. As any clinician with a pile of unread medical journals knows, that massive collection of information includes far too many papers for any one human to read. Watson reads it for them. At the time of the Jeopardy competition, for instance, it was capable of reading 65 million pages of text per second.
With the help of natural language processing (NLP), the computer not only pulls out relevant terms to match the search terms in a clinician's query, but it also actually understands the idioms and other idiosyncratic gibberish we call English. Which means Watson can make sense of the fact that Americans park in driveways and drive on parkways, or the fact that noses run and feet smell.
Put another way, Watson does much more than just locate relevant keywords in its database. With the help of temporal, statistical paraphrasing and geospatial algorithms, it finds meaningful relationships between the clinician's question and its massive collection of medical facts and theories.
Armed with this skill, the supercomputer works through several logical steps to help physicians through their decision-making process. Once it understands the nature of the request, Watson generates a long list of hypotheses in response to the clinician's question. Then it assigns priority ratings to those hypothetical answers based on its analysis of millions of pages of stored data. Next, it generates a confidence level for each of the likely answers so that it can help physicians make an evidence-based decision.
In the final analysis, however, it's the clinician who must review the best solution and choose a course of action. He or she also has the option to ask Watson to supply all of the supporting literature upon which the computer based its answers. Similarly, Watson may ask for additional data, suggesting specific lab tests be done to improve the probability of arriving at a correct diagnosis or treatment regimen.
Of course all this impressive technology would only be an exercise in IBM bravado if there were no real patients and doctors to put Watson to the test. Enter Sloan-Kettering.
As Skarulis pointed out during her presentation, the medical center has about 2,000 order sets it can pull from when choosing a cancer treatment. Finding the best fit for each patient is no easy task. To help, Sloan-Kettering can tap its own massive database, called Darwin, which includes everything that has happened to all of its 1.2 million inpatients and outpatients over 20-plus years. In essence, that database embodies "the thinking patterns of all our experts," she explained.
The medical center decided to collaborate with IBM to "build an intelligence engine to provide specific diagnostic test and treatment recommendations," Skarulis says. The two organizations are now combining all of Darwin's intelligence with all of Watson's NLP capabilities. IBM is using all of the medical center's structured patient data and its NLP tools to convert the medical center's free text consult notes into usable data.
The team will first use this approach to tackle non-small-cell lung cancer. It has brought in Mark Kris, MD, one of MSKCC's top lung cancer experts, to help develop training cases for Watson to work on, focusing on 14 to 20 data elements, including the size and location of a patient's tumor, the presence of any genetic mutations (Sloan-Kettering does a full genomic analysis on all of its lung and colon cancer patients), and whether the tumor has spread to other tissues.
Watson's task is to follow the protocol that Kohn outlined above and come back with a list of diagnostic and treatment options for physicians to choose from, with confidence ratings for each option. Ideally, a treatment regimen that Watson concludes has a 95% confidence rating, for example, would help oncologists choose from the 28 different chemotherapy cocktails they have at their disposal.
Watson's training has prepared it for its role as a clinical decision support system. But now that it has graduated medical school, it's time for a real world residency. Skarulis hopes to launch a pilot program by the end of this year that will allow the supercomputer to work on real cases. It's hard to imagine an attending oncologist who would not want such a resident assisting him at the bedside.