sexta-feira, 20 de maio de 2016
Humans vs. Machines
IBM's Watson -- the language-fluent computer that beat the best human champions at a game of the US TV show Jeopardy! -- is being turned into a tool for medical diagnosis. Its ability to absorb and analyse vast quantities of data is, IBM claims, better than that of human doctors, and its deployment through the cloud could also reduce healthcare costs.
The first stages of a planned wider deployment, IBM's business agreement with the Memorial Sloan-Kettering Cancer Center in New York and American private healthcare company Wellpoint will see Watson available for rent to any hospital or clinic that wants to get its opinion on matters relating to oncology. Not only that, but it'll suggest the most affordable way of paying for it in America's excessively-complex healthcare market. The hope is it will improve diagnoses while reducing their costs at the same time.
Two years ago, IBM announced that Watson had "learned" the same amount of knowledge as the average second-year medical student. For the last year, IBM, Sloan-Kettering and Wellpoint have been working to teach Watson how to understand and accumulate complicated peer-reviewed medical knowledge relating to oncology. That's just lung, prostate and breast cancers to begin with, but with others to come in the next few years). Watson's ingestion of more than 600,000 pieces of medical evidence, more than two million pages from medical journals and the further ability to search through up to 1.5 million patient records for further information gives it a breadth of knowledge no human doctor can match.
According to Sloan-Kettering, only around 20 percent of the knowledge that human doctors use when diagnosing patients and deciding on treatments relies on trial-based evidence. It would take at least 160 hours of reading a week just to keep up with new medical knowledge as it's published, let alone consider its relevance or apply it practically. Watson's ability to absorb this information faster than any human should, in theory, fix a flaw in the current healthcare model. Wellpoint's Samuel Nessbaum has claimed that, in tests, Watson's successful diagnosis rate for lung cancer is 90 percent, compared to 50 percent for human doctors.
Sloan-Kettering's Dr Larry Norton said: "What Watson is going to enable us to do is take that wisdom and put it in a way that people who don't have that much experience in any individual disease can have a wise counsellor at their side at all times and use the intelligence and wisdom of the most experienced people to help guide decisions."
The attraction for Wellpoint in all this is that Watson should also reduce budgetary waste -- it claims that 30 percent of the $2.3 trillion (£1.46 trillion) spent on healthcare in the United States each year is wasted. Watson here becomes a tool for what's known as "utilisation management" -- management-speak for "working out how to do something the cheapest way possible".
Wellpoint's statement said: "Natural language processing leverages unstructured data, such as text-based treatment requests. Eighty percent of the world's total data is unstructured, and using traditional computing to handle it would consume a great deal of time and resources in the utilisation management process. The project also takes an early step into cognitive systems by enabling Watson to co-evolve with treatment guidelines, policies and medical best practices. The system has the ability to improve iteratively as payers and providers use it." In other words, Watson will get better the more it's used, both in working out how to cure people and how to cure them more cheaply.
When Watson was first devised, it (or is it "he"?) ran across several large machines at IBM's headquarters, but recently its physical size has been reduced hugely while its processing speed has been increase 240 percent. The idea now is that hospital, clinics and individual doctors can rent time with Watson over the cloud -- sending it information on a patient will, after seconds (or at most minutes), return a series of suggested treatment options. Crucially, a doctor can submit a query in standard English -- Watson can parse natural language, and doesn't rely on standardised inputs, giving it a more practical flexibility.
Watson's previous claim to fame came from it winning a special game of US gameshow Jeopardy! in 2011. For those unfamiliar, Jeopardy!'s format works like this: the answers are revealed on the gameboard and the contestants must phrase their responses as questions. Thus, for the clue "the ancient Lion of Nimrod went missing from this city's national museum in 2003" the correct reply is "what is Baghdad?". Clues are often based on puns or other word tricks, and while it's not quite on the level of a cryptic crossword, it's certainly the kind of linguistic challenge that would fox most language-literate computers.
Watson's ability to parse texts and grasp the underlying rules has had its drawbacks, though, as revealed last month when IBM research scientist Eric Brown admitted that he had tried giving Watson the Urban Dictionary as a dataset. While Watson was able to understand some of the, er, colourful slang that fills the site's pages, it also failed to understand the different between polite and offensive speech. Watson's memory of the Urban Dictionary had to (regrettably) be wiped.