Natural language processing and clinical decision support for patient care
Using the help of Google and Wikipedia, “emergence” is defined as the way complex systems and patterns arise out of a multiplicity of relativity simple interactions. The word is a favorite in philosophy, systems theory, science and art, and is encroaching on the use of “paradigm shift” and “synergy” in business circles. It is not that the previous buzzwords are falling out of favor, it is that there is a logical progression of “looking at things from a different perspective,” into “noticing that specific parts of a system can act in concert,” and currently “observing how each part of a system can contribute to the emergence of a behavior that can only be performed by the group acting as a whole.”
In linguistics, this concept is closely related to “semantics.” The semantic meaning of a word or phrase is often only derived by observing the other words used along with it and the situational context of the entire message. For instance, the title of this column, “Emergency!”, could be a call for help within a sequence of events that have convolved to create a situation that is harmful to life, limb or property. It also could refer to a combination medical drama, action-adventure series that broadcast on American television in the early 1970s. Or, it could be an intentional pun of the word emergence with the intent to draw the reader’s attention to this important concept.
The complex task of correctly determining context, semantics and meaning is the basis of the game show, Jeopardy! Broadcast on American television in different incarnations since 1964, Jeopardy! challenges contestants to perform the deceptively simple task of recalling trivial facts. At face value, any player that is well-read over a broad array of topics should fare well. However, there are several cognitive gymnastics that separate each question from its answer. First, the answer is presented and the contestant must provide the correct question. Second, the “answers” are presented as nuances, puns, irony and humor in the context of a category that itself may have various semantic meanings.
While rarely performing on a television game show, we humans are constantly using situational awareness to determine the correct meaning of language. Undergraduates in the pre-health professions curriculum often lament that they feel as though they are in training to be a contestant on Jeopardy!, given all of the seemingly unrelated facts they must learn from courses in physics, chemistry and biology. Obviously, they are not studying to be search engines; we already have several of those. As future health professionals, their expertise and training are needed to bring context and semantics into clinical diagnostics and emergency situations.
Using Google to search for the phrase “broken leg” returns Web pages containing definitions, symptoms, treatments and videos relating to humans, animals and furniture, an indie rock song, and even a cocktail recipe containing rum and apple cider. International Business Machines Corporation (IBM) refers to this exercise in semantics as the “Paris Hilton Problem.” When entering the search phrase “Paris Hilton” into an automated search algorithm, a machine returns information about the Hilton hotels located in Paris, France, interspersed with gossip stories about Paris Whitney Hilton, the socialite daughter of Richard and Kathy Hilton. For a computer, as with a human, additional information is needed to disambiguate the phrase.
In computing, this task is known as “natural language processing” (NLP) and has its roots in the development of artificial intelligence in the 1950s. In the mid 2000s, IBM Researchers led by Dr. David Ferrucci set out to develop an NLP system that would be capable of handling the contextual challenges of Jeopardy! in a project call DeepQA — in homage to IBM’s previous chess-playing project, Deep Blue, but one that worked with questions and answers. The project developed code that is based on the open source software projects of Apache UIMA (Unstructured Information Management Architecture), Apache Lucene (a high-performance text search engine), Indri (University of Massachusetts and Carnegie Mellon University query language), and SPARQL (a resource description format query language based on World Wide Web Consortium (W3C) specifications). The DeepQA project integrated the open source software using proprietary methods and hardware to produce an NLP system named “Watson” in honor of IBM’s founder, Thomas J. Watson.
On February 16, 2011, Watson defeated two human all-time champions in the game Jeopardy! on national television. While each of the steps is complex, after receiving the text of a Jeopardy! “answer”:
1. Watson parses the natural language answer to generate a search query.
2. Watson’s embedded search engine searches a large document knowledge-base to find related documents (similar to performing a Google search).
3. Watson parses the natural language-based search results and generates potential answers (the hypotheses).
a. For each hypothesis, Watson constructs and initiates another search to collect evidence that supports this hypothesis.
b. Watson’s embedded search engine searches supporting evidence for each hypothesis in parallel.
c. The search results are again parsed and each piece of evidence is scored for its strength.
d. Each hypothesis is then assigned a score based on the strength of all of its supporting evidence.
4. The hypotheses are turned into a list of answers and returned to the user.
In essence, Watson performs a Web search and then performs additional Web searches on the results to determine which one of the results best fits the semantic context of the original query.
A 1999 study published in the British Medical Journal observed over 100 family doctors throughout their work day. The physicians asked more than 1,100 clinical questions during the 2.5-day study, and 64 percent of the questions were never answered. Of those that were answered, the physicians spent less than two minutes seeking answers, and only two were answered by conducting a literature search, while the others were answered by seeking advice from fellow physicians. A related 2006 study, published in the same journal, examined a year’s worth of diagnostic cases published in the New England Journal of Medicine. The study found that 58 percent of the cases were correctly diagnosed by a trained professional who could provide semantic context to Google search results of the symptoms.
Watson’s current goal is to serve as a semantic search tool to aid physicians as they provide patient care. Watson can continuously update its databases with the latest published medical research and also interface with patient electronic health records and laboratory clinical diagnostic analysis systems. Perhaps in the near future, hospitals will construct an additional “Emergency Room” that will house a local copy of the Watson supercomputer and where correct diagnoses and patient care decisions will emerge from the streams of collected medical informatics.
1. Watson and Healthcare: How natural language processing and semantic search could revolutionize clinical decision support http://www.ibm.com/developerworks/industry/library/ind-watson/index.html
2. IBM Watson: Ushering in a new era of computing http://www-03.ibm.com/innovation/us/watson/
3. Analysis of questions asked by family doctors regarding patient care http://www.ncbi.nlm.nih.gov/pmc/articles/PMC28191/
4. Googling for a diagnosis — use of Google as a diagnostic aid: Internet-based study http://www.bmj.com/
William Weaver is an associate professor in the Department of Integrated Science, Business and Technology at La Salle University. He may be contacted at editor@ScientificComputing.com.