The National Library of Medicine Explores A.I.

The National Library of Medicine is working on a search engine that will scan thousands of medical records to turn up documents related to patient queries. That sounds exactly like what Google and Bing can already do, but it’s different.

It will employ Watson-like search and retrieval techniques to return a very specific set of one to three documents. Watson, IMB’s robot with artificial intelligence capabilities, smashed Jeopardy! records when it appeared on the gameshow last year. One trait that makes Watson intelligent is that the computer can understand natural language as opposed to a Google query (i.e. “cure common cold” instead of “what’s the cure for the common cold?”), something that researchers at the National Library of Medicine (NLM) want to try to replicate with their own machine.

Another characteristic of intelligent computers is that they learn. This occurs over time and through repetition.

“In order to train the machine we had to have a large set of questions in the vernacular, not cleaned up to grammar, and we had to have a lot of associated, or reasonably associated, answers,” Dr. Milton Corn, deputy director for research and education at NLM, said.

Having a database of answers ― not just questions ― is key because when the computer turns up a wrong answer, researchers need to show the computer what the right answer looks like. Corn said NLM gets millions of queries from patients every year, but they’re answered through results generated by NLM’s databases PubMed and MedlinePlus, as opposed to health care professionals.

In his search for a set of published medical questions and answers, Corn came across Ask the Doctor, a Canadian website where doctors answer individual patient questions. Ask the Doctor agreed to share a set of about 8,000 medical questions and their answers with NLM for its artificial intelligence project.

Right now the computer is in Rocky Balboa mode.

Researchers are training it by feeding it questions, then feeding it answers and testing it. Then they repeat the process. The major hurdle for the computer is getting to a point where it can understand natural language.

And if the machine ever becomes ready for public use, patients, too, will have to be trained to approach the Q&A process differently. Patients will need to phrase their questions like they’re talking to their doctor, not Google. This will lay the groundwork for the possible incorporation of voice technology later on.

Is NLM also laying the groundwork for computer replacement of physicians? Not quite.

“We are not contemplating ― unless this were an unbelievably fantastic success ― letting a machine practice medicine,” Corn said.

If this early stage does prove to be successful, the project might expand with a partnership with IBM, which has expressed an interest in working with NLM.

“We would like to see whether Watson could be adapted to health questions,” Corn said.

But first, Watson will be in for some medical education.

“It doesn’t know much about health at the moment,” Corn said.

  • http://www.facebook.com/RebeccaCoelius Rebecca Coelius

    Would be interested to hear how they are giving Watson “a medical education”. I read an interesting study recently suggesting that natural language analysis of EMR records were as accurate as the providers in final diagnosis and predicting the patients state of health. This is possible because medical education does a pretty a good job instilling a consistent language used across providers to descrive everything from the lung exam to a skin lesion. 

    • Laura Montini

      When I followed up with Milt Corn he said: 

      I don’t know in detail how the Watson team goes about its work, but in general “machine learning” involves setting up a situation in which the machine is programmed to solve a problem according to the programmers guesses about what computational approach might work.  For the problem involving answering of questions, the results are compared with the “true” answers, and then the machine’s programming is tweaked to improve the results. The process is repeated on the training questions until the machine is answering at a satisfactory level. Then comes the challenge:  the machine is now tested on a set of questions similar in format and theme to the training questions, but completely different in that the machine has never seen the test questions before.  Commonly, before the training begins, part of a question set is used for training, and another part is reserved for later testing.  The challenge step is essential, because not much has been accomplished in the way of “intelligence” or “learning” if the machine knows only how to recognize familiar questions but can’t do anything with questions it has never seen before. I add the obvious comment that use of words such as “learning” and “intelligence” are just metaphors; there is no implication that the machine is conscious or understands anything in the human sense.  We are programming  it by trial and error to behave in a manner that would be considered intelligent in a human.
      Natural language processing is fundamental to analyzing electronic patient records, but state of the art is not really able to do as well as the reader was told.  I wish all doctors used precisely the same language, but in fact, the variations in terminology and in abbreviations from doctor to doctor and from hospital to hospital are so large that imposing some translation/thesaurus system is getting a lot of attention now, and is essential if we want computers to “read” the records.
      I don’t understand what the machine actually accomplished in diagnosis if all it did was “read” what the doctors had diagnosed and written in the chart.