IBM’s Watson starts its medical career

IBMs Watson starts its medical career

IBM’s Dr. Watson of Jeopardy! fame has finally completed its residency and fellowships and, presumably to its creators’ utter delight, is now a practicing oncologist.

The prodigy “cognitive system” completed its training in less than a year at the illustrious Memorial Sloan-Kettering Cancer Center, and although only proficient in lung cancer right now, Dr. Watson’s career as an advisor to oncologists everywhere is off to a great start.

A recently released video demonstration shows Dr. Watson in action, researching, evaluating and treating a 37 year old woman with newly diagnosed stage IV lung cancer in his advisory capacity to a hurried and pretty uninspiring human oncologist. Regardless of the slightly weird scenario, it is worth noting that in a fraction of a second Dr. Watson, scours 3,469 text books, 69 guidelines, 247,460 journal articles 106,054 other clinical documents and 61,540 clinical trials, and evaluates their contents against the patient’s EMR to identify need for further diagnostic tests and treatment options for this patient. Being an exceedingly helpful advisor, Dr. Watson quickly reads the entire EMR and uses his trained processing power to eliminate all the clutter in the EMR, presenting to the human doctor only information pertinent to this particular diagnosis. Ouch.

On the other side of town, ONC is busy apologizing for the sorry state of what it calls “interoperability”, blaming everything from the lack of standards to people’s inability to agree on a restricted set of vocabularies for the medical profession. According to the ONC philosophy of interoperability, only “computable” data can be exchanged or analyzed in a meaningful way. In other words, all medical professionals must learn to express themselves in a standardized way which computers can understand.  To that end we have ICD-9, ICD-10, SNOMED-CT, LOINC, RxNorm and all sorts of other terminologies and vocabularies aimed at restricting the English language to the limited computational abilities of available EMR software. How do you say “Mr. Smith is a pleasant 82 year old gentleman with a sad demeanor” in SNOMED? You don’t. You dispense with the pleasantries (pun intended) and diagnose Mr. Smith with depression. The Sapir-Whorf linguistic relativity hypothesis is by no means a settled subject, but if it contains any truth and vocabulary does affect cognition, then how will restricting clinical vocabulary affect the cognitive abilities of its users over time? We don’t know, and frankly I am not interested in finding out.

The folks at IBM took a different route. Paraphrasing Sir Francis Bacon, loosely quoted as, “If the mountain won’t come to Muhammad then Muhammad must go to the mountain”, Dr. Watson’s creators must have decided at some point that if the doctor won’t come to the computer then the computer must go to the doctor. Instead of framing the problem by asking how we can change human communications to better enable the current generation of computers to “understand” humans, IBM began by figuring out how to change computers to better enable them to understand current forms of human communications.

Thus, Dr. Watson learned to read books and articles and all sorts of “unstructured” information, because no matter how hard the powers to be are trying to fit the square peg of human language into the round hole of computer language, and tragically vice versa, most information generated by people is in their natural language and Dr. Watson was programmed to process natural language. So if Dr. Watson is able to “read” half a million pieces of text of various heft in a second or so, how long would it take for it to read an old fashioned paper chart, or an electronic rendition thereof? I am pretty sure that if you ask nicely, Dr. Watson would be happy to rearrange it for you in any way you choose, while pointing out the most pertinent parts to your current objective, highlighting discrepancies, missing and redundant information, all in a picosecond or less. And interestingly enough IBM developers thought it wise to take a generalized path to Watson’s education, instead of creating specialized Watsons each with linguistic abilities specific to a domain. Seems more human friendly that way…

The IBM Watson software line is not an EMR, but it can process and analyze information in an EMR. It is really an attempt at artificial intelligence consisting of a gigantic contextual search engine, coupled with lots of very sophisticated and self-generating algorithms to both analyze and inform the search itself. Watson doesn’t need to have the smoking status check box clicked in order to infer that the patient is or is not a smoker and doesn’t need to have a new standard defined before it can read a patient’s family history.

True, Watson is pretty new software and folks have been tinkering with natural language processing and artificial intelligence for half a century without much success, but things are beginning to coalesce now and technology in the next decade will look very different. Is it really wise for our government to spend so much money and invest so much effort in building and enforcing the use of tools that are becoming obsolete faster than they are created? My hat is off to the VA and DoD who gave up on the strange and expensive idea of building their own EMR from scratch (better late than never). I think it’s high time that other governmental agencies got out of the EMR design business as well, because there are companies out there whose core competency is technology and who have large enough innovation budgets to build the next generation of health IT.

Consider this: What if Dr. Watson had a few less educated siblings serving as medical secretaries, summarizing, abstracting and relaying information back and forth, on demand? All of a sudden the shape, form, functionality, standardization and all “meaningful” bells and whistles in an EMR are rendered irrelevant, and using Microsoft Word for typing or dictating your note is as good as using a “certified” EMR, or much better, because the context is so much clearer and so much more forthcoming.

Whether it can pass a Turing test or not yet, Dr. Watson is not a real doctor, and it will not be one in our lifetime. Dr. Watson has no free will and everything it knows is dictated by the corpus of knowledge made available to it by its creators or employers. There will be huge ethical and legal questions raised by software capable of supplanting human decision making processes, and software that can be centrally deployed and manipulated to this end. Even before that future arrives, it is worth noting that Dr. Watson is simultaneously employed in oncology clinics and by payers, and in my opinion, Dr. Watson has one button too many – the direct button to the insurance company, which will automatically approve payment for Dr. Watson’s top recommendations, but presumably not so much for other choices. Like all technologies, Watson embodies hope for the greater good along with great new perils for ordinary people. Leaving these philosophical questions aside for a moment, the only certain thing is that Dr. Watson is starting its brilliant new career by introducing a cure for one very painful disease that is reaching pandemic proportions amongst medical professionals: clicking boxes.

Margalit Gur-Arie is founder of BizMed. She blogs at On Healthcare Technology

Comments are moderated before they are published. Please read the comment policy.

  • mjdemay

    I think this is great news! I envision a day where patients walk into a clinic and are handed a tablet to fill out their symptoms and verify their history. The form is then sent to a “super computer” (like Watson) and analyzed. The staff and physician then receive a detailed report including specific questions to ask relating to the symptoms, a probability chart of a diagnosis and suggested educational resources (in case the staff/physician need educational refreshing and updating with new relavent scholarly articles). This will help direct the time spent by the staff/physician and limited resources to be much more efficiently used.

    • Close Call

      And when the probability of a certain treatment working for your possible diagnosis falls below a certain threshold, your insurance company has a great reason not to pay for said treatment.

      • Margalit Gur-Arie

        That is very true and we should keep an eye on those extra buttons, but in all fairness, you can probably program any EMR to do the same thing, and there are initiatives under way to provide electronic adjudication of claims already.

        • N N

          Um, keep an eye on those extra buttons? Sorry, but they come part of the package. You can expect Watson to memorize and put together some pieces of information and not other pieces of information, as to whether certain treatments fall below a certain threshold, which the insurance company can then use.

          • Margalit Gur-Arie

            Not necessarily. The current implementation of the IBM technology seems to include the ability to preadjudicate prescribed therapies. This does not mean that future implementations of the software should necessarily include the payer point of view.
            If I was customizing and/or developing this software, I would suggest that the “best” clinical recommendations are objectively listed and that some sort of visual cue for what the patient’s plan will pay for is also displayed. Very similar to what you currently see in most eRx formulary advice.
            The cold facts are that insurers pay for some things, but not others, and sometimes they don’t pay for what you or your patient think is the best course of action (particularly the low priced or meagerly funded plans). This is a tragedy that technology cannot change, not even IBM. We just need to make sure that technology, at the very least, does not make things worse.

          • Matt DeMay

            HIPPA definitely has its advantages and disadvantages in protecting patient’s data. We really need the HMOs to work together and allow patient’s data to be transfered through the electronic medical record programs we have.
            It is a shame when unconscious patients with an MI come into a hospital and we can’t obtain their history before intervention!

  • Suzi Q 38

    Thanks for this information. It can save time and lives.

  • Adriana

    Watson is a fantastic idea and I’m all for it. I think it is a bit naive on the part of the original author to suggest that software/electronics shouldn’t supplant human decisions or that it raises ethical issues. In fact, if the author has an automobile, has traveled via airplane, etc then technology already supplants human judgement and with overwhelmingly positive results. Your “mechanic” doesn’t do 25 different tests to determine what’s wrong with your vehicle, no, instead they simply plug in the EMI to their computer and it tells them exactly what’s wrong. That 14 hour cross-atlantic flight is run by computers for oh, about 13 hours and 55 minutes. Ditto many public transit systems, driverless cars, etc. Whether you do so consciously or with benign ignorance, you already consent to having life/death decisions made by something other than a human.

    • Close Call

      The chances of dying in an airplane crash are incredibly small – it’s because there’s good engineering, tough regulations and smart people at work to make air travel safe.

      You can’t can’t reach that level of positive outcomes in medicine. Cancer is cancer. We still age. We still eventually die. The goal of an airplane is to transport you from point A to point B without killing or maiming you. The goal of medicine is to prevent you from dying, to heal you, and take away your pain and comfort you.

      Medicine isn’t perfect and cannot ever hope to achieve the same sort of success as we have with automobile or airline safety.

      Watson may provide better predictions in some instances than a doctor, but what do we do with these better predictions? Does insurance get to stop paying when a potential treatment falls below a certain threshold?

      When Watson says that my grandmother’s chance of dying is 95% in the next 6 months because of kidney disease, does this give Medicare the ability to say “Hey, the chances of your grandma dying are really high and it doesn’t make sense to pay for her dialysis if she’s going to die anyway. If you want that extra treatment, you need to pay on your own.” We have precedent for this in terms of prior authorizations. It’s going to happen.

      The problem with Watson is that it makes explicit the imperfect predictions doctors make every day. It’s an impartial, objective arbiter of data and predictions unmotivated by drug reps, where you trained or if you just saw a commercial for a new drug on tv. But it will never reach 100% certainty, and many times it will fall far below that mark.

      My personal take: Will it be better for our budget and the long term fiscal health if we use something like Watson in clinical decision making? Yes. Is this called rationing? Yes. Will it be incredibly difficult to know how to pay for treatments that have a low predictive success rate? Yes. Are we going to have some tough conversations with people who want to do “everything” even though the chances of success aren’t on their side? Absolutely.

    • Margalit Gur-Arie

      What Close Call said, plus…
      It’s not really a question of letting machines “make decisions” that could result in death in case of malfunction. There are hundreds, perhaps thousands of examples for this. It is about letting qualitative statistical algorithms, based on imperfect science and “other” considerations, preempt, rather than inform, people’s decisions.

  • Margalit Gur-Arie

    Dr. Sherling, I don’t disagree with the possibilities created by collecting data from EHRs, cloud based or otherwise. I do however believe that the natural language processing abilities available in the Watson technology, and in other products now coming to market, will obviate the need for the user (physician) to input data into the computer in a predefined format that is convenient to older products. Watson type software can and does process plain text to extract discrete values for analytics.

    So why not just let the data collection part of the EMR be just a word processor, or voice processor, or picture/video processor, or whatever the user prefers it to be, and let the Watson type software do the heavy lifting of abstraction?

  • Shirie Leng

    Right. IBM has Watson and my EMR can’t talk to the hospital across the street. What is going on here??? Read the article about medical technology in the Atlantic Monthly from March by Jonathan Cohn. That computer utopia looks pretty cold and lonely.

Most Popular