A Social Network for Health Informatics Professionals and Students
11:00am Dr Martin Kohn MD is giving the talk on IBM Watson for healthcare. He's the Chief Medical Scientist for Care Delivery Systems at IBM research. The room is packed with attendees - there must be a lot of interest in this topic!
He starts his talk by introducing Watson as a hardware software combination: a supercomputer with lots of memory and processing power - the analytics to understand language is very computer intensive. Watson links to evidence based medicine sources, websites and even blogs to gather data.
Dr Kohn then showed a video of Watson in use on an iPad - it looks like a really clear, easy to use interface. Watson makes suggestions as information is entered into the software as to the diagnosis and management.
11:15am Dr Kohn says that 80% of medicine is using unstructured data - basic text-like data. Healthcare information in social media is this type of data - it has a great deal of influence, he says - e.g. patients finding out they are in the control arm of a study and dropping out.
Watson was designed for Jeopardy - This influenced it's key features: English only, 3 second response time, single question per system instance, unstructured text, can read 200 million pages of text in 3 seconds (stored in memory as couldn't connect to the internet on the show).
11.20am Healthcare is a more viable market than playing Jeopardy. The more Watson played Jeopardy, the better it gets, the more it "plays" healthcare, the better it gets. Most computers are very good at complex mathematics and database lookup calculations and therefore replace humans doing this. Unstructured text, however, is something most computers are unable to do - first Watson has to understand a complex, natural language question.
Search engines look for keywords in pages which can result in errors. Watson uses parallel probabilistic and temporal algorithms to analyse pages rather than just keyword searches. It returns suggestions with a "confidence level".
Watson uses "answer sources" to come up with possibilities and then links them to "evidence sources" to find the best answer. Temporal reasoning is very important in healthcare to determine progression of disease.
11.30am Watson can collect sources from a wide range of unstructured sources: Journals, medical records, the web, etc. It uses "iterative dialogue" to refine search results - it keeps going back to ask for more information and then starts the search over again. Watson doesn't remember anything about each iteration which is important for sensitive health information.
Dr Kohn refers to a Health Affairs Journal article that say that 16% of practice revenue is spent interacting with payers - this is very expensive and time consuming. Bringing Watson into this process can make it more efficient - it saves money on both sides and improves the patient experience.
Watson is also used in medical education for teaching medical students and residents - it teaches them how to think and analyse data, Dr Kohn says.
11.45am Dr Kohn says that Watson brings out information from the electronic health records. It brings a series of suggestion to the physician and links through to the sources of the suggestion in the literature. Dr Kohn then demonstrated a complicated case published in the NEJM and how Watson goes through the case to make suggestions.
Dr Kohn finishes his talk by reiterating that Watson isn't about making decisions for doctors but helps them work by presenting them with suggestions and information.
A question is asked about why Watson doesn't remember - Dr Kohn says that the job of the EHR - to store the health information. Watson is a tool that sits on top of the EHR.
In answer to another question, Dr Kohn says we don't have a lot of research on patients with 3 or 4 chronic diseases but there is a lot of data out there that hasn't been included in the normal studies as they are excluded from the analysis.
Would Watson be able to manage a hospital? Dr Kohn answers that this kind of data would probably be structured but Watson works best with unstructured data so wouldn't be used for this purpose.
A questioner postulates that Watson may only be as good as the evidence base and much of that is being shown to be flawed and biased. Dr Kohn says that this is a flaw in decision making in general and not specific to Watson - they are looking at trying to validate the literature. Watson also helps by using "confidence levels" to qualify the suggestions it gives.