So, IBM’s Watson won Jeopardy hands down, playing against Jennings and Rutter.
I am sure many NLP researchers will say there’s nothing new in what IBM’s Watson project achieved – all this has been done before. It has also not received the kind of attention that it deserves in the tech and mainstream media. The hoopla surrounding Kasparov’s loss to the chess playing machine Deep Blue was much bigger.
Incremental advancement in the eyes of many, to me, it is a momentous event. Watson is as much a research feat as an engineering marvel. It brought together many of the advancements in software and hardware to to notch this singular, elegant success. I get a sense that IBM’s Watson research team understands the import of this achievement but is downplaying its implications a little bit. By all indications they did treat it as a their own little Manhattan project.
Let me explain why Watson is a daunting technology for me – an informatics researcher’s view, if you will:
Search will never be the same again
Google’s search delivers dumb web pages, but Watson, without being connected to the Web, delivers answers. This is what the all so many search engines that keep cropping up have been promising but failing to deliver.
Watson learns the real connections between facts and that too from all the undisciplined way we humans have been documenting them. If in a matter of two years it could beat the Jeopardy champs, imagine what will it be able to do down the line. It can certainly learn from its own mistakes (and successes) as from your and mine. Sure Google also learns but it learns paltry little from the same resources.
We do not need Structured Data anymore
One major challenge, in particular for healthcare, has been having the data in a form so that we can use the existing powerful database operations like finding the right information, cross referencing different items, from it. It has always needed humans to be disciplined enough to enter the structured data to work around this failing of computers – the inability to work with unstructured data. Humans have to adapt to computers if they are to be fully exploited. This is the reason, why most EMRs are such inadequate tools for the clinicians’ primary tasks; EMRs are willing to accept unstructured data but have little capability do much with it. With Watson like technology under the hood, now you, as a clinician, can keep merrily describing the patient as you would to a resident or a colleague, the data will be ‘understood’ and stored the proper way in its memory for asking all sorts of interesting questions about the patient at a later point.
In fact, if the technology is adapted to include action recognition from other cues (videos, bar codes, sensors etc.) even documentation by narration will become increasingly redundant.
We do not need to author Rules
Many business and clinical solutions get their smarts from rules engines but the rules which provide the actual logic in them, are authored by some human expert. Watson like technology will make that redundant. If you tell it that the patient is a male, presenting with acute abdomen, has right iliac fossa tenderness and fever, it will, with moderate level of confidence, tell you that it is acute appendicitis, and that you better get blood counts and sonography, to clinch the diagnosis irrevocably. The thing is, no one would have fed the rules for differential diagnosis of acute abdomen anywhere in the system – it would have learned that from reading the surgery textbooks it is provided beforehand. For a while, I think there will be back on forth between the doctor and Watson, for one diagnosis that their combined understanding can settle upon, much like discussing a case with a colleague. But Watson would learn and remember much more from those interactions than the doctor would, progressively diminishing the need for the latter.
Information Retrieval Researchers can go home
The information retrieval researchers can also start packing up or quickly find some other problems to solve. Since nearly a decade medical informatics researchers have been trying to develop ways of providing adventitious information to the clinicians, that is highly pertinent to the patient on hand and his/her current problem – something like context sensitive help. There have been several ideas but all of them focused on tagging the data and resources themselves in a particular way to make this possible. These approaches will become redundant, since Watson like tech will not just identify the context much better but also dip into its learning to deliver the right resources. And all of this without any new XML tags having to be created
Knowledge Discovery will take a leap forward
Since it discovers patterns some parts of the technology suite of Watson can help pull out nuggets of unsuspected connections between facts. This will allow identifying things like new causalities for diseases and unsuspected benefits and side-effects of drugs, diets and interventions.
High level professionals should start feeling the heat
First they came for the typists, and I didn’t say anything because I was not a typist, then they came for the clerks, and I didn’t say anything because I was not a clerk ….
In fact, I even caught myself smiling smugly because I was a highly educated medical specialist. I knew the computers posed no risk to me. Suddenly, I am not so sure. Tally all that I have written above and it will be obvious to you that we are about to cede much of the intellectual grounds also to the computers. What will remain will be the contact based part of healthcare. Well, at least until the robots achieve a little more dexterity and are able to feign a better smile, when they say, “And how are you today, Mrs. Patterson” in a calm, reassuring and friendly voice (think HAL 9000).
I look forward to the day when we will be able to deploy Watson as an inference tool for Proteus.
Now, only if they can bring Watson’s size down to fit into my smartphone and for it to understand the Indian accent.