Recently a company called Wellpoint was mentioned by the WSJ as purchasing some Watson technology. According to the article Steven A. Mills, an IBM vice president, says "the system could be used in settings as varied as call centers and offices doing engineering and scientific work." According to the article the "first Watson deployment would come early next year with WellPoint nurses who manage complex patient cases and review treatment requests from medical providers."
Now on the one hand I really have to give it to IBM. Watson is an excellent system and works well for playing Jeopardy. IBM has a long history of technology advancement in this area as I wrote in "Elementary My Dear Watson" back at the beginning of the year.
Unfortunately, on the other hand big business is big business - particularly big medical business. And, according to the article Watson will be busy with more than just Jeopardy questions. "Replacing call centers" for one.
Today most people I know do not reaching a call center in India. They find it a problem at many levels: US jobs sent overseas, a poor level of knowledge, poor support outcomes, those sorts of things. But I guess soon we won't have to worry about reaching India. Instead we will reach Watson.
And for medical advice, no less.
Now personally I have viewed many "medical" websites - WebMD comes to mind as an example. These sites have various diagnostic tools to help you "decide what to do." Should I call a doctor, should I go to the emergency room, those sorts of things.
But these web pages are out of control - mostly causing concern and panic where simple common sense would suffice. Several examples I personally witnessed recently come to mind. One involved an insect sting of some sort - no body was found. Red dot in the center, some swelling, and some pain - but nothing that wasn't easily tolerated. The results of WebMD? Emergency room. A few hours later the incident was all but forgotten.
Another was a "rib pain" in an elderly person. Somebody went on "WebMD" and decided, again, it should be an emergency room visit. This time the victim went only to spend hours and hours, get all sorts of dangerous and expensive tests like CAT scan, and to finally be sent home with "we don't know what it is and its probably not dangerous."
Will Watson do any better after processing all of these WebMD-type pages?
The sort of application Watson represents to healthcare is a far more complex version of the old "Artificial Intelligence" paradigm created thirty or more years ago. The quintessential example of which was the Digitial Equipment Corporation (DEC) R-1 system for configuring VAX computers.
VAX computers were big, giant old-school boxes that sat in computer rooms with arrays of attached devices, disks, terminal controllers and the like. Humans often made mistakes when selling VAX computers because they would forget necessary items, e.g., specific cables, needed to make the entire system work.
So DEC started a project to have an "expert system" assist the humans in configuring VAX computers. The idea being that you'd have rules like:
(need-cables?
(sold: disk drive)
(sold: VAX disk controller)
(required: cable))
which basically said if you sold a disk controller and a disk don't forget the cable.
So collections of rules like the one above were loaded into a computer and the computer applied them.
However, according to "What every engineer should know about artificial intelligence" things didn't work out so well (link to excerpts here, Chapter 9, page 161). In the end it took far, far more resources to maintain and control the system, by some estimates 100 man years and 30 staffers to maintain over the course of a year - all to solve a problem which cost far, far less in the first place.
The bottom line is that, at least in my opinion, these types of systems are useless.
The first reason is that in an "expert domain" like configuring a computer "expert knowledge" is truly needed - if it weren't DEC would have hired $5/hour flunkies to do it - far cheaper than building an expert system.
The second reason is that it always seems as if the rules for these kinds of problems are "simple." But a big part of the problem with AI and "expert systems" in particular is that while its easy to start scribbling down rules about how to do something building a complete system that can handle the nuances of real world applications is not easy.
So having learned all this thirty years ago do we think the Wellpoint experience with Watson will be any different?
For one thing Watson is not performing any sort of "reasoning." Its merely correlating occurrences of text. When it sees "Created the telephone in 1875" it correlates this with "Alexander Graham Bell."
But it doesn't know what a telephone is.
It doesn't understand the concept of invention.
Its just finding things that look like they go together.
And sometimes its wrong.
So now we are applying this to healthcare.
Oh boy.
Wellpoint is looking to save money. So the call center in India is out and Watson with an electronic speech system is in.
Does Wellpoint really thing this is going to make their company more efficient?
No, it will allow them to spend less money per patient so that when they bill the government for Medicare they will profit more.
Some people never learn from the past...
Side Note: I used to work downstairs from an "AI" company in the 1980's. They were busy working on "healthcare" applications.
Like the "Turing Test" for intelligence we had the "Kueny Test" for commercial AI healthcare systems: The AI had to conduct a robotic proctological exam on the various software developers, product managers, domain experts, etc. If the results were indistinguishable from a protological exam by a human expert the AI system was ready for public use...
No comments:
Post a Comment