5 ms·

Watson's Next Feat: Taking on Cancer

43 points by tacon 11y ago

gojomo 11y ago
If you read 'Watson' as 'whatever changing mix of machine learning techniques IBM throws at a particular domain', these PR stories make more sense.
iskander 11y ago
Is Watson a single fixed codebase or just a branding of multiple loosely related IBM research projects?
- andrenotgiant 11y ago
  It's a marketing campaign.
  - mercurialshark 11y ago
    I just had a conversation with an IBM Watson engineer. It was the worst professional conversation of my adult/professional life after I asked "is everything Watson has accomplished so far a PR campaign?" It's not they said, "we are moving into oil and gas." Ah, right.
    - agumonkey 11y ago
      They keep rebranding it. It feels like Java. Based on that, Watson will be a success, but not where it was intended to.
      - shiven 11y ago
        Yep, soon as Rich Hickey gets around to writing a Lisp for it. Wonder what he'll call it though. Clotson? Watjure? Wisp?
  - nkozyra 11y ago
    Ultimately, yes, I think it's unlikely there's any real magic there. If we use "Watson" as shorthand for "supervised learning logic agent" I think it makes more sense. It just means it's not entirely tied to IBM.
Gatsky 11y ago
I had a go using the Watson cancer decision making system. This was the experience: Web interface. Log in. Add new patient. Fill in many many different check boxes/drop down lists about the patient. Some of these seem irrelevant. Push a button. Treatment recommendations come up. Top recommendation is reasonable. Evidence summary provided about the recommendation. This is explicitly marked as being human curated from the MD Anderson, not generated by Watson. Links to possible clinical trials given - these were wrong, not applicable to the patient at all. So in summary, this is really designed for a doctor treating a cancer that they aren't really used to treating. There is no 'magic' that I could see. Maybe I used the most simplistic version, but the major limitation seems to be that it requires very structured input about the patient's case. I can't indicate for example that there was something unusual about the how the cancer looked under the microscope, or that the patient has a rare other medical condition that influences the treatment decision etc. Not surprising, because these are factors that lack an evidence based recommendation. If you read the article this is the key point: "I see technology like this as a way to really break free from our current health-care system, which is very much limited by the community providers. If you want expert care you have to go to an expert center," she said, "but there are never enough of those to go around." This is quite a different goal to 'revolutionising' cancer care, or 'taking on' cancer. There is clearly a lot of hype... But I laud the effort, it will likely end up being genuinely useful in the future. At the moment it seems to be a very thin machine learning layer on top of a highly domain-customised infrastructure.
teekert 11y ago
From the article: But he is emphatic that computers will never truly replace human doctors for the simple reason that the machines lack instinct and empathy. “There are a lot of things you can deduce by what a patient is not telling you, how they interact with their families, their mood, their mannerisms. They don’t look at the patient as a whole,” Husain said. “This is where algorithms fail you.” I don't understand this sentiment, instinct and empathy are just feelings that results from small cues that a physician can't exactly put into words. It does not put them into the realm of magic. And why exactly would a computer not be able to look at the patient as a whole? Of course a computer can (at some point) do that. It can even deduce heart rate an breathing rate from a camera image of a face, something a human can't. Why do people always assume that if something is difficult to put in words it is impossible for a computer somehow? It seems like an invalid argument to me.
- rohit89 11y ago
  > Why do people always assume that if something is difficult to put in words it is impossible for a computer somehow? It seems like an invalid argument to me. Because if we don't really know how we are doing something, how can you tell the computer how to do it? Simply measuring a bunch of values doesn't mean it will be interpreted correctly. For eg, see the lie-detector.
- agumonkey 11y ago
  The less we understand the more we think about it the more we attach ourselves to it ? And beyond that the most amazing part is the bias we have in this regard. Doctors fail too, most of the time accidentally (fatigue, stress, collaboration issues, lack of information,...) but some time out of hubris. Got to see this myself: one doctor told a condition required removing an organ, another one told what was wanted to be heard, guess which doctor had the job and what was the result ?
- gmarx 11y ago
  The hardest thing in medical diagnosis/treatment for a computer to do is not running algorithms to determine diagnosis/treatment; it's taking the patient history. If you ask a patient if he has had any abdominal pain, a lot of patients think about it and decide they have had abdominal pain. It may just have been some slight discomfort from gas a couple of weeks ago, but doc asked so it must be important; wouldn't want to leave anything out. Other patients are stoic and wouldn't really call that raging appendicitis "pain", more like "discomfort". An experienced doc should be able to read the patient and tell these apart. Watson? Maybe someday. As with most tasks in medical analytics, the rate limiting factor is the data.

43 points by tacon 11y ago

If you read 'Watson' as 'whatever changing mix of machine learning techniques IBM throws at a particular domain', these PR stories make more sense.

iskander 11y ago

Is Watson a single fixed codebase or just a branding of multiple loosely related IBM research projects?

andrenotgiant 11y ago
It's a marketing campaign.
- mercurialshark 11y ago
  I just had a conversation with an IBM Watson engineer. It was the worst professional conversation of my adult/professional life after I asked "is everything Watson has accomplished so far a PR campaign?" It's not they said, "we are moving into oil and gas." Ah, right.
  - agumonkey 11y ago
    They keep rebranding it. It feels like Java. Based on that, Watson will be a success, but not where it was intended to.
    - shiven 11y ago
      Yep, soon as Rich Hickey gets around to writing a Lisp for it. Wonder what he'll call it though. Clotson? Watjure? Wisp?
- nkozyra 11y ago
  Ultimately, yes, I think it's unlikely there's any real magic there. If we use "Watson" as shorthand for "supervised learning logic agent" I think it makes more sense. It just means it's not entirely tied to IBM.

Gatsky 11y ago

I had a go using the Watson cancer decision making system. This was the experience: Web interface. Log in. Add new patient. Fill in many many different check boxes/drop down lists about the patient. Some of these seem irrelevant. Push a button. Treatment recommendations come up. Top recommendation is reasonable. Evidence summary provided about the recommendation. This is explicitly marked as being human curated from the MD Anderson, not generated by Watson. Links to possible clinical trials given - these were wrong, not applicable to the patient at all. So in summary, this is really designed for a doctor treating a cancer that they aren't really used to treating. There is no 'magic' that I could see. Maybe I used the most simplistic version, but the major limitation seems to be that it requires very structured input about the patient's case. I can't indicate for example that there was something unusual about the how the cancer looked under the microscope, or that the patient has a rare other medical condition that influences the treatment decision etc. Not surprising, because these are factors that lack an evidence based recommendation. If you read the article this is the key point: "I see technology like this as a way to really break free from our current health-care system, which is very much limited by the community providers. If you want expert care you have to go to an expert center," she said, "but there are never enough of those to go around." This is quite a different goal to 'revolutionising' cancer care, or 'taking on' cancer. There is clearly a lot of hype... But I laud the effort, it will likely end up being genuinely useful in the future. At the moment it seems to be a very thin machine learning layer on top of a highly domain-customised infrastructure.

teekert 11y ago

From the article: But he is emphatic that computers will never truly replace human doctors for the simple reason that the machines lack instinct and empathy. “There are a lot of things you can deduce by what a patient is not telling you, how they interact with their families, their mood, their mannerisms. They don’t look at the patient as a whole,” Husain said. “This is where algorithms fail you.” I don't understand this sentiment, instinct and empathy are just feelings that results from small cues that a physician can't exactly put into words. It does not put them into the realm of magic. And why exactly would a computer not be able to look at the patient as a whole? Of course a computer can (at some point) do that. It can even deduce heart rate an breathing rate from a camera image of a face, something a human can't. Why do people always assume that if something is difficult to put in words it is impossible for a computer somehow? It seems like an invalid argument to me.

rohit89 11y ago
> Why do people always assume that if something is difficult to put in words it is impossible for a computer somehow? It seems like an invalid argument to me. Because if we don't really know how we are doing something, how can you tell the computer how to do it? Simply measuring a bunch of values doesn't mean it will be interpreted correctly. For eg, see the lie-detector.
agumonkey 11y ago
The less we understand the more we think about it the more we attach ourselves to it ? And beyond that the most amazing part is the bias we have in this regard. Doctors fail too, most of the time accidentally (fatigue, stress, collaboration issues, lack of information,...) but some time out of hubris. Got to see this myself: one doctor told a condition required removing an organ, another one told what was wanted to be heard, guess which doctor had the job and what was the result ?
gmarx 11y ago
The hardest thing in medical diagnosis/treatment for a computer to do is not running algorithms to determine diagnosis/treatment; it's taking the patient history. If you ask a patient if he has had any abdominal pain, a lot of patients think about it and decide they have had abdominal pain. It may just have been some slight discomfort from gas a couple of weeks ago, but doc asked so it must be important; wouldn't want to leave anything out. Other patients are stoic and wouldn't really call that raging appendicitis "pain", more like "discomfort". An experienced doc should be able to read the patient and tell these apart. Watson? Maybe someday. As with most tasks in medical analytics, the rate limiting factor is the data.