From: sbharris@ix.netcom.com(Steven B. Harris)
Newsgroups: sci.med,sci.med.cardiology,alt.activism,talk.politics.medicine
Subject: Re: Backlash against HMOs: a declaration of war (was Doctor-bashing)
Date: 17 Apr 1999 11:49:59 GMT

In <371778B5.1C2EFFD0@cs.uoregon.edu> Bret Wood
<bretwood@cs.uoregon.edu> writes:

>"Steven B. Harris" wrote:
>>
>>     I don't know if anyone used the phrase "program in."  I know I used
>> the phrase "data entry."
>
>Well, the data entry for the appropriate variables wouldn't be as
>difficult as the extreme example you used.


    It would if it was honest.  The problem with use of expert systems
in medicine is that it's all theoretical, and relies on assistance of
physicians with a few more variables beyond the great many they acount
for consciously or subconsiously when they actually are in the presence
of a real patient with a real problem.  You can find me studies where a
computer does better at managing a "paper patient" (made up case, or a
selected historical case, presented as already digested data) than a
doctor does.   But that's not how the real world works.   Real problems
in medicine don't present as a paragraph of problems in text.  They
present in the form of a real person who has real complaints.  Which
may or may not be relevent to the problem.  It takes a human mind to
sort that out.

From: sbharris@ix.netcom.com(Steven B. Harris)
Newsgroups: sci.med,alt.activism,talk.politics.medicine
Subject: Re: Backlash against HMOs: a declaration of war (was Doctor-bashing)
Date: 17 Apr 1999 19:23:21 GMT

In <37187FD4.70AE831F@cs.uoregon.edu> Bret Wood
<bretwood@cs.uoregon.edu> writes:

>"Steven B. Harris" wrote:
>>
>>     If you are talking about virtual Hopfield networks (which I assume
>> you are), then, yes, I do.
>
>Most expert systems have nothing to do with neural networks.  So I
>guess you don't.


   I know very well that they don't.  However, those were not the ones
I was talking about.  I was talking about systems in the future capable
of observing (which in practice means learning on their own-- since
observation is so difficult), so that all that data didn't have to get
entered by hand.

   Expert systems of the type we have now do not work in medicine,
except as aids in very narrow areas (antibiotic prescribing, for
example, where the system keeps abreast of culture sensitivity reports
from the lab at your hospital, and gives you advice based on a simple
(but unchanging) algorithm based on those).  Or programs like
INTERNIST, which require physical exam and patholgy report data (which
means a lot of preprocessing by brains).  Even with the first kind of
system, data entry is onerous, and is made up for only by the fact that
the system automatically generates hospital antibiotic orders to sign
off on, and monitors further culture developments and antibiotic level
values (though this is probably going to be less important as levels of
aminogylcocides become less important).

   In the future, we may see some valuable programs in the ICU, where
sensors for continuous and rapidly changing things like vital signs and
blood gasses can be tracked by computer and used to deliver better and
faster recommendations for fluid managment and ventilator setting
changes.  Or even to control ventilators directly, by direct feedback
without a human directly interupting the loop (the FDA will have a fit
at that, of course.  This has actually been possible for years, ever
since computer controlled servo ventilators and in-line blood gas
systems, but has yet to be implemented.  Mainly for political reasons).
As for expert systems that learn, they won't get through the FDA in
medicine until they start saving a lot of lives and money somewhere
else.

From: sbharris@ix.netcom.com(Steven B. Harris)
Newsgroups: sci.med,sci.med.cardiology,alt.activism,talk.politics.medicine
Subject: Re: Backlash against HMOs: a declaration of war (was Doctor-bashing)
Date: 17 Apr 1999 19:47:15 GMT

In <37187F9C.F79E7AF5@cs.uoregon.edu> Bret Wood
<bretwood@cs.uoregon.edu> writes:

>In specialized disciplines, I can find examples of expert systems
>which outperformed real physicians with REAL patients.

   Not by themselves, you can't.  There's always a medically trained
person hidden in the "expert system" somewhere.  Such combinations are
sometimes able to beat a specialist, just as with a computer I can
figure out some problems faster and more accurately than a mathematician
(not just arithmetic and statistics, but even some symbolic math, as
with MATHEMATICA).  But that's not too impressive.  I certainly could
not use a calculator to beat a mathematician if I didn't know any math
at all.  And with some math problems, a computer is no help at all.

   We had computer *assisted* chess 20 years ago, which could
significantly improve play for all but advanced players-- and sometimes
even for them, if they were having a bad day and were tending to make
dumb mistakes.  Computers have not even needed observational assistance
though much of that time, if special boards are used.  Still, it was
not until very recently that the computer became able to play as well
or better as the best players without *judgemental* assistance.
Medicine is far more complicated than chess, and it's likely to be much
longer before computers make many inroads without judgemental
assistance there.

>Obviously this isn't _always_ the case, or else expert systems would
>be replacing specialists in certain fields.

   It isn't even more likely the case, or else expert systems would be
replacing specialists in certain fields.


> But it does tend to show
>that the number of variables a physican deals with aren't as vast as
>previously thought.


   No.  At best, it illustrates that the number of extra variables
which the specialist deals with, over and above that of the average
doctor, is less in some disease states than the SPECIALSTS believed.

    But that's not the same thing at all.  The nonspecialists have long
suspected it, in fact, for a few things.  After you get three
specialist consults for the same sort of thing which all say the same
thing, you begin to get the message.  That's one of the ways teaching
happens informally in medicine.

                                        Steve Harris, M.D.

From: sbharris@ix.netcom.com(Steven B. Harris)
Newsgroups: sci.med,alt.activism,talk.politics.medicine
Subject: Re: Backlash against HMOs: a declaration of war (was Doctor-bashing)
Date: 20 Apr 1999 06:22:03 GMT

In <37192425.2A6A4C61@cs.uoregon.edu> Bret Wood
<bretwood@cs.uoregon.edu> writes:

>A traditional expert system can be linked into a neural network based
>visual system.  Personally, I wouldn't trust a self-learning neural
>network to make any critical decisions until we have a much better
>understanding of _why_ they work.  The general principles are fairly
>well understood at this point, but we really don't have a clue about
>how any individual network does its decision making.  So trusting
>a neural network means you essentially need to trust a "black box",
>without any way of validating its decision making process.

   Just as with any individual human (or, rather, no more and no less).
A computer that learns via neural net will certainly start with big set
of plain ordinary algorithms, which it will then refine under
experience (too time consuming to do it any other way).  If you ask why
it is doing what it is doing, it will give you the relevent borad
algorithm.  If you push it, or ask it why it is not applying the
algorithm in quite the same way to two patients who would appear to fit
it equally, it won't be able to tell you.  But then, a human doctor
won't either.  He or she will just end up saying: "I dunno.  It just
smells like an MI to me.  And I'm not talking about odor."


>I would suspect that considering the relatively primitive state
>of safety critical sofware systems, it will be a LONG time before
>anyone can convince people that a neural network is safe enough
>to rely on for critical judgements.  On the other hand, a more
>traditional expert system, augmented by neural networks for
>perceptual information processing would present a system which
>has a more "transparent" mode of operation, and could be more
>easily verified.  Also, explicit sanity checks could be built into
>the system, and a failure of the perceptual networks would be
>much less likely to result in a catastrophic action being taken
>by the whole system.

   Yes, I suspect you are right,  Whether or not this is really the
safest or fastest way to do it, is not clear.  Part of this lies in
what looks to be "transparent" and what doesn't.  A helicopter crashes
into a mountain at night in fog, and it appears that partly to blame is
the practice of flying with night vision optics,  which use IR to see
through fog, but screw up depth perception.  But it looks like you're
just using a really complicated set of glasses or goggles with snouts.
If the thing was more computerized, and the pilots sat below deck
somewhere with a set of VR glasses, or in a "holodeck" cubicle with
computer-generated images (which might actually cause less depth
perception problem), there'd be a lot more questions.  But there's
really not much difference between a person sitting in the right place
with something opague over his eyes, seeing an electronic image, and
somebody sitting where they "shouldn't" be, seeing an electronic image.
The only difference is that one is scarier to look at for the
passengers.  And won't get approved as soon.  Even if it is safer.



>>    In the future, we may see some valuable programs in the ICU, where
>> sensors for continuous and rapidly changing things like vital signs and
>> blood gasses can be tracked by computer and used to deliver better and
>> faster recommendations for fluid managment and ventilator setting
>> changes. Or even to control ventilators directly, by direct feedback
>> without a human directly interupting the loop (the FDA will have a fit
>> at that, of course. This has actually been possible for years, ever
>> since computer controlled servo ventilators and in-line blood gas
>> systems, but has yet to be implemented. Mainly for political reasons).
>
>More than just political reasons.  I don't recall the name of the
>software right now, but there was an X-Ray system a while ago which
>had a flaw in the software which ended up massively overexposing
>patients.


    Yes, a couple of people died.  After which the FDA decided that now
they control and approve all medical software.

   But you cannot run an industry like this.   Imagine two people dying
in an airplane accident in the early days of aviation.   The Feds now
decide they have compete control over the airplane development, and you
have to get new designs approved before you can build them.   Can you
imagine where we'd be?  The first guy to die in a powered airplane was
about the second passenger in one, a guy named Selfridge who died in a
crash with Wilbur Wright himself in 1908.  And it gets sillier from
there.  I saw on a postage stamp the other day a young woman in a
violet flight suit and goggles: "Women in Aviation" was the series.
Except that this particular woman (though it doesn't say so on the
stamp) is famous for advancing aviation by doing a loop in an airshow
during which her passenger, not belted in, fell out.  After which the
woman herself proceeded to do the same and fall out herself.  So she
and her passenger are dead, and this helped put seatbelts in biplanes.
Duh.   Had FAA investigations stopped all plane manufacture for such
accidents, we'd STILL be in biplanes.  We used test pilots and non-test
pilots and even passengers up at a rapid rate instead, and went from
horse and carriage to the moon in one human lifetime.  Just about all
the men who delivered airmail with Lindebergh didn't make it though the
early years alive.  But eventually, modern aviation arrived.  Just in
time for the country to make rapid progress after sputnik, then chicken
out, decided that every dead astronaut should shut down the entire
space program for months (Apollo 1) then years (the shuttle).

  That is the way not to make progress.  No guts.  There's no telling
how many commerical passenger lives and just plain lives (air, sea and
land rescue, etc) have been saved by having reliable air transport.
Just consider the lives saved by keeping people from traveling those
distances in automobiles.  And then there are the wars we might have
lost, that really couldn't afford to be lost.

   We're losing a war now-- the war against disease and death.   Death
by disease and aging is far nastier than the German armies ever were,
and it kills far more people every year than Auschwitz ever did, and
that just in this country alone.  The careful and prudent track is not
an option against such an array of horror.  But it's what we've decided
we'll do.  In a democracy, everybody gets what only the majority
deserve.  Too bad.  Perhaps the majority does deserve what it is
getting, due to its cowardice.  But not everybody does.



> The problem with relying on computers for more than the
>most basic safety-critical systems is that software engineering has
>not progressed to the point where it is really possible to guarentee
>the reliabilty of any significantly complex program.

   Hello?  When you're dying of cancer, you don't wait until somebody
can absolutely guarantee the safety of the radiotherapy machine
software (or anything else about it).  First, that's not the way you
live the rest of your life, very likely, even when you're well. Second,
 you're far from well.  If you don't do something drastic, you're going
to be dead. The FDA will not take responsibility for your safety, then.




>> As for expert systems that learn, they won't get through the FDA in
>> medicine until they start saving a lot of lives and money somewhere
>> else.
>
>Personally, I think that this is reasonable.  If we can have a learning
>system make decisions usually left to a doctor, wouldn't it make sense
>to field test it in an environment where there aren't enough doctors
>in the first place.


   Yes, but that is often considered unethical.  As when they tested
several anti-AIDS drugs against placebo in African populations which
couldn't afford them anyway.  And even if you give everyone active
treatment, there is always the question of why you're using poor people
as guinnea pigs.   Instead of testing a computer on them, why not fly
them home to Boston for the best medical care at Mass General, instead?
No answer to that.  There's really never any answer to the people who
feel everyone should have the cheap and reliable medical care, but who
don't dare do the experiments it takes to get everyone cheap and
reliable medical care.  Today, when poor people must travel, on many
routes they travel by commercial jet, at 85% of the speed of sound.
That's because it's far cheaper per mile than doing it by bus.  A lot
of people fell out of airplanes to make that possible.  Alas, nobody
learned a damn thing from it (in general).  They have a postage stamp
for this dead aviatrix I mentioned, whose name I've forgotten, but it's
not for the right reason.   There really ought to be one for her
passenger.  And another for the government officials who just shrugged
their shoulders and hoped the flying machines would improve.

                                          Steve Harris, M.D.

From: sbharris@ix.netcom.com(Steven B. Harris)
Newsgroups: sci.med,alt.activism,talk.politics.medicine
Subject: Re: The Patients' Bill of Rights (was Backlash against HMOs: a   
	declaration of war)
Date: 20 Apr 1999 07:20:13 GMT

In <7ff0ud$mco$1@nina.pagesz.net> henryj@nina.pagesz.net (George
Conklin) writes:

>   You raise an interesting point.  The medical business
>seems to be the only business in the world which wants
>government to pay for 100% of the research costs.  Drug
>companies pay for their own, mostly, but for the rest of it,
>doctors don't bother.


    It's not a matter of whether or not doctors "bother."  Research
costs money.  Drug companies (and industry and other patent-and
copyright protected folks) can fund it because they have a mechanism
for payoff from new knowledge and novel ideas.  Figure out a way for
other people in the information industry to profit by results of
research, and you'll see a lot more research.  As it is, we have
information-socialism in many areas, and that's like having
agricultural-socialism.  Lousy farming results.  Allow people to keep
some of the fruits of their labors and you'll see something quite
different.  Even statisticians will be driving BMWs.

    I've previously suggested that the reason so many academics are
socialists is that they envy people in other industries who actually
profit from having new ideas.   After you yourself have been stolen
from a while, as a poor academic, you start to get the idea that maybe
EVERYBODY should have to share.  Bill Gates, too.

                                    Every Man A King...
                                    Steve Harris, M.D.