Hi Steve, thanks for your reply.
"Simon, Steve, PhD" wrote:
> Many of us on the list are skeptical about artificial intelligence methods.
> They have proven useful for some applications, but they are often marketed
> as if they can substitute for consultation with a professional statistician.
Have no fear, my application would by no means hand you or your brethren a pink
slip. ;)
> Putting aside that skepticism, have you thought about how you are going to
> measure how successful your neural net is at mimicking a human pancreas? In
> particular, you do not seem to have data on insulin production, so how will
> you know that the neural net pancreas produces insulin at a rate comparable
> to the human pancreas?
Among the data that I am using are basal and bolus amounts of insulin injected
by the patient(roughly translated, small amounts throughout the day(basal) and a
large amount during meals(bolus)). This information will be used as a baseline
to train the neural net or genetic algorithm on how much insulin produces the
desired result. The data that I need to have "tweaked" or "interpolated", is
the blood sugar readings themselves. The standard practice of an insulin
dependant diabetic is to check blood sugars at various times during the day, and
then apply booster shots of insulin as needed. This is, of course, NOT the way
the human pancreas works. It provides a steady stream of insulin as necessary
as the blood glucose levels in the body rise and fall throughout the day. By
reading a blood sugar on a minute by minute basis and then producing smaller
dosages of insulin over a prolonged period, the pancreas operation may be
effectively simulated.
Since I only have data for patients where they are checking their blood sugars
at most 10 times a day, I need to get a minute by minute best guess as to the
condition of the blood sugar, based on the actual data I have.
The success of the neural net will be measured on how effectively the program
remains within the 80-120 blood sugar reading deemed "normal" for a Type I
diabetic. Too much insulin, the blood sugar readings dip into dangerously low
levels, a bad thing. Too little insulin, the blood sugar readings skyrocket,
also, a bad thing. By comparing the output of the AI to actual readings, I will
be able to guage how effectively the AI is at determining how much insulin to
give, based on previous readings.
Suffice it to say, the amounts of insulin delivered do not need tweaking. Only
the blood sugars.
> If you can quantify a measure of success, it might help us suggest what sort
> of interpolation you need. I suspect that you do not really want or need
> interpolation, because the interpolated values will not add any independent
> information to the training set above and beyond the actual data values.
> Think about it. The neural net uses weighted input to train itself. The
> interpolated values are weighted from the same input. So wouldn't the
> interpolated values be redundant?
The reason I need minute by minute data is because of the manner in which
insulin is produced by the body. Us non-diabetics produce a constant stream of
insulin. Diabetics produce none(or very little). The only viable option at the
moment is to give large doses of insulin at one time(not the best way, but it
has been the standard in diabetic care for a very long time). The ideal
solution would be to have insulin delivered on a constant basis. That's why I
need the interpolated data, so that i can apply the insulin levels over a
shorter time period.
> You might want to try training your net with just the actual data, and with
> some very simple interpolation, such as a straight line interpolation
> between adjacent points. Does the net with the interpolated values perform
> better than the net with just the original data?
I don't know. I haven't trained it yet. I have been working under the
assumption that it would be better to train it for a method that would be ideal
rather than a simulation of the current state of affairs. As such, I've been
waiting to see what kind of minute by minute data I can throw together.
> You might also want to think about what is known about insulin production.
> If it has cycles that are more frequent than every two hours, then any
> attempt at extrapolation will have to miss those cycles, wouldn't it?
If you could elaborate on this point, I'm afraid I don't quite follow you here.
> Try not to get too impatient with our questions. It is perhaps an attempt to
> get you to elaborate more about the fundamental nature of the problem.
Not at all, any help is most definitely appreciated. Any direct questions may
also be directed to me via email, and I will answer them as quickly as possible.
> Also, forgive me if my questions are a bit naive. It is hard to discuss
> complex issues via email.
Again, I understand completely. Thank you for your time.
--
----------------------------------------------------------------------------
| Jeff Goslin - MCSD | "Oh Bentson, you are so mercifully free from the |
| [EMAIL PROTECTED] | ravages of intellect." --Evil, The Time Bandits |
----------------------------------------------------------------------------
| how come everyone elses religion is a cult but your cult is a religion |
----------------------------------------------------------------------------
| XGenetic, the ActiveX Artificial Intelligence Genetic Algorithm |
| http://www.winsite.com/info/pc/win95/demo/xgen-sw.zip/ |
----------------------------------------------------------------------------
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================
