pjhamvs@cs.vu.nl (Summeren van Peter) (09/06/90)
Greetings to the reader,
I am trying to make a neural network which can classify rheumatic diseases.
In the network which I use ("opt": Written by Etienne Barnard at
Carnegie-Mellon University) there are some 150 input neurons, one hidden
layer and 16 classifications. Learning from 600 patientdata goes very well.
(I have 74 data of one patient and for the network I made most of them binary.)
But when testdata are given to the network, the result is that about 46%
is classified allright. Of course this can be due to the inputdata.
The classification doctors give to a conglomerate of data for one patient
does not have to be unique. From their side the classification I use
is a good one, and the problem of this 46% mine, because it is not a bad
result at all. It is very normal that a medical diagnosis can have more
than one possibility and also it is understandable that some classifications
are very difficult to handle, even for specialists in the field.
Yet I would like to add something from the field of neural-nets. It would
be nice to have an automatic classifier, something that looks at the input
and tries to find a classification itself. If such an undertaking would
succeed, than one could maybe have a discussion about the classification
this network makes and about the one of the medical field. I certainly do
not expect too much from such a discussion. It it quite possible that
some rheumatic diseases are due to an attack of viruses on our immume system
and then most of the observed data are just on the surface. But it could
help a little bit. And although this is tedious work, it has to be done.
Also, most of the results will be careful examined: better to check 10fold
than to give results too early. So the number "46%" could already be too
much. It is a so small sample!
About 10% of the population has to do with what is called "rheuma" and it
is a matter of choice to give money to more urgent areas(aids, hart diseases
etc). I am doing this kind of research out of personal interest and for
a thesis at the end of my study in information technology. There is here
in the Netherlands just one doctor (specialist) who does this work also.
(Of course much and much more is happening in the field of rheumatic
diseases, but not in software!!)
As I am very new in this field (I read and read) I think that the adaptive
resonance theory (Carpenter, Grossberg) could help - but there could be
other methods too. I have two questions:
1. can someone give me an ART implementation suited for this kind of research?
(in C perhaps?)
2. are there other automatic classifier systems?
So if someone could help me, it would be very fine.
Because the work I do is in the medical field I can give no further information.
Peter van Summeren