Hi, i am currently doing a project that require my .net website to be able to make some medical diagnosis given a set of symptoms.
Is neural network the way to go?
or shld i try other stuff like prolog?

if neural network is the way to go, how can i go about doing it?
is matlab the correct software to use if i have to link the neural network with my .net website?

thks alot

Well neural networks have certainly been used for medical diagnosis before and the system does lend itself to the sort of thing you are attempting to do here. It may also lend itself to being easily adapted to the way .NET has already been structured in that your neural connections or "the invisible neural layer" of the design could be a set of classes which you design. The user then inputs the symptoms and each class decides the optimal path by weighting the symptoms independently and then passes on those symptoms to the appropriate class until it reaches the end... resulting in a diagnosis.

I recommend taking a binary approach to this network by each class having essentially two classes or "nodes" it can pass to based on one simple decision. Basically you pass it a string of symptoms to one class, which will weight "Does the symptoms include a stomach ache" and if no, pass it to one class for further processing or "yes" and pass it to another class which then either breaks down the severity of the symptom or analyzes the accompanying symptoms to either formulate the conclusion or continue pass it it down to two more classes.

The only problem with this, as with any binary system and even a neural network in general, is that this hidden layer of decision making classes will grow exponentially and thus limit the range of diagnosis. It might be only able to diagnose an appendicitis or upset stomach but not much more. To cut down on this increased complexity, the more accurate you can make each class, the further reductions of decisions will be made and the flatter your overall binary decision tree will be.

I do think it is your best bet to go with a neural network design, but perhaps a limited diagnosis scope.

Hope that helps!

This post has been edited by Martyr2: 17 Sep, 2007 - 02:36 PM

.Net has a neat added language called F# that is a functional language that might be a better platform that Prolog if you are integrating with .Net applications.

As for the direction you will go: I assume you are asking about the difference between classic "expert systems" which essentially use large IF-THEN-ELSE blocks to follow a flow chart (much like you may find in a home medical reference) or more modern "fuzzy" system which may promise a more flexible system.

The expert systems are tried and true. Doctors have been using similar systems to aid in diagnosis for quite a long time. I know that fuzzy logic can be used to extend such systems to help aid in supplementing for "doctors opinion".

I think the question really comes down to how much time/money you have to spend. a fuzzy system would probably be better, but this would take much longer to implement and require much more research. The neural net or fuzzy system also requires a period of "training" which can be time consuming. If there is a legal element to your site (which dealing with medical diagnosis I would imagine it would) I would think that the expert system would be easier to defend.

Truth is. When dealing with such things, it is often better to seek out a 3rd party who specializes in such software. The basic reasons are listed above: time, money, legal, and responsibility.

Neural networks are a subset of machine learning and statistical techniques which are typically trained on a set of historic examples (past cases with symptoms and the known outcome). If you have a collection of past cases with which to train such a system, then neural networks are a viable candidate, as are any number of alternatives (logistic regression, discriminant analysis, kernel regression, tree induction, etc.). Which learning algorithm is most suitable will depend on a variety of factors, not least of which is your familiarity with any of them.

MATLAB is an ideal platform for developing learning algorithms, and in fact is my development tool of choice for this work. Presently, I work for a global bank building systems which predict which customers will pay back their loans and which won't, and MATLAB is my tool of choice. I do not know, though, how one would tie this back to VB.net.


-Will
Data Mining in MATLAB